Citation
Neural network-based control designs for complex industrial process applications

Material Information

Title:
Neural network-based control designs for complex industrial process applications
Creator:
Lefebvre, Wesley Curtis, 1965-
Publication Date:
Language:
English
Physical Description:
x, 187 leaves : ill. ; 29 cm.

Subjects

Subjects / Keywords:
Architectural models ( jstor )
Boilers ( jstor )
Combustion ( jstor )
Datasets ( jstor )
Dynamic modeling ( jstor )
Furnaces ( jstor )
Modeling ( jstor )
Neural networks ( jstor )
Parametric models ( jstor )
Steam ( jstor )
Automatic control ( fast )
Intelligent control systems ( fast )
Neural networks (Computer science) ( fast )
Genre:
bibliography ( marcgt )
theses ( marcgt )
non-fiction ( marcgt )

Notes

Thesis:
Thesis (Ph. D.)--University of Florida, 2000.
Bibliography:
Includes bibliographical references (leaves 181-186).
General Note:
Printout.
General Note:
Vita.
Statement of Responsibility:
by Wesley Curtis Lefebvre.

Record Information

Source Institution:
University of Florida
Rights Management:
The University of Florida George A. Smathers Libraries respect the intellectual property rights of others and do not claim any copyright interest in this item. This item may be protected by copyright but is made available here under a claim of fair use (17 U.S.C. §107) for non-profit research and educational purposes. Users of this work have responsibility for determining copyright status prior to reusing, publishing or reproducing this item for purposes other than what is allowed by fair use or other copyright exemptions. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder. The Smathers Libraries would like to learn more about this item and invite individuals or organizations to contact the RDS coordinator (ufdissertations@uflib.ufl.edu) with any additional information they can provide.
Resource Identifier:
45125835 ( OCLC )
ocm45125835

Downloads

This item has the following downloads:


Full Text









NEURAL NETWORK-BASED CONTROL DESIGNS FOR COMPLEX
INDUSTRIAL PROCESS APPLICATIONS
















By

WESLEY CURTIS LEFEBVRE














A DISSERTATION PRESENTED TO THE GRADUATE
SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL
FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE
OF DOCTOR OF PHILOSOPHY

UNIVERSITY OF FLORIDA 2000













ACKNOWLEDGMENTS

When looking for a graduate school I was told to pick an advisor, not a school. My choice was Dr. Jose Principe, and his mentoring and friendship through a master's, Ph.D., and the startup of two companies has proven this advise to be sound. I would also like to thank Dr. Glen Johnson from IBM Yorktown for giving me this advice and first introducing me to the field of neural networks.

I would also like to thank Charles River Associates and Commonwealth Energy for believing in my work and this field, the members of my committee for the effort and time spent on my behalf, the staff at Canal Electric Generating Station, and the members of the Computational NeuroEngineering Laboratory (CNEL).

I would also like to thank my wife, Anna, for her patience as I struggled to juggle far too many commitments, and my family for their never-ending support referring to us as Dr. and Mr. Lefebvre.


















11














TABLE OF CONTENTS

page

A CKN OW LED GM EN TS ............................................... ..........................................ii

LIST OF TABLES ...................... .. .... .. ....... ........................v

LIST OF FIGURES ......................................................................................................vi

CHAPTERS

1 INTRODU CTION ....................................................................................................1

Proposed W ork .....................................................................................................1..
D ocum ent Organization .........................................................................................7

2 LITERA TURE REV IEW ...........................................................................................9

Optim ization .......................... ..........................................................................9
N eural N etw orks ......................................................................................................16
Control Theory ................................................. ......................................................32
N Ox ......................................................................................................................... 40
Fossil-Fired Pow er Generation .......................................... ....................................45

3 BOILER OPTIM IZATION .......................................................................................63

First Principles ........................................................................................................63
Fuel and Air D istribution ........................................................................................64
Boiler Tuning ..........................................................................................................65
The Role of CO .........................................................................................................66

4 CON TROL D ESIGN S ..............................................................................................67

V ariable D efinitions ................................................................................................67
Optim ization Objectives ........................................................................................68
Operating Constraints .............................................................................................70
Perform ance Criteria ..............................................................................................72
Controller D esigns ........................................................................... .......................73



iii








5 DATA PREPARATION .............................................................................................90

D ata M anagem ent ...................................................................................................91
V ariable Selection ...................................................................................................92
V alidation ................................................ ....... ....................................................... 93
Tim e Constants .......................................................................................................97
N orm alization .........................................................................................................97

6 M OD ELIN G ...........................................................................................................99

M ethodology ............................................................................ ...............................99
M odel D efinitions ...................................................................................................100
D atasets ..................................................................................................................... 104
Perform ance Criteria ...............................................................................................105
Learning A lgorithm .................................................................................................109
V ariable Pruning .....................................................................................................113
Architecture Selection .......................................... ...............................................118
Analysis .................................................................................................................... 136

7 CON TROLLER IM PLEM EN TATION S ...................................................................138

O ffline Q uantification ..............................................................................................138
Online Quantification ..............................................................................................143

8 PARAM ETERIZATION PROBLEM ........................ ............................................148

Search for a V alidation M etric .................................................................................149
Correlation Paradox .................................................................................................154
V alidation M etric .....................................................................................................160
Revised Representation Pruning A lgorithm ....................... ................................161
M odeling ...................................................................................................................168
Control Im plem entation ...........................................................................................169

9 CON CLU SION ..........................................................................................................176

Contributions ............................................................................................................176
A fterword ..................................................................................................................178
Future D irection ........................................................................................................178

APPEND IX .................................................................................................................... 180

REFEREN CES ................ ... .................................................................................181

BIO GRAPH ICA L SKETCH .........................................................................................187




iv














LIST OF TABLES

Table 1: Probability of Z-Score exceeding value .......................................................96

Table 2: Types in order they were rem oved ...............................................................116

Table 3: Final variable selections after pruning ........................................................118

Table 4: Results of auto-regressive tap search algorithm for all dynamic models .....121 Table 5: Results of moving average tap search algorithm for all dynamic models ....122 Table 6: Results of hidden layer #1 PE search algorithm for all steady-state models. 124 Table 7: Results of hidden layer #2 PE search algorithm for all steady-state models. 126 Table 8: Results of tap search algorithm for all dynamic models ..............................128

Table 9: Results of taps search algorithm for all dynamic models ............................130

Table 10: Results of hidden taps search algorithm for all dynamic models ................131

Table 11: Results of hidden states search algorithm for all dynamic models ..............133

Table 12: Results of state hidden PEs search algorithm for all dynamic models ........134 Table 13: Results of output hidden PEs search algorithm for all dynamic models ......136 Table 14: Final variable selections after revised pruning ............................................165

T able 15: E ssensial tag list ..................................................... .....................................180












v














LIST OF FIGURES

Figure 1: Multilayer perceptron model architecture ....................................................18

Figure 2: TDNN input PE connectivity .....................................................................19

Figure 3: Gamma memory processing element ...........................................................20

Figure 4: Nonlinear state space neural network configuration ....................................21

Figure 5: Combustion emissions characteristic versus air flow ..................................64

Figure 6: Effect of lower 02 on combustion emissions ..............................................65

Figure 7: Offline training and retuning configuration for steady-state optimizer ........75 Figure 8: Online control configuration for steady-state optimizer ..............................76

Figure 9: Offline training and retuning configuration for model-inverse controller ....79 Figure 10:Online control configuration for model-inverse controller ..........................80

Figure 11:Offline training and retuning configuration for steady-state optimizer ........82 Figure 12:Offline training and retuning configuration for model reference controller. .86 Figure 13:Online control configuration for model reference controller .......................89

Figure 14:Daily % missing across February dataset ....................................................94

Figure 15:Example of variables with large NMSE but high R .....................................108

Figure 16:Results of type pruning algorithm ...............................................................115

Figure 17:Representation pruning algorithm results ....................................................117

Figure 18:Results of auto-regressive taps search algorithm for NOx ARMA Model ... 120 Figure 19:Results of moving average tap search algorithm for NOx ARMA Model ... 121 Figure 20:Results of hidden layer #1 PE search algorithm for the NOx MLP Model. ..124 vi














Figure 21 :Results of the hidden layer #2 PE search algorithm for NOx MLP Model. ..125 Figure 22:Results of tap search algorithm for NOx TDNN Model ..............................127

Figure 23:Results of taps search algorithm for NOx GNN model ...............................129

Figure 24:Results of hidden taps search algorithm for NOx GNN model ...................130

Figure 25:Results of hidden states search algorithm for NOx NLSS model ................133

Figure 26:Results of state hidden PEs search algorithm for NOx NLSS model ..........134 Figure 27:Results of output hidden PEs search algorithm for NOx NLSS model ........135 Figure 28:Best models for all model definitions by architecture .................................137

Figure 29:Average NOx reduction over testing dataset ..............................................141

Figure 30:Average CO above max over testing dataset ...............................................141

Figure 31:Average NOx reduction over testing dataset using train and test models ....142 Figure 32:Average CO above max over testing dataset using train and test models ....142 Figure 33:Change in NOx for steady-state controller experiments ..............................146

Figure 34:Final CO level for steady-state controller experiments ...............................147

Figure 35: Summary of validation metrics for MLP CV model ....................................150

Figure 36:NMSE and R for all 10 training results for MLP CV model .......................151

Figure 37: Summary of validation metrics for combined SV/CV model ......................153

Figure 38: Sensitivity results for all 10 training results for NOx CV model .................160

Figure 39:NOx CV model sensitivity with 95% confidence intervals .........................161

Figure 40:Results of revised representation pruning algorithm ...................................164

Figure 41:Sensitivity results for all 10 training results for revised NOx CV model .....167 vii














Figure 42:Revised NOx CV model sensitivity with 95% confidence intervals ............168

Figure 43: Best revised models for all model definitions by architecture ......................169

Figure 44:Average NOx reduction over testing dataset using old and revised models. .170 Figure 45:Average CO above max over testing dataset using old and revised models. .170 Figure 46: Change in NOx for revised steady-state controller experiments .................172

Figure 47:Final CO level for revised steady-state controller experiments ...................172

Figure 48:Average percent NOx reduction for 10 online experiments ........................173

Figure 49:Average percent CO reduction above 500ppm for 10 online experiments. ... 174






























viii













Abstract of Dissertation Presented to the Graduate School of
the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy

NEURAL NETWORK-BASED CONTROL DESIGNS FOR
COMPLEX INDUSTRIAL PROCESS APPLICATIONS By

Wesley Curtis Lefebvre

May 2000


Chairman: Jose C. Principe
Major Department: Electrical and Computer Engineering

Neural networks have successfully transitioned from an academic interest into a viable technology which is now being used in everyday products. To date, neural networks have been predominantly applied to forecasting or modeling applications. Based on their success in such applications, there has been significant interest in using neural networks in control applications, creating a new field called neurocontrol. Although there have been significant advances in the theory of neurocontrol, there are very few successful commercial applications using neurocontrollers. Commercial applications often provide the most challenging problems because the controllers are required to function robustly in complex and unknown environments. Real-world processes are complex and difficult to control because they contain a large number of highly interdependent variables, have highly nonlinear responses to these variables, and change their response over time.

This work identified two significant reasons why neurocontrol designs fail in realix














world applications: First, the controllable parameters over most industrial processes are highly correlated, often not for physical reasons but because of our process control strategies. Second, intermediate process states that affect the process output, which are also affected by the controllable parameters, have a significant impact on controller performance. When the controller changes the controllable parameters, the impact that this has on the process states, which will in turn affect the process output, is not accounted for in most neurocontrol designs in the literature.

This dissertation advances the field of neurocontrol by providing the following solutions: first, the use of statistical significance testing on the local linearized relationships extracted from nonlinear neural network models to avoid problems with correlated controllable parameters; second, augmenting neurocontrol designs to incorporate dependent state models. These enhancements have been applied to four distinct neurocontrol architectures. The new control architectures have been applied to the novel application of controlling NOx emission from an oil and gas-fired electric power plant.
















x













CHAPTER 1
INTRODUCTION

Neural networks have found an applications niche as robust predictors that have demonstrated the ability to out-forecast more traditional methods in complex real-world applications. The vast majority of neural network applications to date rely solely on the model's ability to forecast without regard for what the model has learned about the relationships within the underlying process or the ability to affect it. Soulie and Gallinari [59] recently compiled 53 industrial applications of neural networks, of which only 4 make any attempt to make inferences about the underlying process or to control it.

This ratio does not apply to theoretical publications in the literature, however. The research into neural network-based control has recently received widespread attention, coined neurocontrol [40]. Many authors have presented the abstract concepts behind neurocontrol [40][43][31][37], but the literature contains a disproportionately low number of papers presenting real-world neurocontrol applications. The vast majority of these papers have developed controllers for fabricated, simulated, or laboratory controlled processes.



1.1 Proposed Work

The goal of this project is to develop robust neurocontrol design methodologies for

complex industrial process control applications. Industrial control applications are characterized by nonlinear, noisy, non-Gaussian, highly correlated, and nonstationary processes. This is an applications area where classical control designs have proven ineffective.



1






2

1.1.1 Case Study

As a case study, this work develops on-line advisory neurocontrollers designed to minimize the NOx emissions for an oil and gas co-fired power plant. The combustion of fossil fuels inside a large-scale boiler is a highly complex process; this complexity is a direct function of the boiler size. A typical electric power boiler maintains a "fireball" which is 3 to 5 stories tall, and there are hundreds of parameters which affect the injection of fuel and air at different locations within the furnace. The problem is our lack of understanding about how these combustion parameters affect NOx formation. This multivariate optimization problem requires a technology that can look at the process globally and determine the appropriate combination of combustion controls

The neurocontrollers developed in this study will forward control setpoints to operators through the plant's existing distributed control system (DCS). The neurocontrollers will be required to provide setpoints that minimize NOx emissions, while maintaining unit operating constraints. A demonstration system has been completed at Canal Electric's 580-MW tangentially fired Unit 2. Charles River Associates (CRA) and Commonwealth Energy jointly funded this study.


1.1.2 Objectives

The author believes that neurocontrol strategies have not been more successful in realworld applications because they

1) are difficult to implement,

2) fail to account for dependent internal process states, and 3) have difficulty dealing with correlated process variables.

The scientific objectives for this work are to develop:






3
1) an application-based neurocontrol implementation methodology,

2) state-space neurocontrol architectures,

3) methods for dealing with correlated data,

4) accurate combustion models, and

5) a novel combustion controller.

1.1.2.1 Application-Based Neurocontrol Implementation Methodology

Most neurocontrol implementations in the literature have been ad hoc and application specific [40]. In many cases, the process has been simulated and thus known completely. Modem control theory has introduced many sophisticated control designs, but the fact remains that approximately 90% of the industrial control applications apply a simple proportional integral differential (PID) controller [37].

The PID controller only requires the process engineer to specify reasonable knowledge of the process. This application-based implementation methodology is largely responsible for the success of PID control in the industry. If neurocontrollers are to enter the mainstream process control market, there will have to be designs that do not require detailed knowledge of neural network or neurocontrol theory.


1.1.2.2 Application-Based Neurocontrol Implementation Methodology

We begin in Chapter 4 by defining a methodology for categorizing process variables into groups based on a set of objective criteria about their role in the process. Each controller will be implemented in Chapter 7 based on this labeling of the process variables, without additional process knowledge. This will allow process engineers with reasonable process knowledge and without any knowledge of neural network or neurocontrol theory to successfully deploy such technologies. Notice that achieving such a straightforward






4
implementation methodology does not imply that the details behind the neurocontroller implementation are easy, it simply requires that they can be automated.


1.1.2.3 State-Space Neurocontrol Designs

Nonlinear state-space neural network architectures offer the greatest modeling potential, but the difficulties in their training have led investigators to reject their application [31]. In fact, most neurocontrol designs are based on input/output process models [31][43][40]. The use of input/output neural network architectures in the design of neurocontrollers assumes that all input variables are independent, a situation which in not likely to be found in real-world applications. Many of the process variables that affect the plant's output will depend on the same process setpoints that a controller is manipulating. These facts have limited the performance and complicated the implementation of neurocontrol designs. If neurocontrol is to be a viable methodology in industrial process situations, then the controllers will have to be extended with architectures capable of dealing with internal process states.

There has been limited success in applying state-space neural network models [31]. It is widely accepted, however, that state-space representations hold the most promise for modeling and controlling complex processes [34][35]. The literature seems to be treating the viability of state-space architectures as an all-or-nothing affair. Most publications apply purely input/output architectures with overwhelming success, while a few investigates have tested purely state-space architectures where all states are treated as hidden and unknown with limited to no success.

This work will empirically investigate several shades of gray, ranging from purely

input/output to purely state-space controllers. These state-space controller designs are pre-






5

sented in Chapter 4, and their performance is empirically investigated in chapters 7 and 8. The primary difference in the state-space representations proposed here is that the state variables are treated as not hidden and known. This will require an extension to the neurocontrol design strategies presented in the literature [31][43][40].


1.1.2.4 Methods for Dealing with Correlation

Industrial process applications are unique in that there is a massive amount of available data. The input variables for complex process models are typically highly correlated, a situation for which there are few solutions in the literature. This correlation can come from several sources: dependent states (as addressed above), physical linkages, and soft linkages through control strategies the lack of adequate system parameterization. Industrial process control applications will require the process engineer to select representative variables from a large set of available process variables. Because of issues like input correlation, the representative variables selected will have a significant impact on the performance of the resulting neurocontroller. A viable neurocontrol design methodology will have to be able to cope with correlated process variables. Linear control theory deals with this aspect through parameterization of the controller. Nonlinear control theory, however, has not solved this problem in general. In Chapter 8 we propose to compute sensitivities through a committee of trained neural models to select the best variables for system identification and control.

State-space neurocontrol architectures will be able to explicitly deal with one source of correlation present in industrial processes, namely correlation produced by dependent state variables. As mentioned above there are several other sources of correlation, however. Our first goal is to empirically quantify the impact of this correlation on our model-






6

ing and control objectives. We begin by ignoring other sources of correlation and investigate their impact in Chapter 7. Once specific problems have been identified, methods are developed for dealing with correlation during controller implementation in Chapter 8, and the performance of these methods are quantified with respect to the performance of the resulting controllers.


1.1.2.5 Accurate Combustion Models

Little is known about how NOx is formed from air-bound nitrogen during combustion. To date, reliable models for NOx formation in electric power boilers have not been available [39]. There are not adequate models for many real-world industrial processes, a fact which has also limited the acceptance of modern control strategies. One reason that neurocontrol has received such widespread attention is because of its potential ability to deal with very complex processes that have escaped modeling.

Chapter 6 develops accurate combustion models according to accepted modeling performance metrics. Chapter 7 demonstrates the impact that correlation has on neurocontrol designs, and Chapter 8 investigates its impact on the accuracy of the underlying process models. Here it will be shown that accuracy is subjective, and that in fact no good metrics for model accuracy are available in the literature. A new metric is proposed and empirically compared against available metrics in the literature in Chapter 8. Applying this new metric, predictive combustion models are developed and used to shed light on which process variables have the greatest impact on NOx and CO formation.






7
1.1.2.6 Novel Combustion Controller

This project develops four neurocontrollers for the complex industrial process of NOx formation. We begin in Chapter 3 by looking at boiler optimization from a first-principles perspective, focusing on what a NOx controller is expected to achieve and why classical control methods are not able to achieve it. Predictive neural network combustion models are then developed in Chapter 6, and deployed within online neurocontrollers in chapters 7 and 8. The performance of each of these controllers is then quantified to compare and contrast the four control designs in Chapter 8.

To the authors knowledge, this work developed the first NOx controller for a gas and oil co-fired electric power plant. New regulations and the restructuring of the electric power industry have combined to create a NOx trading market. The annual benefits to a gas and oil co-fired electric power plant associated with a 25% NOx reduction will be in the range of $2,000,000 to $5,000,000. Clearly, a control strategy that uses existing plant capital investments and runs on a $2,000 pentium workstation has tremendous value.



1.2 Document Organization

Chapter 2 presents a summary of the required background and a literature review of the relevant work in neurocontrol, along with references to the literature for more detailed treatments. This chapter is the best place for readers to become familiar with the notation used throughout this work.

A strategy for reducing the NOx emissions from a fossil-fired generating unit is provided in Chapter 3. The goal of this section is to provide a physical understanding for what we are asking the controllers to perform, thus providing justification that our objectives are feasible. Chapter 4 develops four detailed neurocontrol designs belonging to the model






8
predictive, model inverse, and model reference control families. The designs are presented as generalized methodologies that are applicable to any control application. Chapter 5 presents a management and preprocessing methodology for collecting data in support of these control designs and the required modeling.

Each control design considered requires accurate process models for its implementation. Chapter 6 presents a modeling methodology for developing these models. The controllers are then implemented in Chapter 7. The performance for each resulting controller is then quantified using offline simulations and online experiments. Significant problems are discovered with the controllers for which there are no solutions in the literature. These problems, along with proposed solutions, are investigated further in Chapter 8. This section additionally demonstrates the validity of these solutions by quantifying the performance of the revised controllers.

The "key learnings" and extensions to this work are summarized in Chapter 9.













CHAPTER 2
LITERATURE REVIEW

This chapter provides background for the rest of this document in the areas of:

1) optimization,

2) neural networks,

3) neurocontrol,

4) NOx, and

5) fossil-fired power plants.


2.1 Optimization

Mathematical optimization methods are at the heart of modem modeling and control applications. Neural networks use optimization methods to facilitate learning, and control applications apply these methods to meet their control objectives. The notation and methods presented in this section will be used extensively throughout the rest of this document.

Optimization is defined as the process of finding the values of Nz decision variables

ffv NNz
z N that minimize a scalar performance objective J e N 91 [15]. Formally, this optimization task will be represented as ArgMin{J(1)} (1) where N is the decision variable space, which is most often taken to be euclidean N = 91. Optimization methods, also known as mathematical programming methods, can be classified according to the amount of a priori information available about the system


9






10

being optimized [54]. The following sections broadly categorize optimization methods into the following:

1) Classical Analytic Optimization: where the system being optimized is
known completely or nearly completely and a tractable analytic solution
exists

2) Descent Optimization: where first and/or second-order partial derivatives are available everywhere for the parameters of the system being
optimized

3) Direct Optimization: where little to no a priori knowledge exists about
the physical structure of the system being optimized

If the optimization problem involves objective functions or constraints which cannot be stated as explicit functions of the design variables or are too complicated to manipulate, we cannot solve it by using classical analytic optimization methods. This work will be dealing with complex systems where little is known a priori and will therefore not consider analytic optimization methods.


2.1.1 Iterative Optimization Methods

All direct and descent optimization methods are iterative in nature, i.e., they start from an initial trial solution and proceed toward the minimum point in a sequential manner. An iterative optimization method is typically judged based on its rate of convergence [54]. In general, an optimization method is said to have convergence of order p if 1(n + 1)- <*11 1 (2) II z(n) Z'l


where k(n) and k(n + 1) denote the points obtained at the end of iterations n and n + 1, respectively, '* represents the optimum point, and II|l denotes the length or norm of the vector &.






11

If p = 1 and 0

2.1.2 Direct Optimization

In problems where analytic solutions are not possible and the design variables are of mixed type there is little choice but to use some variation on a direct search methodology. Direct searches may be broken into the following broad categories.


2.1.2.1 Exhaustive methods

In most practical applications, the optimum solution is known to lie within restricted ranges of the design variables. Exhaustive search methods are applied to problems where the interval in which the optimum is known to lie is finite. Conceptually, these methods evaluate the objective function at a predetermined number of points in this interval and reduce the interval of uncertainty using the assumption of unimodality. Exhaustive methods include [54]:

1) Random Search

2) Grid Search

3) Pattern Directions

2.1.2.2 Elimination methods

The exhaustive search methods are similar to a larger class of algorithms known as

elimination methods, because they search by eliminating parts of the interval. Elimination methods differ in how they search and discard sub-intervals. The more common elimination methods include [54]:

1) Dichotomous Search






12
2) Interval Halving

3) Fibonacci Method

4) Golden Section Method

2.1.2.3 Interpolation methods

Interpolation methods iteratively fit the local performance surface with a simple polynomial form, and then approximate the minimum point of the system as the minimum point of the polynomial [65]. These methods are generally more efficient than elimination methods and can be accelerated if gradient information is available. Some of the more popular interpolation methods include [54]:

1) Quadratic Method

2) Cubic Method

3) Newton Method

4) Quasi-Newton Method

5) Secant Method

2.1.2.4 Unrestricted methods

When the design variable range is not known the search must be performed without restrictions on the values of the variables. Most of these methods use a step size and move from an initial guess in favorable direction (positive or negative) [54]. The step size used must be small in relation to the final accuracy desired. This method is often accelerated by using a variable step size. These methods include [54]:

1) Simplex Method

2) Revised Simplex Method

3) Karmarkar's Method

4) Hook's and Jeeves' Method






13

5) Rosenbrock's Method

In addition, evolutionary computing techniques like genetic algorithms belong to this category.


2.1.2.5 Line search

All of the direct search methods presented above can be applied to both one-dimensional or n-dimensional searches. A one-dimensional search is often referred to as a line search since we are searching along a line. The aim of all line searches is to find i* e 91 such that


n* = ArgMin {J( + 7 d)}, (3) where is the design vector, and d is a known search direction.

One of the most efficient, and hence most popular, line searches uses the Quadratic Method to find 4* [54]. This method has been applied in this work, using the following algorithm:

I: Normalize the search vector zd by dividing each component by the

absolute value of the element of zd with the maximum absolute value

II: Evaluate the function at the points A = 0 and D = o0, where 10 is
an initial step size

III: If JD> JA then set C = D andB = 10/2

IV: Else set B = D and evaluate at the point E = 2rI0. If JE > JD then
set C = E. Else set D = E and T0 = 20, and goto step III

V: Calculate

4JB- 3JA-Jc (4) 4JB 2JC- 2JA






14

VI: If J < AJn' then set T* = and quit 11

VII:If
VIII:Goto step V

where A'in is the minimum change in J to detect early stopping.


2.1.3 Descent-Based Optimization
9z
When all values of e 9I are possible and the function J(1) has first and second partial derivatives everywhere, the necessary conditions for a local minimum are 8J
= 0, (5) by which we mean 8J/8zi = 0, Vi and a_2 J> (6)
2


by which we mean that the m x m -matrix whose components are J/a8zizj must be positive semi definite, i.e., have eigenvalues that are zero or positive [15].

All points that satisfy (5) are called stationary points. Sufficient conditions for a local minimum are (5) and

a2
> 0, (7)
0 ,2


that is all eigenvalues must be positive. If(5) is satisfied but 892 jia2 = 0, that is, the determinant of the matrix is zero (meaning that one or more of its eigenvalues is zero),






15
additional information is needed to establish whether or not the point is a minimum. Such a point is called a singular point.


2.1.3.1 Methods

Classical analytic optimization methods use these conditions to solve for the optimal solution. If the optimization problem involves an objective function or constraints that can not be stated as explicit functions of the design variables, or which are too complicated to manipulate, then descent optimization methods provide efficient alternatives. In general, these methods will have significantly better convergence characteristics than direct methods [54].

Descent search methods are iterative algorithms for improving estimates of the decision variable, ,, so as to come closer to satisfying the conditions for a stationary point. The steps in using the descent method are as follows:

I: Set n = 0 and guess at the initial design vector ,(n), usually random

II: Determine the values of dJ/81(n)

III: Interpreting 8J/a(n) as the gradient vector, determine the search

direction d (n) = f (OJ/8k(n)) as a function of this gradient

IV: Determine the step size to be taken T(n) = f(zd(n)), as a function
of this direction

V: Update the estimates of Z(n + 1) = '(n) + (n)z'd(n) VI: Repeat II until (8H/8t(n))(8H/9(n))T is very small






16
The variations in descent-based optimization can be expressed as variations in the


determination of the direction vector / and the step size f Some of the more common variations include [15][54][65]:

1) Steepest Descent:

dz (n) = -8H/a(n) (8) 1 (n) = constant (9) 2) Steepest Descent with Momentum:

zd(n) = (1 p)aH/a(n) + pz' d(n) (10) T (n) = constant (11) 3) Conjugate Gradients:


d (aH/8(n + 1))T[aH/a(n + 1)- aH/a(n)] (12) (8H/8k(n + 1)) 8H/8~(n + 1)


1 (n) = LineSearch(Ad(n)) (13)


2.2 Neural Networks

Artificial neural networks (ANNs) are biologically motivated data processing structure that consist of a large number of relatively simple highly interconnected neurons or processing elements (PEs) [24]. In general, these structures provide an inductive mathematical model that can be represented by ) = f(, T), (14) where f:9N --+ 9 is the model's input/output map, and e 91 h~ E9 and 9Nw are its outputs, inputs, and coefficients, respectively. The coefficients in an






17

ANN map are commonly referred to as weights. Artificial neural networks infer or learn the relationships between and b by observing actual process data. In this way, ANNs can be applied to generalized regression and classification inference problems.

ANN architectures possess two fundamental properties:

1) They are capable of approximating to arbitrary accuracy any continuous
function, i.e., they are universal mappers [24].
2) They have robust optimization convergence properties with respect to
the optimization of their coefficients, i.e., they are robust learners [26].

These properties make ANNs a useful tool for empirical modeling tasks where little to no a priori information is available about the underlying process.


2.2.1 Model Architecture

There are many types of ANNs in the literature, each with specific advantages when modeling various types of processes [14]. The two primary factors which differentiate between ANN models are their architecture and their learning rule. A model's architecture defines the way in which it processes input information to produce output information, i.e., the form of their mathematical input/output map f.


2.2.1.1 Multi-layer perceptron

Most ANNs presented in the literature are static mappers, i.e., they are only capable of modeling static or steady-state process relationships. By far the most popular and widely applied ANN architecture is called the multilayer perceptron (MLP) [24]. This network consists of fully-interconnected layers of PEs with logistic response characteristics. The MLP network is typically configured with one or two hidden layers ofPEs. A two-layer MLP is illustrated in Figure 1.






18











Figure 1: Multilayer perceptron model architecture.


Formally, using matrix algebra this architecture is given by

f/lp(h, o) = 0y(1h2 1(lTlh + hI)+ h2)+ Ay, (15) where le N,, x N" is the matrix of weights for the first hidden layer, h le N" is a vector of bias values for this layer, 2 C 2 X NhI is the matrix of weights for the second hidden layer, Ah2 e Nh2 is the bias vector for this layer, Y N x N12 is the matrix of weights for the output layer, A 9? NY is the corresponding bias vector, Nhl is the number of PEs in the first hidden layer, Nh2 is the number of PEs in the second hidden layer, acr is the tanh logistic function, and the set P = { T 1, h 1 2, h2, represents the model's weights.

When the process being modeled is dynamic, i.e., its current output is a function of its current state as well as previous process states, static models are not well suited. For such situations, models which are able to extract both static and temporal process relationships are required. The most common method for creating dynamic neural networks is to simply place dynamic PEs in the input layer of a static MLP [17]. These models have been referred to by many researchers as dynamic neural networks (DNNs). The dynamic PEs






19
can have response characteristics based on a priori process knowledge or contain adaptive memory mechanisms or filters.


2.2.1.2 Time-delay neural network

The most common DNN is called the time-delay neural network (TDNN) [24]. This architecture consists of a MLP where each input PE has an adaptive linear FIR filter, as illustrated in Figure 2.



UP)3





(1W

Figure 2: TDNN input PE connectivity. Formally, the TDNN can be described by

ui (t) j = 1 (h(t), N = di j Vi,j (16) ij 1 (t 1) j E (2, N ] /dnn((t), ) = pdl( (t), NT), iP), (17) where :9 N N x N represents the tapped-delay line operator, and N is the number of taps in the delay line.


2.2.1.3 Gamma neural network

The main disadvantage to most DNNs is that they preprocess the input to extract fixed and known dynamics of the process data rather than learn these dynamics from this data. The TDNN can be considered as an exception to this rule, but here the process must have






20
finite impulse response (FIR) dynamics of known order. The Gamma Neural Network (GNN) [17] represents an important class of dynamic ANN models that is able to learn infinite impulse response (IIR) process dynamics without a priori knowledge about the structure or order of these dynamics.

7 )Z- '(1 -y2


(1 -71)z-1 72
71
UI,/
ui(t)
Figure 3: Gamma memory processing element.


The GNN architecture is conceptually an MLP with adaptive Gamma Filters (GF)

placed at the output of its input and hidden layer PEs. A single GF is illustrated in Figure

3. Formally, the GNN is given by

T ui (t) j = 1
j( T(t), N) = Vi, j (18) -l(t)+(1 -yj) .j(t- 1) j E (2,N T] till Wae phIT hl el h2)+A,

g (h(, t) = Ta(Jz2 J 1(hIlJ((t), N ) + ), I + ', (19) where Jg:9N" x 'NT
where : -+ N"x N represents the GF, N is the number of GF taps in the input layer, and N is the number of GF taps in the first hidden layer. This architecture has been presented without a GF in the second hidden layer, but such a configuration would be a straightforward extension.






21
2.2.1.4 Nonlinear state-space model

The GNN uses the Gamma Filter to represent process dynamics. The GF approximates these dynamics from a Gamma memory kernel basis. The Gamma kernels are able to model an important class of dynamics but may not be the best representation for general process dynamics. An alternative approach to using Gamma kernels is to design the ANN architecture with an explicit state and data flow structure that is capable of learning universal process dynamics. This approach is the goal of the nonlinear state-space model (NLSS) which implements process dynamics directly as a nonlinear state evolution equation and an output observation equation [43], as given by


-(t) = /x(t(t 1), h(t), tx) (20) Y(t) = ((t), -(t), 4y), (21) where k(t) e 9< is the models state vector consisting of Nx hidden PEs, fj is an ANN map describing the time evolution of this state, tx are the weights of this state network, f is a second ANN map describing how outputs are produced from this state, and t are the weights of this output network. Figure 4 illustrates the configuration of a NLSS network.










Figure 4: Nonlinear state space neural network configuration.






22

2.2.2 Learning Algorithms

The biological roots of neural networks are responsible for the widespread use of the term learning to describe the process during which the network parameters are changed to improve the performance of the neural-network-based system. An ANN learning algorithm specifies how its weights are updated in response to training data. These algorithms are simply optimization methods applied to the task of finding the best model weights to minimize a specified modeling objective J, i.e., ArgMin,{J} In general, any of the optimization methods presented above can be used to solve this problem.

One of the most significant breakthroughs in the field of ANNs was the realization that the chain rule for ordered partial derivatives provides a mechanism for deriving the firstorder gradients for all weights in a model, even though the modeling objective is only an explicit function of the model's outputs [72]. Recall that the chain rule for ordered partial derivatives is given by

J J + _" Cxi (22) x ax a8x .ax .
J J ij 1 J

Applying the chain rule allows sensitivities to be calculated from the output of the

model back to its input, which is why the resulting algorithm has been coined "backpropagation" in the literature [40]. When the variables are temporally related, the chain rule has the following form

oJ ( J + a tJ xi(t + ) + (23)
8x(t) x(t) at )8xi(t + -) 8xj(t)

Here, in addition to backpropagating sensitivities from the model's output to its input, the sensitivities are backpropagated through time.






23
The most common optimization method is simply steepest-descent with momentum, although many variations have been demonstrated to significantly improve convergence [10]. The issues leading to the selection of one optimization method over another are:

1) Convergence Rate

2) Implementation Complexity

3) Configuration Complexity

4) Avoidance of Local Minima

5) Sensitivity to Correlation

The most common modeling objective used in ANNs is the mean squared error (MSE) between the model's output ) and a specified desired response d e 91 as given by



J 1 (di(t) -yi(t)) (24) Nt= l i= 1

where Nt is the number of samples in the training dataset. Learning rules which use the MSE criterion are commonly classified as supervised learning rules, because of the presence of a "teacher" implied by the explicit specification of a desired response. Learning rules without explicit reference to a desired response for the model in the objective function are commonly referred to as unsupervised learning rules.


2.2.3 Generalization

As universal mappers ANNs are almost always more complex than the relationships that we seek to uncover. The net result is that ANNs are notorious for over-fitting a training dataset, i.e., performing well on training data but poorly on a blind test dataset [24]. It






24

is very important to optimize the complexity of the neural network in order to achieve the best generalization.


2.2.3.1 Bias and variance

Considerable insight into this phenomenon can be obtained by introducing the concept of the bias-variance trade-off. Bishop [13] observes that the generalization error using the Euclidean norm, will depend on a particular dataset D on which the network was trained. The dependence on D can be eliminated by considering an average over the complete ensemble of datasets, which can be written as = ED[((W )f(, ))2], (25) where (Jh) denotes the conditional average, or regression, of the desired data given by,


(dI) = .-p(-Il)dt, (26) and p(d h) is the conditional density of the desired variable d conditioned on the input vector h Bishop [13] demonstrates that this generalization error can be decomposed into the sum of the bias squared plus the variance

= (ED f( t) h))2 + ED[f(, ) ED[f(, )]]. (27)

A model which is too simple, or too inflexible, will have a large bias, while one which has too much flexibility in relation to the particular dataset will have a large variance. Bias and variance are complementary quantities, and the best generalization is obtained when we have the best compromise between the conflicting requirements of small bias and small variance. The variance of the prediction will be further addressed below, Section

2.2.4 "Standard Errors."






25
For any given dataset, there is some optimal balance between bias and variance which gives the smallest average generalization error. In order to improve the performance of the network further we need to be able to reduce the bias while simultaneously reducing the variance. The more straightforward way of achieving this is to use more data samples. As the number of data samples is increased we can afford to use more complex models, hence reducing the bias, while at the same time ensuring that each model is more heavily constrained by the data., thereby also reducing the variance. If the number of data samples is increased rapidly in relation to the model complexity we can find a sequence of models such that both bias and variance decrease. Models such as ANNs can in principle provide consistent estimators of arbitrary accuracy as the number of data points is increased to infinity. Note that, even if both the bias and variance can be reduced to zero, the generalization error will still be nonzero due to the intrinsic noise in the data.

One rarely has infinite data, and practical issues like training time make simply adding more data points impractical. There are several practical and practiced ways to improve model generalization, we start with regularization.


2.2.3.2 Regularization

Regularization was originally proposed by Tikhanov [62] as a method for solving illposed problems. The basic idea is to stabilize the solution by means of some auxiliary nonnegative functional that embeds prior information, e.g. smoothness constraints on the input/output mapping. Regularization is able to transform an ill-posed problem into a well-posed problem [48].






26

Tikhanov's regularization theory uses a regularization penalty term of the form 12
D = 2 Pf 1,||, (28) where P is a linear (pseudo) differential operator. This penalty term is added to the objective function to give

J = J+ FO, (29) where F is the regularization parameter. Prior information about the form of the solution (i.e., the plant) is embedded in the operator P. The operator P is referred to as a stabilizer in the sense that it stabilizes the solution ', making it smooth.

The appropriate choice for P and the solution to (28) requires functional analysis and is beyond the scope of this work. The most commonly used form of regularizer, however, is quite simple to implement. Weight decay regularizer terms consist of the sum of squares of the adaptive parameters in the network = wi, (30)
I

where the sum runs over the weights and biases. In conventional curve fitting the use of this form of regularizer is called ridge regression. It has been found empirically that a regularizer of this form can lead to significant improvements in generalization [29].


2.2.3.3 Growing and pruning algorithms

The topology of a neural network, number of units and interconnections, can have a significant impact on its performance. Regularization helps to minimize this impact when the complexity of the network is larger than required for the particular application. Clearly, however, a better approach is to match the complexity of the model with the com-






27
plexity of the application. Various techniques have been developed for optimizing the topology, in some cases as part of the network training process itself [43]. It is important to distinguish between two distinct aspects of the topology selection problems. First, we need a systematic procedure for exploring some space of possible architectures. Second, we need some way of deciding which of the architectures considered should be selected.

A straightforward approach to network structure optimization involves an exhaustive search through a restricted class of network topologies. This approach requires significant computational effort and only searches a very restricted class of network topologies. Much of the computational burden can be lessened by considering a network which is relatively small and by allowing new units and connections to be added during training. This approach was shown to be successful by Bello [10] who used the weights from one network as the initial guess for training the next network (with the extra weights initialized randomly). Techniques of this form are called growing algorithms. An alternative approach is to start with a relatively large network and gradually remove units; these are known as pruning algorithms. Most of these procedures are ad hoc and tailor to specific applications, that is not to say, however, that they are ineffective.

More recent work has taken advantage of developments in discrete optimization using genetic algorithms [36]. Genetic algorithms provide a methodical way of searching large discrete spaces more efficiently.


2.2.3.4 Cross-validation

An alternative to regularization as a way of controlling the effective complexity of a network is the procedure of cross-validation [13]. The training of a nonlinear model corresponds to the iterative reduction of the error function defined with respect to the training






28
dataset. During training, the error will generally monotonically decrease as a function of the number of presentations of the training dataset, i.e., epochs. However, the generalization error, with respect to an independent dataset called the validation dataset, often shows a decrease at first, followed by an increase as the network starts to over-fit. Training can therefore be stopped at the point of smallest error with respect to the validation dataset as this produces a network with the smallest generalization error (or at least an approximation thereof).


2.2.3.5 Committees of networks

In practice, building neural network models requires the training of many different

candidate networks and then the selection of the best performer. Typically performance is based on the networks performance on a third dataset not used for training or cross-validation. There are two disadvantages to this approach. First, all of the effort involved in training the remaining networks is wasted, and secondly, the generalization performance on the validation dataset has a random component due to the noise on the data [13]. The network which performed the best on this dataset might not be the one with the lowest generalization error. Recall that the generalization error is averaged over all datasets (25).

These limitations can be overcome by combining the networks together to form a

committee [47][46]. This approach was shown to provide significant improvements in the generalization error. Denote the committee prediction as Nc
(, = 1 y(,wit), (31) Nci= ,






29

where Fi are the weights of committee member i, and { i } is the set of all weights for the committee.

Bishop [13] shows that if the errors of the individual committee member are decorrelated, then the committee will always have a lower generalization error than any of its individual members.


2.2.4 Standard Errors

Tibshirani [61] reviews a number of methods for estimating the standard error of predicted values from a multi-layer perceptron. These include direct evaluation of maximum likelihoods based on the Hessian matrix, the "sandwich" estimator and the bootstrap method. Tibshirani offers the following observations:

1) The bootstrap methods provided the most accurate estimates of the standard errors of predicted values.

2) The non-simulation methods (delta and sandwich) missed the substantial variability due to the random initial weights from the multiple training runs.

The non-simulation methods are solved analytically, and therefore require unique

solutions for each network topology. Whereas the bootstrap methods apply to all network topologies, as well as non-neural network paradigms. The additional fact, as noted above, that the bootstrap methods account for local minima, provides strong argument for their use.


2.2.4.0.1 Bootstrap methods

Bootstrap methods work by creating many pseudo-replicates ("bootstrap datasets") from the training dataset and then reestimating the models weights t on each bootstrap dataset. There are two different approaches to bootstrapping [9]. One can consider each






30
training case as a sampling unit, and sample with replacement from the training dataset cases to create a bootstrap sample. This is often called the "bootstrap pairs" method. The bootstrap pairs sampling algorithm is given by:

I: Generate Nb samples, each one of size Ns drawn with replacement

from the Ns training observations { h (i), (i)}i= 1, and the b -th sam~b ,b N'
ple by { Ab(i),d (i) }I 1

II: For each bootstrap sample b e [1, Nb], find


ArgMinb {Jb d -f( b b )) } (32) III: Estimate the standard error of the i th prediction as


2 -Yi) (33) where

-1
Yi fi(d, ) (34) Nj= 1

On the other hand, one can consider the predictors as fixed, treat the model residuals

-) as the sampling units, and create a bootstrap sample by adding residuals to the model fit )5. This is called the "bootstrap residuals" approach:

I: Find ArgMin0 {J(d -f(k, t)) } from the N' training observations

{(i), (i.)} I= 1 and let r(i) = l(i)-f(h(i), T)






31

II: Generate Nb bootstrap samples, each one of size Ns drawn with
b
replacement from { (i) = 1, and the b th sample by { r (i) } i = i letting

b b '
i = {f(=, ) + (i)}i 1 (35) III: For each bootstrap sample be [1, Nb], find


ArgMin { Jb f( b)) } (36) IV: Estimate the standard error of the i th prediction as Nb
2 1 -2 C (yi) 1 f(, ) -i) (37) Nb 11=
j= 1

Note that both of these methods require fitting a model (retraining the network) Nb times. Typically Nb is in the range 20 Nb ___ 200. In simple linear least squares regression, it can be shown that the bootstrap methods both agree with the standard least squares formula as Nb -> o.

The bootstrap methods will arrive at confidence intervals

-2 -2
Yi CconJ2 (i) Yi < i- CcoJ (Yi) (38) where cconf depends on the desired confidence level 1 Ox. The factor cconf can be taken from a table with the percentage points of the Student's t-distribution with the number of degrees of freedom equal to the number of bootstrap runs Nb.






32
2.3 Control Theory

An ANN is capable of modeling any process making them ideal candidates for complex process optimization and control strategies. Neurocontrol is but a sub-field of classical control theory [31]. To put neurocontrol in perspective, it is important to consider its place within this field.


2.3.1 Classical Control Theory

Classical control theory is strongly biased towards linear time-invariant systems [31]. General nonlinear systems simply do not allow us, because of their analytical intractability, to formulate a theory that is as strong as that of linear system theory. On the other hand, nonlinear systems can be qualitatively similar to linear ones under some circumstances.


2.3.1.1 Linear control

Linear control is concerned with systems of the form

& = +I (39) with state t, input h, measurable output

S= (40) and controllers of the form

= + (41) where i* is the reference state, that is, the state to which the plant is to be brought with the help of the controller [33].

The goals of linear control are:

1) Altering the closed-loop behavior of the system to some user-defined
response characteristics.






33
2) Controlling the closed-loop stability, i.e., convergence back to an equilibrium point after disturbance.

The disadvantages of linear control designs:

1) Assume that the world is linear Gaussian and stationary, when in reality
the world is none of the above.

2) Require complete a priori knowledge of the process dynamics.

3) Require that the process is controllable and observable.

4) Cannot follow a reference trajectory produced by a system of lower
order that the process.

2.3.1.2 Robust control

Robust control addresses the problem of controlling a plant whose behavior is slightly different from that of a plant model [37]. The reasons for the difference are predominately the effect of the nonlinear, non-Gaussian and nonstationary world. A popular pragmatic classical approach to robust control is concerned with preserving stability [2]. The closedloop eigenvalues are chosen so that they remain in the stability region even if the plant model should change in a defined range.

Although robust control strategies are primarily designed to compensate for differences between our linear time-invariant assumptions and the real world, they are still developed based predominately on linear system theory. They are therefore, not able to cope with significant deviations from these assumptions.


2.3.1.3 Adaptive control

Adaptive control is another way to reach a goal similar to that of robust control [37]. Instead of designing robust controllers that work under conditions different from those for which they have been designed, adaptive controllers recognize the difference between the assumption and reality and change to perform better in the new conditions.






34
Adaptation schemes can be based on both a reference model and a cost functional [2]. The approach called model reference adaptive control (MRAC) is, by its name, committed to the former. This approach is based on formulating the rules for computing the direction of change of controller parameters as a function of the difference between the behavior of the closed-loop system and a reference model. Controller parameters can be adapted either directly or via the estimation of plant model parameters.

A more general approach is that of self-tuning regulators (STR) [3] which consists in adaptive estimation of a plant model and applying a formalized controller design method to the plant model. This design method can be based on cost function optimization.

Like robust control, however, adaptive control implementations have been based on linear, or simple nonlinear parametric assumptions, about the process. As a result, adaptive control designs have not demonstrated significant successful with complex real-world processes.


2.3.1.4 Nonlinear control

Nonlinear control theory is concerned with general systems of the form [31]

x = f(&, ) (42) with measurable output

= g(t) (43) and controllers of the form

S= c(t' *). (44) The general formulation of nonlinear control holds promise to overcome all of the limitations of the classical control schemes presented above. The fields track record, however, does not deliver on this promise. The problem is that an analytical solution is known






35
only for a restrictive subclass of nonlinear systems. The difficulties with genuine nonlinear controller designs have typically lead to a linearization approach, also known as gain scheduling.


2.3.1.5 Optimal control

The topic of optimal control theory is to design controllers that are optimized to a certain performance criterion [15]. Classical optimal control has primarily focused on applications where such optimality could be proven analytically. For example, for a linear plant and a quadratic performance criterion the Ricatti controller represents an explicit and global solution [31]. Dynamic optimization provides another example for state evaluation and selection of the optimal action, which can be proven optimal in certain applications. Alternatively, if each state at each sampling period is represented by a node in a directed graph and actions are represented by connecting edges of the subsequent states, then the task can also be transformed to the critical graph problem of graph theory [15].

In its most general form, the optimal control problem can be formulated as an optimization problem [15]. The plant can be generalized as in (42) and the goal is to find a controller described by (44) that minimizes

E[J(.* )], (45) where E[ ] is the mean value over time.


2.3.2 Neurocontrol

Most neurocontrol architectures are either explicit or disguised analogies of classical control design such as optimal control or numerical lyapunov-function-based design methods. It has been argued [14] that it is only the representation of functions by neural






36
networks that defines the field of neurocontrol in the broad sense and differentiates neurocontrol form classical control methods.

The author agrees that the use of nonparametric models does differentiate neurocontrol from classical controls, but would argue that the primary departure from, and extension to the potential of, control theory is neurocontrol's willingness to depart from a requirement for analytic solutions. Neurocontrol designs seek to realize the promise of general nonlinear control replacing analytic optimization methods with numeric ones. Researchers in the field of neural networks are accustomed to working in an intractable world, and have been willing to resolve important questions like stability, robustness and consistency empirically. The result has been, and will continue to be an important extension to control theory.

There are many types of neurocontrol architectures in the literature, each with specific advantages and disadvantages. The following sections review some of the conceptual neurocontrol strategies which have been proposed in the literature.


2.3.2.1 Model-predictive control

When a model is used indirectly and offline the control scheme is usually referred to as model-predictive control (MPC) [37]. In most industrial process control applications a priori knowledge about the process is hard to obtain and black-box models must be used. The offline training phase performs supervised learning to develop an ANN model for the process to be controlled, i.e., the ANN attempts to mimic the process after being exposed to actual process data. This phase can be stated as ArgMin 0 { J(fp -ffff.013 (46)






37

where fY and e are the process outputs to be controlled and inputs to be manipulated, respectively.

At the online control phase, the ANN model cannot be used alone; it must be incorporated with a model-based control scheme [31]. This control scheme is once again an optimization problem, which can be stated as A rgMin { Ju* -f(,P, P)) }, (47) where 9* 0 denotes the desired closed-loop process output. This optimization is performed repeatedly at each time interval during the course of feedback control.


2.3.2.2 Model-inverse control

An ANN can be trained to develop an inverse model of a process [40]. Here, the model's input is the process output, and the model's output is the process input. The offline training phase can be stated as


A rgMin f{ J(P -f(f", 0)) }. (48) Clearly, the inverse model is a steady-state model or the resulting controller would be non-causal. Given a desired process setpoint f*, the appropriate online control signal can be immediately calculated as

e = ff*, 4). (49) Successful applications of inverse modeling are discussed in [40] and [58]. Obviously, an inverse model exists only when the process behaves monotonically as a "feed-forward" function at steady-state. If not, this approach is not applicable.






38

2.3.2.3 Controller modeling

Another simple direct neurocontrol scheme is to use a neural network to model an

existing controller. The input to the existing controller is used as training input to the ANN model, and the controller output serves as the desired response. This approach is similar to the model-inverse control except that the desired response here is not a process but a controller. This approach can be formulated as ArgMinof{J(hc -f(', W)) }, (50) where h c are the decision variables generated by an existing controller in response to the plant states p.

Like a process, a controller is generally dynamic and often comprises integrators or

differentiators. If an algebraic feed-forward network is used to model the existing controller, dynamic information must be explicitly provided as input to the ANN model.

In general, this approach can result in controllers that are faster and/or cheaper than traditional controllers. Using this approach, for example, Pomerleau [50] presented an intriguing application where a neural network was used to replace a human operator, i.e., an existing controller.


2.3.2.4 Model-free direct control

Without an existing controller or process knowledge, controllers have to be adapted or learn the way a human operator learns to control/operate a process for the first time. A model-free neurocontrol design objective can be stated as ArgMinc{Jo* -1 ), P (P, c) (51)






39

where jc is an ANN that is directly controlling the process inputs, and tc are the weights of this network. Notice that the optimization criterion J is only a function of the actual and desired process outputs. This means that the optimization methodology employed must be able to learn Oc without an explicit desired response or even a mathematical linkage to the criterion.

The key feature of this direct adaptive approach is that a process model is neither known in advance nor explicitly developed during control design. This most common learning algorithm for this type of control design is referred to as reinforcement learning. The first work in this area was the "adaptive critic" algorithm proposed by Barto et al. [7]. Such an algorithm can be considered as an approximate version of dynamic programming [73][8], later coined as Neuro-Dynamic Programming [12].

Despite its historical importance and intuitive appeal, model-free adaptive neurocontrol is not appropriate for most real world applications. The plant is most likely out of control during the learning process, and few industrial processes can tolerate the large number of failures required to adapt the controller.


2.3.2.5 Model-reference direct control

From a practical perspective, one would prefer to let failures take place in a simulated environment with a process model rather than in a real plant. Even if failures are not disastrous they can cause substantial losses. The performance of a controller could be evaluated based on a model for the process, rather than the process itself. The training stage of the control design can be given by


ArgMin m{J(p-fn(, I C m)}, (52)






40
and the control design becomes


ArgMin p{Jo* _(e ,m A )), Ic = pc) }. (53)

In the course of modeling the plant, the plant must be operated "normally" instead of being driven out of control. After the modeling stage, the model can be used for controller design. If a process model is already available, an ANN controller can be developed in a simulation in which failures cannot cause any loss but that of computer time. A neural network controller after extensive training in the simulation can then be installed in the actual control system.

Model-Reference direct control schemes have not only proven effective in several studies [41][63], but have also already produced notable economic benefits [60]. These approaches can be used for both off-line control and for on-line adaptation.



2.4 NOx

The Clean Air Act Amendments of 1990 require that electric utilities make significant reductions in nitrogen oxide (NOx) emissions from their fossil-fired power plants. To date, most efforts to reduce NOx emissions have come from expensive hardware retrofits with less than satisfactory performance. Further complicating matters, conditions that decrease NOx formation (lower temperature, excess fuel) result in the formation other polluting compounds, mainly carbon monoxide (CO). Similar emissions reductions are being required in Europe through local and European Economic Community (ECC) initiatives.

Nitrogen monoxide (NO) and nitrogen dioxide (NO2) are by-products of the combustion process of virtually all fossil fuels. Historically, the quantity of these inorganic compounds in the products of combustion was not sufficient to affect boiler performance; their






41
presence was largely ignored. In recent years, oxides of nitrogen have been shown to be key constituents in the complex photochemical oxidant reaction with sunlight to form smog. Today, the emission of NO2 and NO (collectively referred to as NOx) is regulated by the 1990 Clean Air Act Amendments and has become an important consideration in the design of fuel firing equipment.

NOx is formed by two primary mechanisms: thermal NOx and fuel-bound NOx. Thermal NOx formation occurs only at high flame temperatures when dissociated nitrogen from combustion air combines with oxygen atoms to produce oxides of nitrogen such as NO and NO2. The formation of thermal NOx increases exponentially with combustion temperature and increases by a square-root relationship with the presence of oxygen in the combustion zone. Fuel-bound NOx formation is not limited to high temperatures, but is dependent upon the nitrogen content of the fuel. The best way to minimize NOx formation is to reduce flame temperature, reduce excess oxygen, and/or to burn low nitrogen-containing fuels. Conditions that decrease NOx formation (lower temperature, excess fuel) can result in incomplete combustion. These conditions result in the formation other polluting compounds, mainly carbon monoxide (CO).


2.4.1 Reduction

The available NOx reduction technologies can be categorized into one of the following:

Before Combustion: Nitrogen is extracted from the fuel. This is relatively ineffective, since most of the nitrogen in the formation of NOx comes from the air

(containing N2).






42
After Combustion: NOx is chemically reduced before leaving the stack. This process is also expensive, requiring hardware retrofits.


During Combustion: Altering fuel and air flows and introducing them at different

points of the furnace can create several zones with different temperatures and

stoichiometry. These parameters significantly effect the rate of NOx formation.


The following section reviews available NOx reduction strategies and technologies for combustion sources.

Fuel Switching: Fuel-bound NOx formation is most effectively reduced by

switching to a fuel with lower nitrogen content. No. 6 fuel oil or another residual fuel having a relatively high nitrogen content can be replaced with No. 2 fuel oil, another distillate oil or natural gas (which is essentially nitrogen-free) to reduce

NOx emissions.


Flue Gas Recirculation (FGR): Flue gas recirculation involves extraction of some

of the flue gas from the stack, and recirculation with the combustion air supplied to the burners. The process reduces both the oxygen concentrations at the burners and the temperature by diluting the combustion air with flue gas. CO can

become a significant problem here.


Low NOx Burners: Installation of burners especially designed to limit NOx formation can reduce NOx emissions. Higher reduction efficiencies can be

achieved by combining a low NOx burner with FGR. Low NOx burners are

designed to reduce the peak flame temperature by inducing recirculation zones,

staging combustion zones, and reducing local oxygen concentrations.






43
* Derating: Some industrial boilers may be derated to produce a reduced quantity

of steam or hot water. Derating will decrease the flame temperature within the

unit, reducing formation of thermal NOx. Derating can be accomplished by

reducing the firing rate or by installing a permanent restriction, such as an orifice

plate, in the fuel line. Clearly this solution would have significant economic

impact on the unit.


* Steam or Water Injection: By injecting a small amount of water or steam into the

immediate vicinity of the flame, the flame temperature will be lowered and the local oxygen concentration reduced. The result would be to decrease the formation of thermal and fuel-bound NOx. This process generally lowers the combustion efficiency of the unit by one or two percent.


* Staged Combustion: Either air or fuel injection can be staged, creating either a

fuel-rich zone followed by an air-rich zone, or an air-rich zone followed by a

fuel-rich zone. A low NOx burner utilizing staged combustion can be installed,

or the furnace itself can be retrofitted for staged combustion.


* Fuel Reburning: Staged combustion can be achieved through the fuel reburning

process. A Gas Reburning Zone (GRZ) is created above the primary combustion

zone. In the GRZ, additional natural gas is injected, creating a fuel-rich region

where hydrocarbon radicals react with NOx to form molecular nitrogen.


* Reduced Oxygen Concentration: Decreasing excess air reduces the oxygen available in the combustion zone and lengthens the flame, resulting in a lower heat release rate per unit flame volume. NOx emissions are reduced in an approxi-






44
mately linear fashion with decreasing excess air. However, as excess air is reduced beyond a threshold value, combustion efficiency will decrease due to incomplete mixing, and CO emissions will increase. Selective Catalytic Reduction (SCR): Selective catalytic reduction (SCR) is a post-formation NOx control technology that uses a catalyst to facilitate a chemical reaction between NOx and ammonia to produce nitrogen and water. An ammonia/air or ammonia/steam mixture is injected into the exhaust gas, which then passes through a catalyst where NOx is reduced. To optimize the reaction, the temperature of the exhaust gas must be in a certain range when it passes through the catalyst bed. Among its disadvantages, SCR requires additional space for the catalyst and reactor vessel, as well as ammonia storage, distribution, and injection system. Precise control of ammonia injection is critical. An inadequate amount of ammonia can result in unacceptable high NOx emission rates, while excess ammonia can lead to ammonia "slip", or the venting of undesirable ammonia to the atmosphere.


SSelective Non-Catalytic Reduction (SNCR): Selective non-catalytic NOx reduction involves injection of a nitrogenous agent, such as ammonia or urea, into the flue gas. The optimum injection temperature when using ammonia is 1850 degrees F, at which 60 percent NOx removal can be approached. The optimum temperature range is wider when using urea. Below the optimum temperature range, ammonia is formed, and above, NOx emissions actually increase. The success of NOx removal depends not only on the injection temperature, but also






45
on the ability of the agent to mix sufficiently with flue gas.



2.5 Fossil-Fired Power Generation In general Canal Unit 2 is a large fossil fuel combustion engine. From an abstract perspective, the combustion process takes in air and fuel, and produces energy and exhaust; as described by:

1) Air: Fossil fuel combustion requires air, or more specifically the oxygen
contained in air. Subsystems within the plant measure, prepare and
introduce this air.

2) Fuel: Combustion also requires fuel. In the case of Canal Unit 2, the fuel
can be either #6 residual oil (leftover from the refining process) or natural gas. Canal Unit 2 can fire oil only, gas only or a mixture of both.
Both fuels must be measured, prepared and introduced to the furnace.

3) Energy: The energy released by the oxidation of fossil fuels during combustion is used to make steam. The properties of water still make it the
best choice when converting thermal energy to work. Canal uses the
radiative and convective heat from the combustion process to transform ultra-clean water into superheated steam. The expansion of this steam is
used to turn a turbine that turns a coil in a magnetic field, producing
electric potential. The steam, having done this work flows through
ocean water filled condensers that convert it back to super-clean water.

4) Exhaust: The gaseous products of combustion having contributed much
of their heat content to the production of steam are cleaned electrostatically and ejected into the atmosphere.

2.5.1 Process Variables

The specific process variables as they apply to the Canal generating unit are described in more detail in the following sections. These variables are also listed in Appendix, and will be referred to throughout this work.


2.5.1.1 Air

The air required for fossil fuel combustion is prepared and introduced in two ways.

Two large symmetrical Fans called Forced Draft fans push ambient air through a series of






46
preheaters that warm this air to between 80 and 180 degrees F. This hot oxygen rich air is then pressed into a windbox that surrounds the furnace enclosing the burner ports. Through an array of vents called Primary and Secondary Air Shrouds around each burner and through secondary ports called Overfire Air Ports this pressurized air is vented into the combustion zone. In addition to this oxygen rich air, Canal Unit 2 has the ability to recycle exhaust gas into the combustion zone through a Gas Recirculation System.

The measurements of all this air are a function of boiler design, and fan capacities. To increase the output of the engine, additional air must be throttled through these devices.


2.5.1.1.1 Forced draft system

The forced draft fans are 2500 horsepower, 624,000 cfmn centrifugal fans with inlet vane throttles. They are constant speed fans meaning that the fan shaft turns at a constant speed while more or less air with more or less initial spin can be dumped into the blades by opening or closing the vanes. If the vanes are only slightly opened the flow volume of air available to the fans is small, and it takes less work to move it. Fan amps will be correspondingly low. If the inlet vanes are opened wider the flow is greater. Still the fan moves at a constant speed. More work is being done, and the amperage must increase. The output of the FD fans is derived from the boiler master signal. Forced draft output is specified along with fuel flow by the fuel-air curve of the boiler. The fuel-air curve gives a total air flow requirement, as well as a total fuel flow requirement for a given load.

These fans are symmetrical to the furnace like many other systems and they operate symmetrically, through their respective ducts unless biased. Bias represents an addition or subtraction of signal to the B side FD fan. These fans can also be trimmed to meet slightly less or slight more than the Total Air Flow demanded by the fuel-air curve of the Boiler.






47
The FD fans are the principal air throttles of the Boiler and so have a fundamental effect on nearly every other system.


2.5.1.1.2 Forced draft fan inlet vanes

Since the inlet vanes' positions represent the work being done by the fan and are the control most familiar to the operators, these tags were used to represent the FD fans.


2.5.1.1.3 02 trim

This tag represents the bias that operators set into the airflow demand predetermined by load. Functionally this control trims the response of the FD Fan to Air Demand. This tag gives the operators the ability to run the furnace slightly lean or rich overall.


2.5.1.1.4 Induced draft system

As mentioned in the section on the Forced Draft Fans, the function of the ID Fans is to take whatever gasses are present in the furnace, including air that has been introduced by the load following Forced Draft Fans, plus all products of combustion, and pull them out, maintaining a constant under pressure in the furnace of -.5 inches of water column. The FD Fans' speed is kept constant while the volume of air they move is throttled with inlet vane controls. Canal is limited by the power of these fans. Current unit maximum output is frequently limited by the power of these fans to keep up with the increased air flows of the recently installed low NOx shroud and overfire air system.


2.5.1.1.5 Induce draft fan inlet vanes

The induced draft fan inlet vanes are the inlet throttles to the fans, they open in

response to request for increased output and as in the case of the FD fans, represent the work being done.






48
2.5.1.1.6 Combustion air temps

The combustion air temperature tags represent the temperature of the incoming air

after the FD Fans. The temperature of this air is a direct result of energy added to ambient air by the Glycol Air Preheater (GAH), and the Combustion Air Preheater (CAH). Since density is a function of temperature, the temperature of this air can impact the combustion process that is sensitive to the Oxygen content of air as well as the operation of other volumetric systems like the Induced Draft (ID) Fans. It also has a primary impact on exhaust gas temperature and resultant stack gas velocity.


2.5.1.1.7 Primary air shrouds

The Primary Air (PA) Shrouds represent the circular articulating vents that surround the individual burner orifices. These are closest to the fuel gun concentrically inside of the Secondary Air (SA) Shrouds. They are responsible for supplying primary combustion air to the flame front. These tags represent actuator positions.

The PA shrouds are controlled by the Burner Management System (BMS) and they

move as a group from minimum position (5% open to provide cooling air) toward open as load increases. The signal that controls them is called the Primary Air Master Demand (PAMD). Separate PAMD signals exist for fuel gas primary air demand and for fuel oil primary air demand. Each specific burner effectively listens to the current fuel state. Having received this signal each burner's own PA shrouds responds to the PAMD in accordance with one of two functions that are unique to it a Burner Primary Air Shroud Function for oil operation and a Burner Primary Air Shroud Function for gas operation. The correct unique local shroud function is changed according to the correct master signal depending on the fuel state of the burner. These burner and fuel specific response func-






49

tions were set up to give roughly appropriate air flow to combustion at all load points and fuel states based on the air flow inherent to the furnace.

Aside from normal operation the PA shrouds can be biased from the fuel specific master signal or on an individual basis from their respective unique functions.


2.5.1.1.8 Secondary air shrouds

The secondary air shroud tags represent the broadcast actuator positions of the second, outer concentric set of circular articulating vents that surround the individual burner orifices.

The first function of the Secondary Air (SA) Shrouds is to introduce combustion air to the flame front following load. Their second function is to balance windbox pressure, and therefore total airflow, against the actuation of the Overfire Air Ports and the PA shrouds.

The SA Shrouds have a master signal against which a master bias can be set. In addition they have individual actuating functions and individual biases that can be set against these individual functions.


2.5.1.1.9 Over fire air ports

The Overfire Air (OFA) ports are rectangular louvered ports that pass combustion air from the Windbox to the Furnace above the top burner level. In doing this they re-oxygenate the oxygen depleted flame front. The tags themselves represent the positions broadcast from the actuators that control the articulating louvers.

The OFA ports were installed as a part of the low NOx retrofit of 1996. The Forney low NOx burner system is designed to burn more coolly and incompletely than normal. NOx formation has been positively linked with time exposure to higher temperatures. After partial combustion has taken place, low in the flame front, extra oxygen rich com-






50
bustion air is introduced through the OA ports to complete the process. In this way the low NOx burner system stages off-stoichiometric combustion to manage combustion products.

The OA port actuators receive their master signal from load. This signal can be biased. Each actuator's response is based on a unique function that was parametrically determined, in concert with the Primary and Secondary Air Shrouds during installation to give best airflow to combustion at all load points.


2.5.1.1.10 Air preheater temps

These represent the temperature of the exhaust gasses entering and leaving the ljungstrom combustion air heat exchanger. The ljungstrom is a large (30 ft. dia.) rotating wheel, arranged perpendicular to the gas flow. It is half enclosed by the exhaust ducts and half enclosed by fresh air ducts. As this wheel slowly rotates, heat is absorbed by a given area of the wheel exposed to exhaust gas. The absorbed heat is then imparted to the incoming air while that same section traverses the fresh air duct. Elaborate seals and pressurized sealing air keep the two gasses from mingling across this device.

The air preheater tags are somewhat redundant. The "In Temps" represent the temperature of the gas on its way in, while the "Out Temps" represent the temperature of the gas on the way out. The heat exchange of the air preheater is a function of the device and of the temperatures of the two gasses and is not controllable in the least. The gas temp after the air pre heater heat exchange was a more familiar control to the operators, however our ability to collect these signals was compromised by a failing thermocouple during a large part of the data collection for phasel. The gas temp before the air preheater was used to represent exit gas temp for the modeling instead.






51
2.5.1.1.11 Windbox and furnace

The windbox is an enclosed volume that surrounds the waist of the furnace and the burner openings. Preheated, oxygen rich air is pressurized in this volume by the FD fans. From here this air can pass only into the furnace and only through vanes that surround the burner openings called primary and secondary sir shrouds, or through the overfire air ports above the burners. Canal Unit 2 is a balanced draft furnace which means that air flow through the furnace is controlled around a desired furnace pressure by both pushing and pulling fan systems. The pushing fans are the FD fans, while the pulling fans are the induced draft fans. The FD fans have the primary responsibility of getting the combustion zone all the oxygen it requires. The introduction of this pressurized air is accomplished not only by positively pressurizing the windbox but also by negatively pressurizing the furnace. With the windbox driven to a positive pressure and the furnace kept at a fixed relative negative pressure, the velocity of combustion airflow is assured. The induced draft fans have primary responsibility for maintaining the furnace at a negative pressure relative to the windbox. In the course of increasing unit output the FD fans increase air flow. Their aim is to maintain windbox pressure at +2 inches of water column while air transfer to the furnace increases through the widening overfire ports and primary and secondary air shrouds. The induced draft fans, trying to maintain a constant pressure of -.5 inches of water column in the furnace despite this increasing flow of air from the windbox, also ramp up. The opposite happens for decreasing load. When the forced draft fans decrease their output in step with the fuel-air demand, air flow from the windbox to the furnace decreases. In order to maintain a constant -.5 inwc in the furnace the induced draft fans






52

throttled back. Transient changes in the windbox to furnace pressure differential can also produce automated changes in the FD and ID fan flows.


2.5.1.1.12 WindBox pressure

This tags represents the positional average windbox air pressure. It is controlled around + 2 inwc


2.5.1.1.13 Furnace pressure

This tag represents the actual furnace air pressure.


2.5.1.2 Fuel

The Fuel required for Combustion may be either #6 Fuel Oil or Natural Gas. In both cases the fuel is taken from storage, filtered, heated to greater or lesser degree, pressurized, and injected. In the case of #6 Fuel Oil, the temperature required to achieve a pumpable consistency is usually around 200 degrees. Natural Gas comes from high pressure transmission lines and once stepped down to usable pressure is warmed up to around 80 degrees F. Both fuels are then pressurized in their respective headers. It is from these headers that burners, when they are lit, tap their fuel.


2.5.1.2.1 Burners-on/fuel

These tags represent the readings of an array of air cooled optical flame scanners

located in the furnace itself that observe the respective burner flames. Since each burner can fire either natural gas or fuel oil. A scanner, calibrated for each fuel specific flame is permanently assigned to each burner. Although these scanners are analog devices, their primary function is to confirm that the flame emanating from each lit burner is of a threshold quality. If the flame they are monitoring is not of a threshold quality the scanner has






53
the will to declare a Master Fuel Trip and cut off all fuel to the furnace. This is to prevent the introduction of unburned fuel to the furnace. These are analog devices but because they are calibrated with the single purpose of either positively or negatively confirming this threshold they essentially read either 1 or 0. This specific set represents the flame quality of its burner if that burner is on natural gas. 2.5.1.2.2 Burner cells 1-8A & 1-8B MN gas flame

These are the signals for gas flame status for each burner.


2.5.1.2.3 Burner cells 1-8A & 1-8B MN oil flame

These are the signals for oil flame status for each burner.


2.5.1.2.4 Fuel type

As the Boiler Master request increased output BTUs are requested from the Fuel Supply Systems. As a default this request is divided evenly in proportion to burners in service, each of which have BTU content per unit of fuel settings. The total BTUs entering the furnace via the burners in service must equal this demand.


2.5.1.2.5 Fuel oil

The fuel oil introduction system consists of a main pressure generating pump that

ramps up in output as the unit master demand requests more output in the form of BTUs. This pump supplies an operating pressure to the fuel oil header. All oil burners once they are lit and placed into service tap a fixed orifice from this header. Since fuel oil pressure is fixed by the number of BTUs requested by load, and the orifice of each burner tip is a fixed diameter if open, the number of burners in service will dramatically affect Fuel Oil






54
Pressure. Changes in the number of burners lit can vary the fuel oil pressure in the header between 65 and 150 PSIG.

Fuel Temp Fired must be at least the temp required for pumpability, which is specific to the viscosity of the fuel oil being used.


2.5.1.2.6 Natural gas

In a fashion similar to the fuel oil introduction system, the unit master demand

requests BTUs from the gas system. Fuel gas from the pipeline is stepped down to operating pressure, filtered, warmed and supplied to a main gas header. All gas burners when lit tap a fixed orifice from this header. The number of burners lit on gas can affect the actual gas pressure indicated at the header.


2.5.1.2.7 Burner atomization

These tags represent the essential fuel oil atomizing steam parameters. Atomizing

steam is dry superheated steam extracted from the turbine or the reboiler and injected into the oil burner tips to atomize the fuel oil as it is introduced to the combustion zone.

Burner Atomizing Steam pressure runs at a specified 20psig over fuel oil pressure. Burner atomizing steam flow is modulated to maintain this constant difference from fuel oil pressure while the actual temperature fluctuates somewhat at the point of extraction. pv = nrt connects these three variables with temperature being somewhat variable, flow being the control, and pressure being the set point.


2.5.1.2.8 Fuel oil / fuel gas flow differential

This tag represents the ratio of BTUs contributed by the fuel oil system vs. the BTUs contributed by the fuel gas system to the total BTUs required for a given load.






55
2.5.1.2.9 Energy

During operation at Canal Unit 2 feedwater, pressurized by a large parasitic turbine

driven pump, is circulated through series of preheaters and then through the very walls of the furnace. During this passage it is converted to steam. This steam is then collected in a pressure vessel called a Steam Drum located at the top of the boiler where it is "dried". From the Steam Drum this dry saturated steam is passed through radiator like Primary Superheater and Secondary Superheaters that hang at the top of the furnace where convective, and radiative heat transfer occurs. From the outlet of the Secondary Superheater the steam goes directly to the High Pressure inlet of the Turbine. Unit 2 is a single reheat boiler which means that the exhaust from the high pressure turbine, instead of being condensed, is passed back to the boiler and re superheated. This re superheated steam then turns the Intermediate and Low Turbine Stages. Attemperating sprays inject cool feedwater into the steam cycle between the Primary and Secondary Superheaters and also before the Reheat Superheater. These cooling sprays dampen thermal dynamics and keep steam temperature at the turbine roughly constant around 1000 degrees.


2.5.1.2.10 Generation

The Westinghouse turbine generator converts the expansion energy of superheated

steam to create rotational momentum in the turbine. This rotational energy is imparted to a coil enclosed in an induced electromagnetic field. The rotation of this coil in this excited field creates electric potential at the ends of the coil. This electric differential has roughly 560 megawatts of power with which to do work. Under normal operating conditions, and aside from throttling effects, the output of the turbine generator is in direct relationship to boiler output.






56

This tag represents the actual instantaneous unit output in units of power.


2.5.1.2.11 Heat rate

This is a simple calculated tag representing the sum of BTUs flowing into combustion from oil and gas combined divided by the amount of power created. It can show the relative efficiency of combustion-steam-power system in an energy in vs. energy out relationship. As load increases heat rate decreases due to the thermal properties of the steam loop.


2.5.1.2.12 Main steam

The main steam temperature, in concert with the throttle pressure is related via steam tables to volume, enthalpy and entropy and describes the output state of the steam generating system. Unit 2 is a sliding throttle unit capable of modulated steam temp output across different throttle valve configurations. Steam output is essentially controlled by flow. As the unit ramps up in load, more steam is generated from increased combustion. Steam temperature is held (roughly) steady via modulation of flow through the turbine throttle valves, which are sequentially opened. Once the unit reaches a certain level of output (@480MW) all throttle valves are set in the fully open position and steam flow is modulated by continuing to increasing steam output through combustion throttling. At all levels of output Steam temperature is controlled around 1000 degrees F for optimum turbine operation


2.5.1.2.13 Temperature

These tags represent the temperature of superheated steam as it exits the secondary superheater header and heads to the high pressure turbine throttle valves.






57
2.5.1.2.14 Attemperation spray

These represent the amount of cool feedwater that is sprayed into main steam between the primary and secondary superheaters to control the temperature of the steam at the secondary superheater outlet to the turbine.

U28300 represents fine control. This valve responds automatically and in analog fashion to all changes in steam temperature at the secondary superheater outlet. U28301 represents bulk control. It responds only to changes in SSH outlet temp that are exceed preset deadband. These coarse and fine cooling controls are combined to dampen and control steam outlet temp against oscillations or imbalances inherent in the steam system.

All desuperheating sprays receive their volume of feedwater from total feedwater flow.


2.5.1.2.15 Reheat steam

Exhaust from the high pressure turbine stage is cycled back to the furnace via the reheat steam loop where it is sprayed then re-introduced to heat exchange in the reheat superheater. Through the reheat superheater this steam is brought back up to 1 000degF and 580pisa upon which it is sent to the intermediate stage of the turbine. Exhaust from the intermediate stage turbine flows to the low pressure turbine stage.


2.5.1.2.16 Temperature

This temperature represent the temperature of re superheated steam as it heads to the intermediate turbine stage inlet.


2.5.1.2.17 Attemperation sprays

These sprays function like the superheater sprays. They inject relatively cool feedwater into the reheat steam after it has been extracted from the turbine and before it is






58

reheated. They function to control the temperature of the steam at the outlet of the reheat superheater. Unlike the superheat desuperheaters, these sprays do not have separate fine and coarse control functions.


2.5.1.2.18 Furnace metal temps

These tags represent an array of thermocouples installed on the vertical legs of the

pendant superheaters. Especially in gas burning the fire side material temperature of these heat exchangers can become problematic. Unit 2 has an especially large area of superheater, which is the heat exchange closest to the fire itself. Because gas burns at a cooler temperature than oil less radiant heat is absorbed by the waterwalls of the furnace and for the same output of steam more heat must be passed to the steam loop through the gas stream and the superheaters. This superheater weighted heat transfer zone in gas burning, combined with air flow stratification that seems to be inherent to this unit, make careful monitoring of these thermocouples necessary. Extended temps above 1100 degrees can increase material fatigue signifigantly.


2.5.1.2.19 Secondary superheater metal temps top-bottom L-R

These represent the temperature of the firesides of selected evenly spaced legs of the secondary superheater, which encounters hot gas second, after the primary superheater. They are alphabetized horizontally across the superheater surface with upper representing the trailing side and lower representing the leading side.






59
2.5.1.2.20 Primary superheater metal temps top-bottom L-R

These represent the temperature of the firesides of selected evenly spaced legs of the primary superheater, which encounters hot gas first and is closest to the flame front. They are alphabetized horizontally across the superheater surface.


2.5.1.2.21 Reheat superheater metal temps top-bottom L-R

These represent the temperature of the firesides of selected evenly spaced legs of the reheat superheater, which encounters hot gas third, after the secondary superheater and before the feedwater economizer. They are alphabetized horizontally across the superheater surface with upper representing the trailing side and lower representing the leading side.


2.5.1.2.22 Exhaust

The gasses created by combustion flow upward through the furnace gas path across the primary and secondary superheaters, the reheat superheater, and a superheater- like feedwater preheater called an economizer. In this pass all steam loop heat transfer occurs. After leaving the furnace these exhaust gasses flow into a ljungstrom air heat exchanger where heat is traded to the incoming combustion air. Under the pull of the induced draft fans this now 350 degree gas passes through the units robust electrostatic precipitator array, through the induced draft fans themselves and then up the stack.

The makeup of the fluegas at the point it leaves the furnace represents the overall quality of combustion. Key parameters include how much oxygen has been left by the combustion process, and how much CO has been created. The richness or leanness of combustion is directly evident.






60

2.5.1.2.22.1 Flue gas

2.5.1.2.22.1.1 CO

These tags represent the CO contained in exhaust gasses as measured in the side A (U27814) and side B (U27813) furnace outlets to the exhaust ducts just after the economizer.

These are point measures of CO in a very large duct and may not capture exact CO content. They also display extreme side to side bias with side B showing higher CO content. Although peculiar, this side to side bias is believed to be a real feature of the Canal Unit 2 furnace draft. These tags are directly related to the quality of combustion and can serve as a non delayed approximation of CO as it will be seen at the stack.


2.5.1.2.22.1.2 02

These tags represent the 02 contained in exhaust gasses as measured in the side A and side B furnace outlets to the exhaust ducts just after the economizer. These tags are used in modeling to represent the richness or leanness of combustion. They are impacted by and can be used as a control reference for forced draft fan output trim on air demand. In addition these tags are used by Canal as a part of the CEM NOx calculation.


2.5.1.2.22.1.3 Temps

The temperature of the air being forced through the boiler at Canal Unit 2 impacts and represents many process parameters, from combustion quality, to heat transfer distribution, to induced draft fan output. It also is control reference for the temperature and velocity of exhaust leaving the stack.






61
2.5.1.2.22.2 Stack

The CEM (Continuous Emissions Monitoring Unit) consists of an array of extraction gas analyzers in a computer room at the base of Canal's 500 ft. stack. The pitots of these analyzers sniff mixed exhaust from the top of the 18 foot wide Unit 2 flue. The specific amounts of certain compounds measured in this gas are entered into a database. This database serves as a binding legal history of Canals environmental compliance. Each violation of emissions limits placed on certain compounds like NOx and CO is recorded. If the unit is in danger of breaking its allowed daily average output of these regulated pollutants, measured from midnight to midnight, all steps must be taken to regain compliance, including dropping load. The cost of such a sacrifice is immense and in effect these hourly and daily emissions limits have become control variables of primary importance.


2.5.1.2.22.2.1 CO

This tag represents the CO content of stack gas in parts per million.

It is worth noting that CO and NOx represent conflicting states of combustion as they are currently understood and managed. To reduce NOx production, combustion is kept cool and rich. NOx formation has been shown to positively relate to increased exposure to combustion and increased temperature. Over fire air is used to complete this off- stoichiometric combustion. Unfortunately such rich and cool (incomplete) combustion inherently produces increased CO.


2.5.1.2.22.2.2 NOx

This tag is calculated using a regulatory approved method and is used to represents the pounds of NOx produced by Canal Unit 2 per million BTUs.






62
2.5.1.2.22.2.3 Temp

Stack temp is important to Canal for several reasons. Keeping stack temperature at a certain point guarantees that no condensation of sulphur products can occur in the exhaust ducts, precipitators, or in the stack itself. The products of sulphur condensation are acidic and over extended periods of time can be damaging to expensive capital equipment. As long as sulphur emissions are within limits, and they are not a problem at Canal, since Canal uses low sulphur fuel oil, it is beneficial to push them all the way out of the stack before they can condense. This requires sufficient stack gas temperatures and velocities.

Stack Temperature is controlled primarily by the amount of preheating that is done to the air before it even enters the windbox. Because of the relationship of final stack gas temperature to combustion air temperature, and the relationship of combustion air temperature to other properties of combustion, stack temperature can be an important and tricky control point. Since the heat losses to the exhaust through the exhaust ducts, precipitators and induced draft fans are fixed, fluegas temperature also represents stack temperature but without gas path travel delay.














CHAPTER 3
BOILER OPTIMIZATION

The most efficient method for reducing NOx emissions is clearly during the combustion process [39]. As presented in Section 2.4 "NOx," simply changing the combustion temperature and fuel/air distribution can dramatically affect NOx emissions. The combustion of fossil fuels inside a large-scale boiler, however, is a highly complex process; this complexity is a direct function of the boiler size. A typical electric power boiler maintains a "fireball" which is 3 to 5 stories tall, and there are hundreds of parameters which affect the injection of fuel and air at different locations within the furnace.

The problem is our lack of understanding about how these combustion parameters

affect NOx formation. This multivariate optimization problem requires a technology that can look at the process globally and determine the appropriate combination of combustion controls.



3.1 First Principles

The concepts behind boiler optimization are relatively simple:

If the boiler operates in an oxygen-rich environment, i.e., with unnecessary

excess air, boiler efficiency will decrease due to the loss of sensible heat up the

stack; NOx emissions will increase concurrently.


If the boiler operates in a fuel-rich environment, i.e., with insufficient air, boiler

efficiency will decrease due to the loss of unburned fuel. In addition, insufficient


63






64

air leads to CO formation which causes slagging and water wall corrosion, ultimately shortening boiler life.


Between these two airflow conditions there is a zone of optimum combustion. This is shown as a dark gray band in Figure 5.



3.2 Fuel and Air Distribution

Boilers for electric power and industrial steam typically have poor distribution of fuel and air within the furnace. This causes some regions of the firebox to be fuel-rich and other regions to be oxygen-rich. This situation is clearly undesirable as it leads not only to unnecessary NOx production and reduced efficiency, but reduced boiler life expectancy due to water wall corrosion and slagging.

The variability of the fuel-air ratios at different locations throughout the boiler is represented as a light gray band in Figure 5. This variability determines the amount of aggregate air required to ensure that all regions inside the boiler avoid fuel-rich combustion.


. .. . .. . . . ................. . !". .. .. .. . ... z - O f . . . . .. . . . ..
Zone of
Fuel-rich environment Combustion Oxygen-rich environment ....... ... .. . . . .. ....... ... ... .. ........ . ... .. ... .... .. . . . . ...........

i ... .. ... ... .. ... ... .. ... ... .. ..... .. .. .

CO Efficiency
. ...... ......... . ..... -----Unburnedi
........... uel N
... ............. .. . .. .....


Excess Fuel Increasing Air Excess Air ......... ...~ ~~ 1] 711 _...... .- .-.. .... ......i.... . ....
Aggregate Air
Flow
Figure 5: Combustion emissions characteristic versus air flow.






65

By improving the distribution of fuel and air in all parts of the firebox, it is possible to reduce the aggregate airflow while maintaining the same safety margin. This improvement is illustrated in Figure 6. The narrower darkly-shaded band which represents the improved distribution of air and fuel moves closer to the zone of optimum combustion. Reducing the aggregate airflow simultaneously increases boiler efficiency and reduces NOx emissions.


Zone of ..... .........
Optimum
Fuel-rich enironmen Combustion Oxygen-rich environment . ... ......... i ciency

Efficiency
Unburned CO
Fuel NOx .... ........ .Nox reduced
............... ....... .. ............... .. .. . ... ..... ... ....... .... . ..............

!' ~ ~ i.............. . ..... ... ... ,i ..... ... .... ....... .............. . .............. .........E ; i r
Excess Fuel Air Excess Air Aggregate Air
Flow Reduced
Figure 6: Effect of lower 02 on combustion emissions.


The key challenge in boiler optimization is identifying which of the many controls affect performance and how they need to be manipulated to ensure optimal performance as process and economic conditions change.



3.3 Boiler Tuning

Boiler manufacturers and service companies offer boiler-tuning methodologies that use the above principles of combustion to identify a limited set of control settings which help lower NOx and increase efficiency without the need for substantial capital expenditure. Such boiler tuning improves unit performance but does not begin to generate the savings achievable through improved control.






66
Unfortunately, the number of control variables available to optimize performance is too large for offline boiler tuning to predict the optimum settings. Optimum settings vary with load, fuel quality, boiler conditions, weather, and other factors making offline tuning difficult if not impossible.



3.4 The Role of CO

Figure 5 and 6 both show an exponential rise in CO as excess air is reduced and the boiler approaches peak efficiency. The steepness of the CO curve depends upon the degree of mixing of fuel and air within the furnace. Poor mixing broadens the CO curve by creating pockets of fuel-rich and oxygen-rich combustion. Together with 02, CO levels provide the best indication about combustion quality.

A model for CO will provide valuable information about:

how well mixed the fuel and air are in the furnace,


how individual setpoints can be used to improve this mixing, and


conditions which lead to slagging and water-wall corrosion.


The CO measurement serves as a key safety constraint when optimizing the boiler. By controlling to CO levels, the boiler can be optimized without compromising safety margins. Improved air and fuel distribution will merely tighten the CO curve, resulting in improved efficiency and lower NOx.














CHAPTER 4
CONTROL DESIGNS

This research investigates the applicability of neurocontrol techniques to complex process control problems, and develops a methodology for implementing them. Towards this end, this work will develop several detailed neural network-based control designs and apply them to the reduction of NOx and the maintenance of acceptable CO levels in electric power plants. Subsequent sections implementation these control designs and use our NOx case study to compare and contrast them. The control methodology will be presented as follows:

1) A methodology for categorizing key process variables into groups that
are required for all control designs.

2) A methodology for formally stating the control optimization objectives
and operating constraints using the aforementioned variable definitions.

3) Performance criteria by which the various control designs will be
judged, based on these formal objectives and constraints.

4) Four formal control designs with explicitly account for state variable
dependencies.


4.1 Variable Definitions

When designing a controller for large-scale industrial processes, there are a large number of variables to be considered. The physical processes are typically considered to have inputs, disturbances, states and outputs. The following variable definitions are proposed as a methodology for categorizing all process variables into subsets; these subsets will prove useful when designing controllers in general:


67






68

1) Manipulated Variables (MVs): process inputs which have been selected
for our controller to manipulate. The MVs should be independent of one another, i.e., manipulating one will not cause a change in any of the others.

2) Disturbance Variables (DVs): process inputs or disturbances that affect
the state or output of the process, but we either cannot or have chosen not to manipulate. The DV should be independent of both each other
and the MVs.

3) Control Variables (CVs): the process state or output variables that the
controller will be designed to control. The CVs should be a function of
the MVs and DVs or there is little hope of the controller being able to
control them.

4) State Variables (SVs): process state variables, which are a function of
the MVs and/or DVs, that affect the CVs. Alternatively, the SVs may be process output variables that have not been selected for control but need
to be considered as constraints.

Notice that the MV, DV, SV and CV definitions categorize the process logically and not physically. These definitions divide variables based on how the controller will be configured, rather than how the physical process is configured. The MVs will always be process inputs, i.e., can be manipulated by operators, but the DVs can contain both process inputs and disturbances depending on which inputs are being manipulated. Likewise, SVs and CVs can each consist of any combination of process states and/or outputs, based on which will ultimately be controlled.

Notation: The categorization of variables into CVs, SVs, DVs and MVs

will be used extensively throughout this work, and is conceptually consistent with the literature on optimization and control [37].



4.2 Optimization Objectives

The control objective is to lower NOx. Formally this objective needs to be stated as an objective function for optimization. Since several of the controllers developed here are tra-






69
jectory (multi-stage) controllers, this objective function will be a function of time. Consider the single control variable NOx(to) e 9 as the measured value of NOx at time to. An optimal control objective with fixed terminal time T for minimizing NOx(t) over the interval t e (to, to + T] can be given by to+ T
J = NOx(t). (54) t = to

In general, there will be more than one CV. If all CVs are to have equal impact on this objective function, then two effects will have to be removed from the optimization objective: 1) the effect of power differences between these CVs, and 2) the current value of each CV. The following objective function extends (54) to multiple control variables

1 to + T N
J TC Pi(Z(cvi(t)) Z(cvi(to))), (55) I= toi= 1

where Ncv is the number of CVs, and pi e 91 is a priority weighting factor and Z(x) = (x tx)/cx is the z-score statistic [66]. Assuming that our controller is designed to minimize J; for pi > 0 the CV cvi will be minimized over the trajectory, while setting pi < 0 will maximize the output.

Equation (55) considers the case where CVs are to be maximized or minimized. In

general, the goal is to design a controller capable of maintaining a control setpoint. A generalized optimization objective is therefore presented as to+ T NcV

TC PiD (56)
t=toi= 1






70

where Di is a desirability function that can be tailored for each CV to Diax = Z(cvi(t)) Z(cvi(to)), (57) Di = Z(cvi(to)) Z(cvi(t)), or (58) Dp = (Z(cvi(t)) Z(sp(t))) (59)


4.3 Operating Constraints

Constraints will be used to ensure that the optimizer produces a feasible solution. By feasible we mean: 1) the MV moves can be made, and 2) that when these MVs are applied the plant will end up in a desirable state. Feasible solutions will be guaranteed by designing controllers which are able to maintain MV and SV constraints.


4.3.1 Manipulated Variable Constraints

To ensure that the MV moves can be made, the controllers will maintain simple range constraints. A range constraint consists of the upper and lower limits that an MV will be allowed to move. Formally the range constraint for MV mvi will be given by C V = [C C' ], (60) where Cn is the MV's absolute minimum and Cipax is its maximum. Controllers will be to +T-1
required to provide an optimal MV trajectory { mv (t) } t = to such that mvi*(t) e- C'v Vi, t. (61)






71
4.3.2 State Variable Constraints

Similarly, to ensure that controllers drive the plant to a desirable state, SV constraints will also have to be addressed. Formally, controllers will be required to provide optimal to+ T
MV trajectories that result in SV trajectories {fsv*(t)}t = to0+ 1 such that svi*(t) E Vi, t. (62)

4.3.3 Penalty Functions

Each control design considered will employ an optimization algorithm during some phase of its development. Some optimization algorithms are able to deal with constraints directly, i.e., given knowledge of the constraints they can ensure a feasible solutions. Others, however, will have to treat constraints indirectly by addressing them with the objective function. The most common method for addressing operating constraints in an objective function are through the use of penalty functions [54]. For example, SV constraints can be stated as penalty functions of the form {" 2ax 2 max
(svi- Ca" ) svi> Cax
X(svi, CV) = 2 in (63) (svi C ) svi K 0 else

Generalizing the penalty function to multiple SV constraints, differences in the energy of the respective signals will once again have to be normalized out. These effects can be compensated for using a generalized penalty function of the form (Z(svi) Z(Cmax))2 sVi > nax ,(svi, iv) = (Z(sI Z(cmin))2 svi < cin (64) 0(Z(sv) Z(Celse)) i
L0 else






72

Given a set of N"v SV constraints, an optimizer may satisfy these constraints by appending their respective penalty functions to its criterion J#v
S J+ p "(svi, C'V) (65) i=1

where psv allows constraints to be individually prioritized. Similarly, both MV constraints can be appended to the optimizer's criterion by defining the penalty functions {k2(mvi nv) IV

Note that implementing constraints with penalty functions will not guarantee that the constraints are met precisely. If the constraints are properly prioritized relative to the optimization objectives, however, these constraints are easily maintained within the desired level of accuracy.



4.4 Performance Criteria

For the case study, controllers will be judged based on their ability to lower NOx while maintaining desired CO emissions. To this end, subsequent sections will measure the performance of controllers as a plant operator moves MVs according to their control laws. Comparing controller performance, however, will prove a difficult task, since the operator can only take the advice from one controller at a time and the plant is constantly changing state. Although the controllers may be able to deal with non-steady-state conditions, it will be nearly impossible to separate the process responses to the state changes versus the control action.






73
Further complicating matters, while one-time tests will provide useful results with which to judge the controllers, they are not the only criteria. The controllers studied will be judged by the following criteria:

1) Ability to control NOx and CO.

2) Ability of the operators to perform the recommended MV moves

3) Flexibility with respect to changing performance objectives and operating constraints

4) Ability to deal with changing operating states, e.g. load changes


4.5 Controller Designs

Four controller designs will be developed. The controller designs considered, fall into the broad categories of:

1) Model-Predictive Control

2) Model-Inverse Control

3) Model-Based Direct Control

There are, however, no standard recipes for building these controllers. The field is still immature, and neurocontrol designs presented in the literature tend to be ad hoc. This work seeks to not only develop and test four neurocontrol designs, but also to develop a generalized methodology for implementing control designs belonging to the above abstract categories. Each controller must be able to deal with the MV and SV constraints, and will be judged by the performance criteria described above.

For each of the control designs considered, there are two distinct phases in the implementation:

1) Offline training.

2) Online control.






74
4.5.1 Steady-State Optimizer

The simplest, and most prevalent, neurocontroller in the literature is the steady-state optimizer [43][40][3 1]. This controller belongs to the model-predictive control family. Model-predictive control (MPC) is not new to commercial applications in the process control industry. The advance proposed here is the application of neural network reference models within this controls methodology.

The concept of MPC is straight forward: combine a model for the process with an optimizer to obtain real-time optimal setpoints. Model predictive controllers can be steadystate or dynamic, depending on characteristics of their underlying process models. This section details the design of a neural network-based steady-state MPC controller to meet the problem specifications presented in Sections 4.2, 4.3 and 4.4.


4.5.1.1 Offline training

Training a MPC controller follows the schematic outlined in Figure 7. Notice that

there are actually two reference models being trained: one SV model and one CV model. The details for how to train these models will be covered in Chapter 6. Notice, however, that the criterion J" is a model training criterion to be presented in Chapter 6, and not the control performance objective I presented in Section 4.4 "Performance Criteria."

The model definitions required by the steady-state optimizer are:

1) Steady-State SV Model: sv = ssSVModel(mv, dv)

2 Stead-State CV Model: c = ssCVModel(m. dv. sv
2) Steady-State CV Model: cv =ssCVModel(mv, dv, sv)






75



ssS/odel

dv sv

Cmy
01 Plant


sv dv

ssC ode v


Figure 7: Offline training and retuning configuration for steady-state optimizer.


The reason to have a CV model is obvious, it will provide the reference model that the optimizer uses to figure out its optimal MV setpoints. The motivation for having a SV model, however, is somewhat less apparent. The problem is that changes made to the MVs by the optimizer will not only change the CVs, but also the SVs. The optimizer will have to consider the effect that MVs will have on the SVs, if it is to accurately predict their effect on CVs. Note that the CV model has an input space that consist of MVs, DVs and SVs.


4.5.1.2 Online control

The online control configuration is illustrated in Figure 8. Here an optimizer calculates ArgMin {f } using the SV and CV reference models developed during model trainm v

ing. The optimizer starts with the current value of the MVs myv = my, uses the SV model to estimate the current SVs sv which are then used, along with the current value






76

of the DVs d to estimate the current CVs cv The optimizer then iteratively updates its estimate for the optimal MVs myv* to minimize its objective function J.


{dy, sv }







MVs is small, then direct optimizer provides an efficient alternative. As the number ofc







MVs grows, however, direct optimization quickly becomes impractical. Descent-based optimization is possible because the SV and CV models are capable, via backpropagation, of calculating the gradient of f with respect to their inputs, i.e., their inputs sensitivities given the sensitivities at their outputs. In this manner, the optimizer calculates the CV sen-A ArgMin ,/












sitivities /sSVM from which the ssCVModel is able to calculate SV sensitivities










af/.~v* and partial MV sensitivities afv/av*, from which the SV model calculates
the remainFigure 8: Online control configuratio* and finor steady-state optimizer is able to updaBoth direct and descent-based optimalization can be used for MPC. If the number ofgradient of MVs is small, then direct optimizer provides an efficient alternative. As the number of MVs grows, however, direct optimization quickly becomes impractical. Descent-based optimization is possible because the SV and CV models are capable, via backpropagation, of calculating the gradient of Ic with respect to their inputs, i.e., their inputs sensitivities given the sensitivities at their outputs. In this manner, the optimizer calculates the CV sensitivities 8//cla--' from which the CV Model is able to calculate SV sensitivities 8//8asv* and partial MV sensitivities 87 /8am ,_ from which the SV model calculates the remaining partial MV sensitivities 8/sv/8my ,V and finally the optimizer is able to update its optimal MV estimate using the MV gradient of af + aft (66) amy amy amy






77
This is really just the backpropagation of backpropagations, a.k.a. more fun with the chain rule.

Several optimization methods were tested for the optimizer, along with various techniques for dealing with the constraints. The most effective combination identified was to use the unconstrained conjugate gradients method in combination with an objective function which included the SV and MV constraint penalty functions
NfV NVv Nv
i = 1 1 SV V p mvi, i (67)
cvi= 1 1viv i= I

where the MV, SV, DV and CV variable sets, along with their corresponding constraints, are defined in Section 6.6.3 "Final Variable Sets," and all priorities have been set to 1. The details of the conjugate gradients method will be presented in Section 6.5 "Learning Algorithm."

Penalty function can negatively impact the performance of a descent-based optimizer by adding complexity to the performance surface having little to do with the underlying problem. This is particularly true when the constrained variables lie outside of their constrained values. For the SV constraints, there is no choice but to use penalty functions for constraints. For MVs, however, there are alternatives, because the MVs always start at their current values which are always within the constraints. Hence, there is little to no overhead to using MV constraints for our online optimizer. It terms of the performance surface, the constraints can be thought of as placing a guardrail on both sides of our current position in weight-space along our path, while having little impact on the local topography of the road.






78

4.5.2 Steady-State Model-Inverse Controller

The next controller design belongs to the model-inverse control (MIC) family. Conceptually, model-inverse control is straightforward: train a model to predict the MVs from the current and known DVs, SVs and CVs, then, given a desired CV setpoint, this model can be used directly to obtain the required MVs. Implementing a MIC controller is also straightforward and can work reasonably well, given that the relationship between MVs and CVs is in fact invertible. This sections details the design of a neural network-based MIC controller, designed to meet the problem specifications presented in Sections 4.2, 4.3 and 4.4.


4.5.2.1 Offline training

Training a MIC controller follows the schematic outlined in Figure 9. Once again, notice that the MV model is being implemented by separate inverse-SV (ISV) and inverse-MV (IMV) models. Once again, the details for how to train these models will be covered in Chapter 6. The model definitions required by the steady-state model-inverse controller are:

1) Steady-State ISV Model: sv = sslSVModel(cv, dv)

2) Steady-State IMV Model: my = sslMVModel(cv, sv, dv)






79



S-4-- ssSV del




my Plant cy N
sv dv





7M ssI odel -+mv H Plant C









Figure 9: Offline training and retuning configuration for model-inverse controller.


Analogous to our MPC controller, two models have been developed which when combined can invert the process. The reason to have a IMV model is obvious, it provides the inverse-model that the controller uses to figure out optimal MV setpoints. The problem is that not all CV-SV combinations are feasible. Given a specified CV target, the ISV model estimates the corresponding SVs which are presented to the IMV model.


4.5.2.2 Online control

The online control configuration is illustrated in Figure 10. If a known target existed for the CVs, the online control implementation would actually be quite trivial. One complexity is that the exact value for the lowest achievable NOx from the controller for a given set of conditions is not known. Another complication with MIC is how to deal with constraints. If one applies a target CV to the input of the inverse-model, it will predict a set of inputs which it believes would have achieved this target. The problem is that the model does not understand the MV or SV constraints, and if one of the inputs it predicts falls out-






80
side these constraints the controller can not provide the required setpoints. This is analogous to the problem faced with SV or CV constraints for MPC. The implementation outlined in Figure 10, uses an optimizer in order to overcome both of these issues. Clearly, CV constraints are straightforward.



{dv, sv}

Op timizer mv* cv Plant
ArgMin_* P{ } "I
cyV



L sslSVModel ssIMVModelFigure 10: Online control configuration for model-inverse controller.



The MIC controller uses an optimizer to calculate ArgMin_* {J} using the ISV and
CV

IMV reference models developed during model training. The optimizer starts with the current value of the CVs cv = cv, uses the ISV model to estimate the current SVs sv which are then used, along with the current value of the DVs dv by the IMV model to estimate the current MVs my The optimizer then iteratively updates its estimate for the optimal MVs m v*, to minimize its objective function I.

Once again both direct and descent-based optimization can be used for MIC, and once again a conjugate gradients-based optimizer was selected. Descent-based optimization is possible because the ISV and IMV models are capable, via backpropagation, of calculating the gradient of J with respect to their inputs, i.e., their inputs sensitivities given the






81
sensitivities at their outputs. In this manner, the optimizer calculates the MV sensitivities a/mv*, from which the IMV Model is able to calculate SV sensitivities 8//sv* and partial CV sensitivities 8/Jf/acv*, from which the ISV model calculates the remaining partial CV sensitivities 8 sv/&* and finally the optimizer is able to update it optimal CV estimate using the CV gradient of / aJs', 8/'"
a +SV a (68) 8cy 8cy 8cy

The optimizer's objective function, which includes the SV and MV constraint penalty functions, is the same objective function used by our steady-state optimizer. The only difference is how the sensitivities flow through the system, as outlined above.


4.5.3 Dynamic Model-Predictive Controller

The steady-state optimizer considered above is a member of the MPC family. The vast majority of MPC applications use models which are first-principles based [37]. Since it is not possible to build an accurate first-principles model ofNOx, a new steady-state optimizer for MPC using neural network models was developed. The vast majority of MPC applications are dynamic, however. The steady-state optimizer only considers the effect that MV changes will have on the unit in steady-state conditions.

This section develops a dynamic neural network-based MPC controller. The main differences between this controller and our steady-state optimizer is that it:

1) Understands the dynamics of the process.

2) Provides a trajectory of MV setpoints designed to optimize the path of
the unit into the future, rather than a optimal steady-state position. In
other words, the controller not only considers where your going but how
you'll get there.






82
The concept behind this controller's operation is identical to that of the steady-state optimizer: combine a model for the process with an optimizer to obtain real-time optimal setpoints. The only difference is that the models are now dynamic, and the optimal setpoints become optimal setpoint trajectories.

This sections details the design of a neural network-based dynamic MPC controller to meet the problem specifications presented in Sections 4.2, 4.3 and 4.4.


4.5.3.1 Offline training

Training a dynamic MPC controller follows a similar schematic as outlined in Figure 7, with the inclusion of each variables explicit dependence on time t, as illustrated in Figure 11. Here the SV and CV models are performing single-stage prediction with respect to the MVs and DVs; notice that the CV model uses the current value of the SVs sv(t + 1). The reasons for this configuration will become clear when we consider the online control implementation in the next section.





dS/Vodel f

dv(t) sv(t + 1)

m i (t) rPlant cy(t + 1) sV(t+ 1) d(t)

d1 Nodel E



Figure 11: Offline training and retuning configuration for steady-state optimizer.






83
Refer to Chapter 6 for details on training the dynamic SV and CV reference models used by the dynamic MPC controller. For now we simply state the model definitions required by the dynamic MPC controller:

1) Dynamic SV Model: sv(t+ 1) = dSVModel(mv(t), dv(t))

2) Dynamic CV Model:
cv(t + 1) = dCVModel(mv(t), dv(t), sv(t + 1))

4.5.3.2 Online control

The online control configuration follows a similar configuration to the steady-state

optimizer presented in Figure 8. Here a dynamic optimizer is required, however. The optimizer calculates ArgMin- {f(t) } using the dynamic SV and CV models developed mv*(t)

during model training. The steady-state optimizer used an application of the chain rule for ordered partial derivatives, which has been coined "backpropagation" [40]. From the perspective of the chain rule, our new optimizer is identical and only the criterion changes. From the perspective of the literature, this algorithm has been coined "backpropagation through time" [40][70].

The optimizer starts with the current value of the MVs m *(t) = mv(t); uses the SV model to estimate the resulting SVs sv (t + 1); which are then used, along with the current value of the DVs dv(t), to estimate the resulting CVs cv (t + 1). Notice that each estimate can rely on both present and past values of the inputs. The optimizer will then repeat this process over the time interval t e (t0, to + T] to produce the MV, SV and CV to+T- ___, to +T to+ T trajectories {mv*(t)} = to {sv (t)}t = to+ 1 and I{cv*(t)}t = to + respectively.






84
The objective function which included the SV and MV constraint penalty functions can now be calculated

1 +Tr IV +,
(t) = pi Z(cvi*(t)) p k(svi*(t), Civ) + p i v(mvi*(t), Civ) (69)
t=o i= 1 i=1 N i=

where the MV, SV, DV and CV variable sets, along with their corresponding constraints, are defined in Section 6.6.3 "Final Variable Sets," and all priorities have been set to 1.

The optimizer then iteratively updates its estimate for the optimal MV trajectories,

to+ T-1 I
{m (t) = to to minimize its objective function f(t). Each step in the iteration performs the following, starting with t = to + T- 1 and iterating down to t = to: first, the optimizer calculates the CV sensitivities af(t)/acv* (t), from which the CV Model is able to calculate SV sensitivities 8f(t)/8sv (t) and partial MV sensitivities afv(t)/lav*(t 1), from which the SV model calculates the remaining partial MV sensitivities oJsv(t)/8m-v*(t 1), and finally the MV sensitivity at time t 1 can be calculated as

8/(t) Jv(t) + ajcv(t) (70) am* (t 1) m-v* (t 1) m*(t l) Once the backward pass is complete, the optimizer is now able to update it's optimal to+ T-1
MV trajectory estimate { m v*(t) } = to using the MV gradient trajectory
(t) t+ T- 1
m V (t) t to(71)
8my (t)t=t






85

Notice that the sensitivities at time t depend on the sensitivities in the future. This is because the models variables at time t depend on the variables in thepast, i.e., the models are dynamic. Hence the term "backpropagation through time."

This entire optimization cycle is run at each time step to. The optimizer derives the next T- 1 MV moves, and the first MV setpoint is applied to the unit mv*(to). At this point the entire process is repeated.


4.5.4 Model-Reference Adaptive Controller

The final controller design considered belongs the model-reference adaptive control family (MRAC). Like the dynamic MPC controller, the MRAC controller understands process dynamics and provides a trajectory of MV setpoints which optimize both where you are going and how you get there. The fundamental difference between these two controllers is how this optimal trajectory is derived. The MPC design utilized an online optimizer to calculate this trajectory, while the MRAC design develops a neural-network based controller which is able to calculate the optimal trajectory directly. Hence this is our first direct controller, i.e., calculates MV setpoints directly.

Notice that the MIC design would have provided a direct online controller, if it wasn't for:

1) the lack of a known target NOx level, and

2) the requirements for MV and SV constraints.

The MRAC design is able to overcome both of these hurdles by building knowledge of the best achievable NOx level and by building all of the constraints directly into the controller. The main advantage to the MRAC design is online response time. There is no opti-






86
mization to run, one simply presents the controller with the current, and past, state of the process, and it generates a MV setpoint as quickly as a neural network can think. These benefits do not come for free, however. The main drawbacks to the MRAC design are:

1) Extensive offline training and retuning requirements.

2) Inflexible online configuration, with respect to changing optimization
objectives and operating constraints.

This sections details the design of a neural network-based dynamic MIRAC controller to meet the problem specifications presented in Sections 4.2, 4.3 and 4.4.


4.5.4.1 Offline training

Training a MRAC controller requires two stages. The first stage is identical to training and returning the dynamic MPC controller. Here, dynamic SV and CV models are developed using the same steps outlined in Figure 11. The second stage uses these models to train the controller with offline data, as illustrated in Figure 12. The offline training is similar to the online optimization which is performed for the MPC design, except this optimization is performed across the training dataset rather than online.




>vsv (t + 1)
dSVModel
di~t) cy(t + 1) Con ol dCVModel

W-V
rm My (t)

-11
zz


ArgMin CL{ (t)}

Figure 12: Offline training and retuning configuration for model reference controller.






87
Once again a dynamic optimizer is required, and the objective function is given by

(69). To train the controller, the optimizer calculates A rgMin /L { f(t) } where ~CL are the weights of the control law neural network.

Training the control law (CL) model starts with the actual values for the DVs, SVs and CVs, and random initial weights for its CL model tCL Starting at time to = Ti + 1, where Ti is the first sample in the training dataset; training uses the CL model to estimate the resulting MVs mv*(t); which are then used to estimate the resulting SVs sv* (t + 1) and CVs cv (t + 1). This process is repeated over the time interval t e (to, to + T] to to+ T- 1 to+ T
produce the MV, SV and CV trajectories {mv (t)} t = to {sv*(t)}t= t + 1 and

to+ T
{ cv (t) } t = to + I, respectively.

The training algorithm then iteratively updates its estimate for the optimal CL model weights, TPCL, to minimize its objective function 1(t). Each step in the iteration performs the following, starting with t = to + T and iterating down to t = to + 1: First, the training algorithm calculates the CV sensitivities 8f(t)/Oacv* (t), from which the CV Model is able to calculate SV sensitivities af (t)/8sv* (t) and partial MV sensitivities 8/v(t)/Dmv*(t 1), from which the SV model calculates the remaining partial MV sen-






88

sitivities afv(t)/Dmv*(t- 1), and finally the MV sensitivity at time t- 1 can be calculated as


Of(t) 8 ov(t) + afV(t) (72) 8vm (t- 1) av (t- 1) av (t-1)

The MV sensitivities are finally passed to the CL model which backpropagates them to derive its weight gradients af /80cL, which the training algorithm is able to use to update its control law's weight estimate.

The training algorithm then increments to and repeats the entire process, until the

training algorithm has converged. When to = Tf T, to is reset to to =Ti + 1 The reason for training the CL model in increments of T is because the SV and CV have a limited prediction horizon, the time before their estimates are no longer valid. By resetting the state of these models to the actual state of the unit after T samples, we are able train the CL model within the prediction horizon of the SV and CV models.


4.5.4.2 Online control

The online control configuration for the MRAC design is straight forward, as illustrated in Figure 13. Simply supply the controller with the current SVs, DVs and CVs, and it outputs the next MV setpoint. This setpoints contains knowledge about the optimal achievable NOx, SV constraints, MV constraints and the trajectory through which it will drive the process into the future.






89

S{sv(t), dv(t)}

Control my (t) cv(t) Law W-1 Plant
r Law

Figure 13: Online control configuration for model reference controller.


Clearly, the controller is only as good as its underlying reference models. In addition, considerable care must be taken to ensure that the training data contains regions of the input space where the SV and MV constraints have been exercised. The design is easily augmented with limiters to guarantee that MV constraints are maintained. However, there is little that can be done to guarantee that the SV constraints are maintained.













CHAPTER 5
DATA PREPARATION

Given the detailed control designs just presented, the next step is to implement the

actual controllers by developing the required reference models. Both reference model and controller implementations require a significant amount of process data. Data collection is the most important aspect of any modeling or optimization project. There is a common saying "junk in, junk out," this study was relentless in reenforcing this lesson. Applying the most sophisticated modeling and/or optimal control algorithms in the world will not make up for problems with data preparation.

With the advanced distributed control systems (DDS) and supervisory control and data acquisition (SCADA) systems readily available in today's process plants, the relative quantity and quality of available data is overwhelming. Much of the statistics and modeling literature has been dedicated to the problems faced when drawing inferences from small sample spaces. Modem processing plants are anything but data limited. The relevant problems are just the opposite, how to draw meaningful inferences from a massive sample space.

The following section presents solutions for the most significant challenges faced in preparing data for modeling and optimization. Much of what is presented in this section was learned the hard way, during modeling and optimization.







90




Full Text
NEURAL NETWORK-BASED CONTROL DESIGNS FOR COMPLEX
INDUSTRIAL PROCESS APPLICATIONS
By
WESLEY CURTIS LEFEBVRE
A DISSERTATION PRESENTED TO THE GRADUATE
SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL
FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE
OF DOCTOR OF PHILOSOPHY
UNIVERSITY OF FLORIDA
2000

ACKNOWLEDGMENTS
When looking for a graduate school I was told to pick an advisor, not a school. My
choice was Dr. Jose Principe, and his mentoring and friendship through a master’s, Ph.D.,
and the startup of two companies has proven this advise to be sound. I would also like to
thank Dr. Glen Johnson from IBM Yorktown for giving me this advice and first introduc¬
ing me to the field of neural networks.
I would also like to thank Charles River Associates and Commonwealth Energy
for believing in my work and this field, the members of my committee for the effort and
time spent on my behalf, the staff at Canal Electric Generating Station, and the members
of the Computational NeuroEngineering Laboratory (CNEL).
I would also like to thank my wife, Anna, for her patience as I struggled to juggle
far too many commitments, and my family for their never-ending support referring to us
as Dr. and Mr. Lefebvre.
n

TABLE OF CONTENTS
page
ACKNOWLEDGMENTS ii
LIST OF TABLES v
LIST OF FIGURES vi
CHAPTERS
1 INTRODUCTION 1
Proposed Work 1
Document Organization 7
2 LITERATURE REVIEW 9
Optimization 9
Neural Networks 16
Control Theory 32
NOx 40
Fossil-Fired Power Generation 45
3 BOILER OPTIMIZATION 63
First Principles 63
Fuel and Air Distribution 64
Boiler Tuning 65
The Role of CO 66
4 CONTROL DESIGNS 67
Variable Definitions 67
Optimization Objectives 68
Operating Constraints 70
Performance Criteria 72
Controller Designs 73
iii

5DATA PREPARATION
90
Data Management 91
Variable Selection 92
Validation 93
Time Constants 97
Normalization 97
6 MODELING 99
Methodology 99
Model Definitions 100
Datasets 104
Performance Criteria 105
Learning Algorithm 109
Variable Pruning 113
Architecture Selection 118
Analysis 136
7 CONTROLLER IMPLEMENTATIONS 138
Offline Quantification 138
Online Quantification 143
8 PARAMETERIZATION PROBLEM 148
Search for a Validation Metric 149
Correlation Paradox 154
Validation Metric 160
Revised Representation Pruning Algorithm 161
Modeling 168
Control Implementation 169
9 CONCLUSION 176
Contributions 176
Afterword 178
Future Direction 178
APPENDIX 180
REFERENCES 181
BIOGRAPHICAL SKETCH 187
IV

LIST OF TABLES
Table 1: Probability of Z-Score exceeding value 96
Table 2: Types in order they were removed 116
Table 3: Final variable selections after pruning 118
Table 4: Results of auto-regressive tap search algorithm for all dynamic models 121
Table 5: Results of moving average tap search algorithm for all dynamic models 122
Table 6: Results of hidden layer #1 PE search algorithm for all steady-state models. 124
Table 7: Results of hidden layer #2 PE search algorithm for all steady-state models. 126
Table 8: Results of tap search algorithm for all dynamic models 128
Table 9: Results of taps search algorithm for all dynamic models 130
Table 10: Results of hidden taps search algorithm for all dynamic models 131
Table 11: Results of hidden states search algorithm for all dynamic models 133
Table 12: Results of state hidden PEs search algorithm for all dynamic models 134
Table 13: Results of output hidden PEs search algorithm for all dynamic models 136
Table 14: Final variable selections after revised pruning 165
Table 15: Essensial tag list 180
v

LIST OF FIGURES
Figure 1: Multilayer perceptron model architecture 18
Figure 2: TDNN input PE connectivity 19
Figure 3: Gamma memory processing element 20
Figure 4: Nonlinear state space neural network configuration 21
Figure 5: Combustion emissions characteristic versus air flow 64
Figure 6: Effect of lower 02 on combustion emissions 65
Figure 7: Offline training and retuning configuration for steady-state optimizer 75
Figure 8: Online control configuration for steady-state optimizer 76
Figure 9: Offline training and retuning configuration for model-inverse controller 79
Figure 10: Online control configuration for model-inverse controller 80
Figure 11: Offline training and retuning configuration for steady-state optimizer 82
Figure 12: Offline training and retuning configuration for model reference controller. .86
Figure 13:Online control configuration for model reference controller 89
Figure 14: Daily % missing across February dataset 94
Figure 15:Example of variables with large NMSE but high R 108
Figure 16: Results of type pruning algorithm 115
Figure 17: Representation pruning algorithm results 117
Figure 18: Results of auto-regressive taps search algorithm for NOx ARMA Model. ...120
Figure 19:Results of moving average tap search algorithm for NOx ARMA Model. ...121
Figure 20:Results of hidden layer #1 PE search algorithm for the NOx MLP Model. ..124
vi

Figure 21 :Results of the hidden layer #2 PE search algorithm for NOx MLP Model. ..125
Figure 22:Results of tap search algorithm for NOx TDNN Model 127
Figure 23: Results of taps search algorithm for NOx GNN model 129
Figure 24:Results of hidden taps search algorithm for NOx GNN model 130
Figure 25: Results of hidden states search algorithm for NOx NLSS model 133
Figure 26: Results of state hidden PEs search algorithm for NOx NLSS model 134
Figure 27:Results of output hidden PEs search algorithm for NOx NLSS model 135
Figure 28:Best models for all model definitions by architecture 137
Figure 29: Average NOx reduction over testing dataset 141
Figure 30: Average CO above max over testing dataset 141
Figure 31: Average NOx reduction over testing dataset using train and test models 142
Figure 32: Average CO above max over testing dataset using train and test models 142
Figure 33: Change in NOx for steady-state controller experiments 146
Figure 34:Final CO level for steady-state controller experiments 147
Figure 35: Summary of validation metrics for MLP CV model 150
Figure 36:NMSE and R for all 10 training results for MLP CV model 151
Figure 37: Summary of validation metrics for combined SV/CV model 153
Figure 38: Sensitivity results for all 10 training results for NOx CV model 160
Figure 39:NOx CV model sensitivity with 95% confidence intervals 161
Figure 40: Results of revised representation pruning algorithm 164
Figure 41: Sensitivity results for all 10 training results for revised NOx CV model 167
vii

Figure 42:Revised NOx CV model sensitivity with 95% confidence intervals 168
Figure 43: Best revised models for all model definitions by architecture 169
Figure 44: Average NOx reduction over testing dataset using old and revised models. .170
Figure 45: Average CO above max over testing dataset using old and revised models. .170
Figure 46: Change in NOx for revised steady-state controller experiments 172
Figure 47:Final CO level for revised steady-state controller experiments 172
Figure 48: Average percent NOx reduction for 10 online experiments 173
Figure 49:Average percent CO reduction above 500ppm for 10 online experiments. ...174
viii

Abstract of Dissertation Presented to the Graduate School of
the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy
NEURAL NETWORK-BASED CONTROL DESIGNS FOR
COMPLEX INDUSTRIAL PROCESS APPLICATIONS
By
Wesley Curtis Lefebvre
May 2000
Chairman: Jose C. Principe
Major Department: Electrical and Computer Engineering
Neural networks have successfully transitioned from an academic interest into a
viable technology which is now being used in everyday products. To date, neural net¬
works have been predominantly applied to forecasting or modeling applications. Based on
their success in such applications, there has been significant interest in using neural net¬
works in control applications, creating a new field called neurocontrol. Although there
have been significant advances in the theory of neurocontrol, there are very few successful
commercial applications using neurocontrollers. Commercial applications often provide
the most challenging problems because the controllers are required to function robustly in
complex and unknown environments. Real-world processes are complex and difficult to
control because they contain a large number of highly interdependent variables, have
highly nonlinear responses to these variables, and change their response over time.
This work identified two significant reasons why neurocontrol designs fail in real-
IX

world applications: First, the controllable parameters over most industrial processes are
highly correlated, often not for physical reasons but because of our process control strate¬
gies. Second, intermediate process states that affect the process output, which are also
affected by the controllable parameters, have a significant impact on controller perfor¬
mance. When the controller changes the controllable parameters, the impact that this has
on the process states, which will in turn affect the process output, is not accounted for in
most neurocontrol designs in the literature.
This dissertation advances the field of neurocontrol by providing the following
solutions: first, the use of statistical significance testing on the local linearized relation¬
ships extracted from nonlinear neural network models to avoid problems with correlated
controllable parameters; second, augmenting neurocontrol designs to incorporate depen¬
dent state models. These enhancements have been applied to four distinct neurocontrol
architectures. The new control architectures have been applied to the novel application of
controlling NOx emission from an oil and gas-fired electric power plant.
x

CHAPTER 1
INTRODUCTION
Neural networks have found an applications niche as robust predictors that have dem¬
onstrated the ability to out-forecast more traditional methods in complex real-world appli¬
cations. The vast majority of neural network applications to date rely solely on the
model’s ability to forecast without regard for what the model has learned about the rela¬
tionships within the underlying process or the ability to affect it. Soulie and Gallinari [59]
recently compiled 53 industrial applications of neural networks, of which only 4 make any
attempt to make inferences about the underlying process or to control it.
This ratio does not apply to theoretical publications in the literature, however. The
research into neural network-based control has recently received widespread attention,
coined neurocontrol [40]. Many authors have presented the abstract concepts behind neu¬
rocontrol [40][43][31 ] [37], but the literature contains a disproportionately low number of
papers presenting real-world neurocontrol applications. The vast majority of these papers
have developed controllers for fabricated, simulated, or laboratory controlled processes.
1.1 Proposed Work
The goal of this project is to develop robust neurocontrol design methodologies for
complex industrial process control applications. Industrial control applications are charac¬
terized by nonlinear, noisy, non-Gaussian, highly correlated, and nonstationary processes.
This is an applications area where classical control designs have proven ineffective.
1

2
1.1.1 Case Study
As a case study, this work develops on-line advisory neurocontrollers designed to min¬
imize the NOx emissions for an oil and gas co-fired power plant. The combustion of fossil
fuels inside a large-scale boiler is a highly complex process; this complexity is a direct
function of the boiler size. A typical electric power boiler maintains a "fireball" which is 3
to 5 stories tall, and there are hundreds of parameters which affect the injection of fuel and
air at different locations within the furnace. The problem is our lack of understanding
about liow these combustion parameters affect NOx formation. This multivariate optimi¬
zation problem requires a technology that can look at the process globally and determine
the appropriate combination of combustion controls
The neurocontrollers developed in this study will forward control setpoints to opera¬
tors through the plant's existing distributed control system (DCS). The neurocontrollers
will be required to provide setpoints that minimize NOx emissions, while maintaining unit
operating constraints. A demonstration system has been completed at Canal Electric’s
580-MW tangentially fired Unit 2. Charles River Associates (CRA) and Commonwealth
Energy jointly funded this study.
1.1.2 Objectives
The author believes that neurocontrol strategies have not been more successful in real-
world applications because they
1) are difficult to implement,
2) fail to account for dependent internal process states, and
3) have difficulty dealing with correlated process variables.
The scientific objectives for this work are to develop:

3
1) an application-based neurocontrol implementation methodology,
2) state-space neurocontrol architectures,
3) methods for dealing with correlated data,
4) accurate combustion models, and
5) a novel combustion controller.
1.1.2.1 Application-Based Neurocontrol Implementation Methodology
Most neurocontrol implementations in the literature have been ad hoc and application
specific [40], In many cases, the process has been simulated and thus known completely.
Modem control theory has introduced many sophisticated control designs, but the fact
remains that approximately 90% of the industrial control applications apply a simple pro¬
portional integral differential (PID) controller [37],
The PID controller only requires the process engineer to specify reasonable knowl¬
edge of the process. This application-based implementation methodology is largely
responsible for the success of PID control in the industry. If neurocontrollers are to enter
the mainstream process control market, there will have to be designs that do not require
detailed knowledge of neural network or neurocontrol theory.
1.1.2.2 Application-Based Neurocontrol Implementation Methodology
We begin in Chapter 4 by defining a methodology for categorizing process variables
into groups based on a set of objective criteria about their role in the process. Each con¬
troller will be implemented in Chapter 7 based on this labeling of the process variables,
without additional process knowledge. This will allow process engineers with reasonable
process knowledge and without any knowledge of neural network or neurocontrol theory
to successfully deploy such technologies. Notice that achieving such a straightforward

4
implementation methodology does not imply that the details behind the neurocontroller
implementation are easy, it simply requires that they can be automated.
1.1.2.3 State-Space Neurocontrol Designs
Nonlinear state-space neural network architectures offer the greatest modeling poten¬
tial, but the difficulties in their training have led investigators to reject their application
[31]. In fact, most neurocontrol designs are based on input/output process models
[31][43][40], The use of input/output neural network architectures in the design of neuro¬
controllers assumes that all input variables are independent, a situation which in not likely
to be found in real-world applications. Many of the process variables that affect the plant’s
output will depend on the same process setpoints that a controller is manipulating. These
facts have limited the performance and complicated the implementation of neurocontrol
designs. If neurocontrol is to be a viable methodology in industrial process situations, then
the controllers will have to be extended with architectures capable of dealing with internal
process states.
There has been limited success in applying state-space neural network models [31]. It
is widely accepted, however, that state-space representations hold the most promise for
modeling and controlling complex processes [34][35]. The literature seems to be treating
the viability of state-space architectures as an all-or-nothing affair. Most publications
apply purely input/output architectures with overwhelming success, while a few investi¬
gates have tested purely state-space architectures where all states are treated as hidden and
unknown with limited to no success.
This work will empirically investigate several shades of gray, ranging from purely
mput/output to purely state-space controllers. These state-space controller designs are pre-

5
sented in Chapter 4, and their performance is empirically investigated in chapters 7 and 8.
The primary difference in the state-space representations proposed here is that the state
variables are treated as not hidden and known. This will require an extension to the neuro¬
control design strategies presented in the literature [31][43][40].
1.1.2.4 Methods for Dealing with Correlation
Industrial process applications are unique in that there is a massive amount of avail¬
able data. The input variables for complex process models are typically highly correlated,
a situation for which there are few solutions in the literature. This correlation can come
from several sources: dependent states (as addressed above), physical linkages, and soft
linkages through control strategies the lack of adequate system parameterization. Indus¬
trial process control applications will require the process engineer to select representative
variables from a large set of available process variables. Because of issues like input cor¬
relation, the representative variables selected will have a significant impact on the perfor¬
mance of the resulting neurocontroller. A viable neurocontrol design methodology will
have to be able to cope with correlated process variables. Linear control theory deals with
this aspect through parameterization of the controller. Nonlinear control theory, however,
has not solved this problem in general. In Chapter 8 we propose to compute sensitivities
through a committee of trained neural models to select the best variables for system iden¬
tification and control.
State-space neurocontrol architectures will be able to explicitly deal with one source
of correlation present in industrial processes, namely correlation produced by dependent
state variables. As mentioned above there are several other sources of correlation, how¬
ever. Our first goal is to empirically quantify the impact of this correlation on our model-

6
ing and control objectives. We begin by ignoring other sources of correlation and
investigate their impact in Chapter 7. Once specific problems have been identified, meth¬
ods are developed for dealing with correlation during controller implementation in Chap¬
ter 8, and the performance of these methods are quantified with respect to the performance
of the resulting controllers.
1.1.2.5 Accurate Combustion Models
Little is known about how NOx is formed from air-bound nitrogen during combustion.
To date, reliable models for NOx formation in electric power boilers have not been avail¬
able [39], There are not adequate models for many real-world industrial processes, a fact
which has also limited the acceptance of modem control strategies. One reason that neuro¬
control has received such widespread attention is because of its potential ability to deal
with very complex processes that have escaped modeling.
Chapter 6 develops accurate combustion models according to accepted modeling per¬
formance metrics. Chapter 7 demonstrates the impact that correlation has on neurocontrol
designs, and Chapter 8 investigates its impact on the accuracy of the underlying process
models. Here it will be shown that accuracy is subjective, and that in fact no good metrics
for model accuracy are available in the literature. A new metric is proposed and empiri¬
cally compared against available metrics in the literature in Chapter 8. Applying this new
metric, predictive combustion models are developed and used to shed light on which pro¬
cess variables have the greatest impact on NOx and CO formation.

7
1.1.2.6 Novel Combustion Controller
This project develops four neurocontrollers for the complex industrial process of NOx
formation. We begin in Chapter 3 by looking at boiler optimization from a first-principles
perspective, focusing on what a NOx controller is expected to achieve and why classical
control methods are not able to achieve it. Predictive neural network combustion models
are then developed in Chapter 6, and deployed within online neurocontrollers in chapters 7
and 8. The performance of each of these controllers is then quantified to compare and con¬
trast the four control designs in Chapter 8.
To the authors knowledge, this work developed the first NOx controller for a gas and
oil co-fired electric power plant. New regulations and the restructuring of the electric
power industry have combined to create a NOx trading market. The annual benefits to a
gas and oil co-fired electric power plant associated with a 25% NOx reduction will be in
the range of $2,000,000 to $5,000,000. Clearly, a control strategy that uses existing plant
capital investments and runs on a $2,000 pentium workstation has tremendous value.
1.2 Document Organization
Chapter 2 presents a summary of the required background and a literature review of
the relevant work in neurocontrol, along with references to the literature for more detailed
treatments. This chapter is the best place for readers to become familiar with the notation
used throughout this work.
A strategy for reducing the NOx emissions from a fossil-fired generating unit is pro¬
vided in Chapter 3. The goal of this section is to provide a physical understanding for what
we are asking the controllers to perform, thus providing justification that our objectives
are feasible. Chapter 4 develops four detailed neurocontrol designs belonging to the model

8
predictive, model inverse, and model reference control families. The designs are presented
as generalized methodologies that are applicable to any control application. Chapter 5 pre¬
sents a management and preprocessing methodology for collecting data in support of these
control designs and the required modeling.
Each control design considered requires accurate process models for its implementa¬
tion. Chapter 6 presents a modeling methodology for developing these models. The con¬
trollers are then implemented in Chapter 7. The performance for each resulting controller
is then quantified using offline simulations and online experiments. Significant problems
are discovered with the controllers for which there are no solutions in the literature. These
problems, along with proposed solutions, are investigated further in Chapter 8. This sec¬
tion additionally demonstrates the validity of these solutions by quantifying the perfor¬
mance of the revised controllers.
The "key learnings" and extensions to this work are summarized in Chapter 9.

CHAPTER 2
LITERATURE REVIEW
This chapter provides background for the rest of this document in the areas of:
1) optimization,
2) neural networks,
3) neurocontrol,
4) NOx, and
5) fossil-fired power plants.
2.1 Optimization
Mathematical optimization methods are at the heart of modem modeling and control
applications. Neural networks use optimization methods to facilitate learning, and control
applications apply these methods to meet their control objectives. The notation and meth¬
ods presented in this section will be used extensively throughout the rest of this document.
Optimization is defined as the process of finding the values of N2 decision variables
z e N that minimize a scalar performance objective J e 'is —> 9? [15]. Formally, this
optimization task will be represented as
ArgMini{J(z)}, (1)
where hi is the decision variable space, which is most often taken to be euclidean
N = 91. Optimization methods, also known as mathematical programming methods, can
be classified according to the amount of a priori information available about the system
9

being optimized [54], The following sections broadly categorize optimization methods
into the following:
10
1) Classical Analytic Optimization: where the system being optimized is
known completely or nearly completely and a tractable analytic solution
exists
2) Descent Optimization: where first and/or second-order partial deriva¬
tives are available everywhere for the parameters of the system being
optimized
3) Direct Optimization: where little to no a priori knowledge exists about
the physical structure of the system being optimized
If the optimization problem involves objective functions or constraints which cannot
be stated as explicit functions of the design variables or are too complicated to manipulate,
we cannot solve it by using classical analytic optimization methods. This work will be
dealing with complex systems where little is known a priori and will therefore not con¬
sider analytic optimization methods.
2.1.1 Iterative Optimization Methods
All direct and descent optimization methods are iterative in nature, i.e., they start from
an initial trial solution and proceed toward the minimum point in a sequential manner. An
iterative optimization method is typically judged based on its rate of convergence [54], In
general, an optimization method is said to have convergence of order p if
||£(n+ 1)-HI ¿>0 p> 1, (2)
i¿(rc) - nr
where z(n) and + 1) denote the points obtained at the end of iterations n and n + 1,
respectively, z* represents the optimum point, and ||x|| denotes the length or norm of the
vector x.

11
If p = 1 and 0 < k < 1, the method is said to be linearly convergent (corresponds to
slow convergence). If p = 2, the method is said to be quadratically convergent (corre¬
sponds to faster convergence) [65].
2.1.2 Direct Optimization
In problems where analytic solutions are not possible and the design variables are of
mixed type there is little choice but to use some variation on a direct search methodology.
Direct searches may be broken into the following broad categories.
2.1.2.1 Exhaustive methods
In most practical applications, the optimum solution is known to lie within restricted
ranges of the design variables. Exhaustive search methods are applied to problems where
the interval in which the optimum is known to lie is finite. Conceptually, these methods
evaluate the objective function at a predetermined number of points in this interval and
reduce the interval of uncertainty using the assumption of unimodality. Exhaustive meth¬
ods include [54]:
1) Random Search
2) Grid Search
3) Pattern Directions
2.1.2.2 Elimination methods
The exhaustive search methods are similar to a larger class of algorithms known as
elimination methods, because they search by eliminating parts of the interval. Elimination
methods differ in how they search and discard sub-intervals. The more common elimina¬
tion methods include [54]:
1) Dichotomous Search

12
2) Interval Halving
3) Fibonacci Method
4) Golden Section Method
2.1.2.3 Interpolation methods
Interpolation methods iteratively fit the local performance surface with a simple poly¬
nomial form, and then approximate the minimum point of the system as the minimum
point of the polynomial [65]. These methods are generally more efficient than elimination
methods and can be accelerated if gradient information is available. Some of the more
popular interpolation methods include [54]:
1) Quadratic Method
2) Cubic Method
3) Newton Method
4) Quasi-Newton Method
5) Secant Method
2.1.2.4 Unrestricted methods
When the design variable range is not known the search must be performed without
restrictions on the values of the variables. Most of these methods use a step size and move
from an initial guess in favorable direction (positive or negative) [54], The step size used
must be small in relation to the final accuracy desired. This method is often accelerated by
using a variable step size. These methods include [54]:
1) Simplex Method
2) Revised Simplex Method
3) Karmarkar’s Method
4) Hook’s and Jeeves’ Method

13
5) Rosenbrock’s Method
In addition, evolutionary computing techniques like genetic algorithms belong to this
category.
2.1.2.5 Line search
All of the direct search methods presented above can be applied to both one-dimen¬
sional or n-dimensional searches. A one-dimensional search is often referred to as a line
search since we are searching along a line. The aim of all line searches is to find q* e 91
such that
r|* = ArgMinr[{J(z + r\zd)},
(3)
where z is the design vector, and zd is a known search direction.
One of the most efficient, and hence most popular, line searches uses the Quadratic
Method to find r|* [54]. This method has been applied in this work, using the following
algorithm:
I: Normalize the search vector z,d by dividing each component by the
absolute value of the element of zd with the maximum absolute value
II: Evaluate the function at the points A = 0 and D = q0, where q0 is
an initial step size
III: If JD>JA then set C = D and B = q0/2
IV: Else set B = D and evaluate at the point E = 2q0. If JE > JD then
set C = E. Else set D = E and q0 = 2q0, and goto step III
V: Calculate
rf
4Jb 3 JA Jc
4JB ~ 2JC - 2JA
(4)

14
VI: If JB - < A/"" then set r|* = rj* and quit
VII: If f|* < B then set C = B and B = rj* . Else set A = B and B = fj*
VIII: Goto step V
where A fu" is the minimum change in J to detect early stopping.
2.1.3 Descent-Based Optimization
z
When all values of z e 91 are possible and the function J(z) has first and second par¬
tial derivatives everywhere, the necessary conditions for a local minimum are
^-0.
Si
by which we mean dJ!dzi = 0, Vi and
(5)
(6)
by which we mean that the m x m -matrix whose components are dJ2/dzfizj must be pos¬
itive semi definite, i.e., have eigenvalues that are zero or positive [15],
All points that satisfy (5) are called stationary points. Sufficient conditions for a local
minimum are (5) and
q>o, (7)
sr
2 2
that is all eigenvalues must be positive. If (5) is satisfied but d J/dz~ = 0, that is, the
determinant of the matrix is zero (meaning that one or more of its eigenvalues is zero),

15
additional information is needed to establish whether or not the point is a minimum. Such
a point is called a singular point.
2.1.3.1 Methods
Classical analytic optimization methods use these conditions to solve for the optimal
solution. If the optimization problem involves an objective function or constraints that can
not be stated as explicit functions of the design variables, or which are too complicated to
manipulate, then descent optimization methods provide efficient alternatives. In general,
these methods will have significantly better convergence characteristics than direct meth¬
ods [54],
Descent search methods are iterative algorithms for improving estimates of the deci¬
sion variable, z, so as to come closer to satisfying the conditions for a stationary point.
The steps in using the descent method are as follows:
I: Set n = 0 and guess at the initial design vector z(n), usually random
II: Determine the values of dJ/dz(n)
III: Interpreting dJ/dz(n) as the gradient vector, determine the search
d
direction z (n) = f (dJ/dz(n)) as a function of this gradient
IV: Determine the step size to be taken r|(«) = , as a function
of this direction
V: Update the estimates of z(n + 1) = z(n) + r\(n)zd(n)
VI: Repeat II until (dH/dz(n))(dH/dz(n))T is very small

16
The variations in descent-based optimization can be expressed as variations in the
determination of the direction vector f and the step size/1. Some of the more common
variations include [15] [54][65]:
1) Steepest Descent:
zd{n) = -dH/dz(n) (8)
r|(«) = constant (9)
2) Steepest Descent with Momentum:
zd(n) = - (1 - p)dH/di(n) + pzd(n) (10)
r\ (n) = constant (11)
3) Conjugate Gradients:
= (dH/dzjn + 1 ))T\dH/dz(n + 1 )-dH/di(n)]
(dH/dz(n + 1 ))TdH/dz{n + 1)
r|(«) = LineSearch(z/«))
(12)
(13)
2.2 Neural Networks
Artificial neural networks (ANNs) are biologically motivated data processing structure
that consist of a large number of relatively simple highly interconnected neurons or pro¬
cessing elements (PEs) [24], In general, these structures provide an inductive mathemati¬
cal model that can be represented by
(14)
y^
where/:91 -» 91 is the model’s input/output map, and j> e 91 , ú e 91 , and
tv e 91^ are its outputs, inputs, and coefficients, respectively. The coefficients in an

17
ANN map are commonly referred to as weights. Artificial neural networks infer or learn
the relationships between p and w by observing actual process data. In this way, ANNs
can be applied to generalized regression and classification inference problems.
ANN architectures possess two fundamental properties:
1) They are capable of approximating to arbitrary accuracy any continuous
function, i.e., they are universal mappers [24],
2) They have robust optimization convergence properties with respect to
the optimization of their coefficients, i.e., they are robust learners [26],
These properties make ANNs a useful tool for empirical modeling tasks where little to
no a priori information is available about the underlying process.
2.2.1 Model Architecture
There are many types of ANNs in the literature, each with specific advantages when
modeling various types of processes [14]. The two primary factors which differentiate
between ANN models are their architecture and their learning rule. A model’s architecture
defines the way in which it processes input information to produce output information,
i.e., the form of their mathematical input/output map /.
2.2.1.1 Multi-layer perceptron
Most ANNs presented in the literature are static mappers, i.e., they are only capable of
modeling static or steady-state process relationships. By far the most popular and widely
applied ANN architecture is called the multilayer perceptron (MLP) [24], This network
consists of fully-interconnected layers of PEs with logistic response characteristics. The
MLP network is typically configured with one or two hidden layers of PEs. A two-layer
MLP is illustrated in Figure 1.

18
Figure 1: Multilayer perceptron model architecture.
Formally, using matrix algebra this architecture is given by
(15)
where Jp^1 e 91^'1
is the matrix of weights for the first hidden layer, t)h* e 91 ^ is a
vector of bias values for this layer, e 91 ^ Nl'' is the matrix of weights for the sec¬
ond hidden layer, t)'2 e 91^'2 is the bias vector for this layer, JpV e 91 ^ ^hl is the matrix
NyxNh2
At
of weights for the output layer, b e 91 ; is the corresponding bias vector, N,, is the
number of PEs in the first hidden layer, Nhl is the number of PEs in the second hidden
layer, a is the tanh logistic function, and the set Tv = %hl, JPy, b' } repre¬
sents the model’s weights.
When the process being modeled is dynamic, i.e., its current output is a function of its
current state as well as previous process states, static models are not well suited. For such
situations, models which are able to extract both static and temporal process relationships
are required. The most common method for creating dynamic neural networks is to simply
place dynamic PEs in the input layer of a static MLP [17]. These models have been
referred to by many researchers as dynamic neural networks (DNNs). The dynamic PEs

19
can have response characteristics based on a priori process knowledge or contain adaptive
memory mechanisms or fdters.
2.2.1.2 Time-delay neural network
The most common DNN is called the time-delay neural network (TDNN) [24]. This
architecture consists of a MLP where each input PE has an adaptive linear FIR filter, as
illustrated in Figure 2.
Figure 2: TDNN input PE connectivity.
Formally, the TDNN can be described by
/'(«(/), Nr) =
'IJ
Uj (0
J = 1
j e (2,Nr]
V/,y
J
fdn'\ü{t), NT), u>)
nip, Jdl.
(16)
(17)
Adi Nu x j1
where / : 91 —» 91 x represents the tapped-delay line operator, and N
ber of taps in the delay line.
is the num-
2.2.1.3 Gamma neural network
The main disadvantage to most DNNs is that they preprocess the input to extract fixed
and known dynamics of the process data rather than learn these dynamics from this data.
The TDNN can be considered as an exception to this rule, but here the process must have

20
finite impulse response (FIR) dynamics of known order. The Gamma Neural Network
(GNN) [17] represents an important class of dynamic ANN models that is able to leam
infinite impulse response (HR) process dynamics without a priori knowledge about the
structure or order of these dynamics.
Figure 3: Gamma memory processing element.
The GNN architecture is conceptually an MLP with adaptive Gamma Filters (GF)
placed at the output of its input and hidden layer PEs. A single GF is illustrated in Figure
3. Formally, the GNN is given by
jfjWO, nt)
«/ (0
j = 1
V/J (18)
*1_,(<) + (!- r¡)ffp -1) j e (2, NT]
tí-) = /Vr) T l/"' ),Nr ) -i t!’2) . U , (19)
of N“ N“ X N‘ . V .
where /:iK —> 5? represents the GF, N is the number of GF taps in the input
f .
layer, and N is the number of GF taps in the first hidden layer. This architecture has
been presented without a GF in the second hidden layer, but such a configuration would be
a straightforward extension.

21
2.2.1.4 Nonlinear state-space model
The GNN uses the Gamma Filter to represent process dynamics. The GF approximates
these dynamics from a Gamma memory kernel basis. The Gamma kernels are able to
model an important class of dynamics but may not be the best representation for general
process dynamics. An alternative approach to using Gamma kernels is to design the ANN
architecture with an explicit state and data flow structure that is capable of learning uni¬
versal process dynamics. This approach is the goal of the nonlinear state-space model
(NLSS) which implements process dynamics directly as a nonlinear state evolution equa¬
tion and an output observation equation [43], as given by
i(o =/(io-i),¿(o,tfg
(20)
y(t) = /(l(0. *K0> fy).
(21)
where 1(7) e 91^ is the models state vector consisting of N* hidden PEs,/* is an ANN
map describing the time evolution of this state, tvv are the weights of this state network,/
is a second ANN map describing how outputs are produced from this state, and $>y are the
weights of this output network. Figure 4 illustrates the configuration of a NLSS network.
Figure 4: Nonlinear state space neural network configuration.

22
2.2.2 Learning Algorithms
The biological roots of neural networks are responsible for the widespread use of the
term learning to describe the process during which the network parameters are changed to
improve the performance of the neural-network-based system. An ANN learning algo¬
rithm specifies how its weights are updated in response to training data. These algorithms
are simply optimization methods applied to the task of finding the best model weights to
minimize a specified modeling objective J, i.e., ArgMin^{J). In general, any of the
optimization methods presented above can be used to solve this problem.
One of the most significant breakthroughs in the field of ANNs was the realization that
the chain rule for ordered partial derivatives provides a mechanism for deriving the first-
order gradients for all weights in a model, even though the modeling objective is only an
explicit function of the model’s outputs [72], Recall that the chain rule for ordered partial
derivatives is given by
iio r XT x° t fix.
(22)
Applying the chain rule allows sensitivities to be calculated from the output of the
model back to its input, which is why the resulting algorithm has been coined “backpropa-
gation” in the literature [40], When the variables are temporally related, the chain rule has
the following form
(23)
Here, in addition to backpropagating sensitivities from the model’s output to its input,
the sensitivities are backpropagated through time.

23
The most common optimization method is simply steepest-descent with momentum,
although many variations have been demonstrated to significantly improve convergence
[10]. The issues leading to the selection of one optimization method over another are:
1) Convergence Rate
2) Implementation Complexity
3) Configuration Complexity
4) Avoidance of Local Minima
5) Sensitivity to Correlation
The most common modeling objective used in ANNs is the mean squared error (MSE)
a.
between the model’s output y and a specified desired response d e 91 , as given by
J =
1
tvW
N' Ñ*
X iwo-MO)2,
t = 1 i = 1
(24)
where TV* is the number of samples in the training dataset. Learning rules which use the
MSE criterion are commonly classified as supervised learning rules, because of the pres¬
ence of a “teacher” implied by the explicit specification of a desired response. Learning
rules without explicit reference to a desired response for the model in the objective func¬
tion are commonly referred to as unsupervised learning rules.
2.2.3 Generalization
As universal mappers ANNs are almost always more complex than the relationships
that we seek to uncover. The net result is that ANNs are notorious for over-fitting a train¬
ing dataset, i.e., performing well on training data but poorly on a blind test dataset [24], It

24
is very important to optimize the complexity of the neural network in order to achieve the
best generalization.
2.2.3.1 Bias and variance
Considerable insight into this phenomenon can be obtained by introducing the concept
of the bias-variance trade-off. Bishop [13] observes that the generalization error £,, using
the Euclidean norm, will depend on a particular dataset D on which the network was
trained. The dependence on D can be eliminated by considering an average over the com¬
plete ensemble of datasets, which can be written as
<; = £0[«á|a>-/(á,d>))2], (25)
where (d\u) denotes the conditional average, or regression, of the desired data given by,
(d\Ú) = fíp(Ü\ñ)dt, (26)
and p(d\u) is the conditional density of the desired variable d conditioned on the input
vector u . Bishop [13] demonstrates that this generalization error can be decomposed into
the sum of the bias squared plus the variance
; = (ED[f(ú,ñ)]-$\ú))2 + ED[f(ñ,ñ)-ED[f(ü,ñ)]]. (27)
A model which is too simple, or too inflexible, will have a large bias, while one which
has too much flexibility in relation to the particular dataset will have a large variance. Bias
and variance are complementary quantities, and the best generalization is obtained when
we have the best compromise between the conflicting requirements of small bias and
small variance. The variance of the prediction will be further addressed below, Section
2.2.4 "Standard Errors.

25
For any given dataset, there is some optimal balance between bias and variance which
gives the smallest average generalization error. In order to improve the performance of the
network further we need to be able to reduce the bias while simultaneously reducing the
variance. The more straightforward way of achieving this is to use more data samples. As
the number of data samples is increased we can afford to use more complex models, hence
reducing the bias, while at the same time ensuring that each model is more heavily con¬
strained by the data., thereby also reducing the variance. If the number of data samples is
increased rapidly in relation to the model complexity we can find a sequence of models
such that both bias and variance decrease. Models such as ANNs can in principle provide
consistent estimators of arbitrary accuracy as the number of data points is increased to
infinity. Note that, even if both the bias and variance can be reduced to zero, the generali¬
zation error will still be nonzero due to the intrinsic noise in the data.
One rarely has infinite data, and practical issues like training time make simply adding
more data points impractical. There are several practical and practiced ways to improve
model generalization, we start with regularization.
2.2.3.2 Regularization
Regularization was originally proposed by Tikhanov [62] as a method for solving ill-
posed problems. The basic idea is to stabilize the solution by means of some auxiliary
nonnegative functional that embeds prior information, e.g. smoothness constraints on the
input/output mapping. Regularization is able to transform an ill-posed problem into a
well-posed problem [48],

26
Tikhanov’s regularization theory uses a regularization penalty term of the form
® - jFfJ2
(28)
where P is a linear (pseudo) differential operator. This penalty term is added to the objec¬
tive function to give
j= y+ro,
(29)
where Y is the regularization parameter. Prior information about the form of the solution
(i.e., the plant) is embedded in the operator P. The operator P is referred to as a stabilizer
in the sense that it stabilizes the solution y, making it smooth.
The appropriate choice for P and the solution to (28) requires functional analysis and
is beyond the scope of this work. The most commonly used form of regularizer, however,
is quite simple to implement. Weight decay regularizer terms consist of the sum of squares
of the adaptive parameters in the network
(30)
where the sum runs over the weights and biases. In conventional curve fitting the use of
this form of regularizer is called ridge regression. It has been found empirically that a reg¬
ularizer of this form can lead to significant improvements in generalization [29].
2.2.3.3 Growing and pruning algorithms
The topology of a neural network, number of units and interconnections, can have a
significant impact on its performance. Regularization helps to minimize this impact when
the complexity of the network is larger than required for the particular application.
Clearly, however, a better approach is to match the complexity of the model with the com-

27
plexity of the application. Various techniques have been developed for optimizing the
topology, in some cases as part of the network training process itself [43]. It is important
to distinguish between two distinct aspects of the topology selection problems. First, we
need a systematic procedure for exploring some space of possible architectures. Second,
we need some way of deciding which of the architectures considered should be selected.
A straightforward approach to network structure optimization involves an exhaustive
search through a restricted class of network topologies. This approach requires significant
computational effort and only searches a very restricted class of network topologies. Much
of the computational burden can be lessened by considering a network which is relatively
small and by allowing new units and connections to be added during training. This
approach was shown to be successful by Bello [10] who used the weights from one net¬
work as the initial guess for training the next network (with the extra weights initialized
randomly). Techniques of this form are called growing algorithms. An alternative
approach is to start with a relatively large network and gradually remove units; these are
known as pruning algorithms. Most of these procedures are ad hoc and tailor to specific
applications, that is not to say, however, that they are ineffective.
More recent work has taken advantage of developments in discrete optimization using
genetic algorithms [36]. Genetic algorithms provide a methodical way of searching large
discrete spaces more efficiently.
2.2.3.4 Cross-validation
An alternative to regularization as a way of controlling the effective complexity of a
network is the procedure of cross-validation [13]. The training of a nonlinear model corre¬
sponds to the iterative reduction of the error function defined with respect to the training

28
dataset. During training, the error will generally monotonically decrease as a function of
the number of presentations of the training dataset, i.e., epochs. However, the generaliza¬
tion error, with respect to an independent dataset called the validation dataset, often shows
a decrease at first, followed by an increase as the network starts to over-fit. Training can
therefore be stopped at the point of smallest error with respect to the validation dataset as
this produces a network with the smallest generalization error (or at least an approxima¬
tion thereof).
2.2.3.5 Committees of networks
In practice, building neural network models requires the training of many different
candidate networks and then the selection of the best performer. Typically performance is
based on the networks performance on a third dataset not used for training or cross-valida¬
tion. There are two disadvantages to this approach. First, all of the effort involved in train¬
ing the remaining networks is wasted, and secondly, the generalization performance on
the validation dataset has a random component due to the noise on the data [13]. The net¬
work which performed the best on this dataset might not be the one with the lowest gener¬
alization error. Recall that the generalization error is averaged over all datasets (25).
These limitations can be overcome by combining the networks together to form a
committee [47][46]. This approach was shown to provide significant improvements in the
generalization error. Denote the committee prediction as
Y(u, Jfr) = ~ £ y,(u, tv(.) ,
A
(31)

29
where tv,- are the weights of committee member i, and jfr = ] is the set of all
weights for the committee.
Bishop [13] shows that if the errors of the individual committee member are decorre-
lated, then the committee will always have a lower generalization error than any of its
individual members.
2.2.4 Standard Errors
Tibshirani [61] reviews a number of methods for estimating the standard error of pre¬
dicted values from a multi-layer perceptron. These include direct evaluation of maximum
likelihoods based on the Hessian matrix, the “sandwich” estimator and the bootstrap
method. Tibshirani offers the following observations:
1) The bootstrap methods provided the most accurate estimates of the stan¬
dard errors of predicted values.
2) The non-simulation methods (delta and sandwich) missed the substan¬
tial variability due to the random initial weights from the multiple train¬
ing runs.
The non-simulation methods are solved analytically, and therefore require unique
solutions for each network topology. Whereas the bootstrap methods apply to all network
topologies, as well as non-neural network paradigms. The additional fact, as noted above,
that the bootstrap methods account for local minima, provides strong argument for their
use.
2.2.4.0.1 Bootstrap methods
Bootstrap methods work by creating many pseudo-replicates (“bootstrap datasets”)
from the training dataset and then reestimating the models weights tv on each bootstrap
dataset. There are two different approaches to bootstrapping [9], One can consider each

30
training case as a sampling unit, and sample with replacement from the training dataset
cases to create a bootstrap sample. This is often called the “bootstrap pairs” method. The
bootstrap pairs sampling algorithm is given by:
I: Generate Nb samples, each one of size TV5 drawn with replacement
* ^ AT5
from the A' training observations {u(i), d(i)}¡ = i, and the b -th sam¬
ple by {ub(i),d\i)}i= ,
II: For each bootstrap sample b e [ 1, //], find
ArgMin^h{jCdb-f{Ãœb,A)} (32)
III: Estimate the standard error of the i th prediction as
- 2
c2(y¡) = -J1— 2 )-y,) , (33)
" - 'y-1
where
Nb
f/ = -7 ^ <34)
On the other hand, one can consider the predictors as fixed, treat the model residuals
A
d
-y as the sampling units, and create a bootstrap sample by adding residuals to the
model fit p. This is called the “bootstrap residuals” approach:
I:
Find ArgMin^{J(d-f(x, tv))} from the /Vs training observations
{ü(i), d(i)}f= i and let r(i) = d(i) -f(ú(i), tv)

31
II: Generate A^ bootstrap samples, each one of size ]\f drawn with
replacement from {r(i)= x, and the b th sample by {r (/)},■ = ] let¬
ting
yb = {Añ,ñ) + rb(i)}f=l (35)
III: For each bootstrap sample b e [ 1, A^], find
ArgMin^h{J(yb -f(u, Tv¿))} (36)
IV: Estimate the standard error of the i th prediction as
V
^(F/) - -J— X (W -yt) (37)
A' _ = i
Note that both of these methods require fitting a model (retraining the network) V*
times. Typically A^ is in the range 20 < 7Va < 200. In simple linear least squares regres¬
sion, it can be shown that the bootstrap methods both agree with the standard least squares
formula as A^ —» co.
The bootstrap methods will arrive at confidence intervals
y¡ ~ cconf2(y¡) - y i - cCOnf¡2 (y i). (38)
where cconj- depends on the desired confidence level 1 - a . The factor cC0nj- can be taken
from a table with the percentage points of the Student’s t -distribution with the number of
degrees of freedom equal to the number of bootstrap runs A^.

32
2.3 Control Theory
An ANN is capable of modeling any process making them ideal candidates for com¬
plex process optimization and control strategies. Neurocontrol is but a sub-field of classi¬
cal control theory [31]. To put neurocontrol in perspective, it is important to consider its
place within this field.
2.3.1 Classical Control Theory
Classical control theory is strongly biased towards linear time-invariant systems [31].
General nonlinear systems simply do not allow us, because of their analytical intractabil¬
ity, to formulate a theory that is as strong as that of linear system theory. On the other
hand, nonlinear systems can be qualitatively similar to linear ones under some circum¬
stances.
2.3.1.1 Linear control
Linear control is concerned with systems of the form
x = + Éú (39)
with state x, input u, measurable output
P = £*, (40)
and controllers of the form
u=-Fx + Óx*, (41)
where is the reference state, that is, the state to which the plant is to be brought with the
help of the controller [33].
The goals of linear control are:
1) Altering the closed-loop behavior of the system to some user-defined
response characteristics.

33
2)Controlling the closed-loop stability, i.e., convergence back to an equi¬
librium point after disturbance.
The disadvantages of linear control designs:
1) Assume that the world is linear Gaussian and stationary, when in reality
the world is none of the above.
2) Require complete a priori knowledge of the process dynamics.
3) Require that the process is controllable and observable.
4) Cannot follow a reference trajectory produced by a system of lower
order that the process.
2.3.1.2 Robust control
Robust control addresses the problem of controlling a plant whose behavior is slightly
different from that of a plant model [37], The reasons for the difference are predominately
the effect of the nonlinear, non-Gaussian and nonstationary world. A popular pragmatic
classical approach to robust control is concerned with preserving stability [2]. The closed-
loop eigenvalues are chosen so that they remain in the stability region even if the plant
model should change in a defined range.
Although robust control strategies are primarily designed to compensate for differ¬
ences between our linear time-invariant assumptions and the real world, they are still
developed based predominately on linear system theory. They are therefore, not able to
cope with significant deviations from these assumptions.
2.3.1.3 Adaptive control
Adaptive control is another way to reach a goal similar to that of robust control [37].
Instead of designing robust controllers that work under conditions different from those for
which they have been designed, adaptive controllers recognize the difference between the
assumption and reality and change to perform better in the new conditions.

34
Adaptation schemes can be based on both a reference model and a cost functional [2].
The approach called model reference adaptive control (MRAC) is, by its name, committed
to the former. This approach is based on formulating the rules for computing the direction
of change of controller parameters as a function of the difference between the behavior of
the closed-loop system and a reference model. Controller parameters can be adapted either
directly or via the estimation of plant model parameters.
A more general approach is that of self-tuning regulators (STR) [3] which consists in
adaptive estimation of a plant model and applying a formalized controller design method
to the plant model. This design method can be based on cost function optimization.
Like robust control, however, adaptive control implementations have been based on
linear, or simple nonlinear parametric assumptions, about the process. As a result, adap¬
tive control designs have not demonstrated significant successful with complex real-world
processes.
2.3.1.4 Nonlinear control
Nonlinear control theory is concerned with general systems of the form [31]
i = /(*, Ó) (42)
with measurable output
P = g(*) (43)
and controllers of the form
úc = c(^, £*). (44)
The general formulation of nonlinear control holds promise to overcome all of the lim¬
itations of the classical control schemes presented above. The fields track record, how¬
ever, does not deliver on this promise. The problem is that an analytical solution is known

35
only for a restrictive subclass of nonlinear systems. The difficulties with genuine nonlin¬
ear controller designs have typically lead to a linearization approach, also known as gain
scheduling.
2.3.1.5 Optimal control
The topic of optimal control theory is to design controllers that are optimized to a cer¬
tain performance criterion [15]. Classical optimal control has primarily focused on appli¬
cations where such optimality could be proven analytically. For example, for a linear plant
and a quadratic performance criterion the Ricatti controller represents an explicit and glo¬
bal solution [31]. Dynamic optimization provides another example for state evaluation and
selection of the optimal action, which can be proven optimal in certain applications. Alter¬
natively, if each state at each sampling period is represented by a node in a directed graph
and actions are represented by connecting edges of the subsequent states, then the task can
also be transformed to the critical graph problem of graph theory [15].
In its most general form, the optimal control problem can be formulated as an optimi¬
zation problem [15]. The plant can be generalized as in (42) and the goal is to find a con¬
troller described by (44) that minimizes
£[/(**-*)], (45)
where E[ ] is the mean value over time.
2.3.2 Neurocontrol
Most neurocontrol architectures are either explicit or disguised analogies of classical
control design such as optimal control or numerical lyapunov-function-based design
methods. It has been argued [14] that it is only the representation of functions by neural

36
networks that defines the field of neurocontrol in the broad sense and differentiates neuro¬
control form classical control methods.
The author agrees that the use of nonparametric models does differentiate neurocon¬
trol from classical controls, but would argue that the primary departure from, and exten¬
sion to the potential of, control theory is neurocontrol’s willingness to depart from a
requirement for analytic solutions. Neurocontrol designs seek to realize the promise of
general nonlinear control replacing analytic optimization methods with numeric ones.
Researchers in the field of neural networks are accustomed to working in an intractable
world, and have been willing to resolve important questions like stability, robustness and
consistency empirically. The result has been, and will continue to be an important exten¬
sion to control theory.
There are many types of neurocontrol architectures in the literature, each with specific
advantages and disadvantages. The following sections review some of the conceptual neu¬
rocontrol strategies which have been proposed in the literature.
2.3.2.1 Model-predictive control
When a model is used indirectly and offline the control scheme is usually referred to
as model-predictive control (MPC) [37], In most industrial process control applications a
priori knowledge about the process is hard to obtain and black-box models must be used.
The offline training phase performs supervised learning to develop an ANN model for the
process to be controlled, i.e., the ANN attempts to mimic the process after being exposed
to actual process data. This phase can be stated as
ArgMin${J(f -/(if, tv))},
(46)

37
where and up are the process outputs to be controlled and inputs to be manipulated,
respectively.
At the online control phase, the ANN model cannot be used alone; it must be incorpo¬
rated with a model-based control scheme [31]. This control scheme is once again an opti¬
mization problem, which can be stated as
ArgMin^Jip* -/(if, *>))}, (47)
where y* g denotes the desired closed-loop process output. This optimization is per¬
formed repeatedly at each time interval during the course of feedback control.
2.3.2.2 Model-inverse control
An ANN can be trained to develop an inverse model of a process [40]. Here, the
model’s input is the process output, and the model’s output is the process input. The
offline training phase can be stated as
ArgMin^Jitf -f(f, . (48)
Clearly, the inverse model is a steady-state model or the resulting controller would be
non-causal. Given a desired process setpoint y* , the appropriate online control signal tip
can be immediately calculated as
(49)
Successful applications of inverse modeling are discussed in [40] and [58]. Obviously,
an inverse model exists only when the process behaves monotonically as a “feed-forward”
function at steady-state. If not, this approach is not applicable.

38
2.3.2.3 Controller modeling
Another simple direct neurocontrol scheme is to use a neural network to model an
existing controller. The input to the existing controller is used as training input to the ANN
model, and the controller output serves as the desired response. This approach is similar to
the model-inverse control except that the desired response here is not a process but a con¬
troller. This approach can be formulated as
ArgMin^jtf (50)
where uc are the decision variables generated by an existing controller in response to the
plant states y?.
Like a process, a controller is generally dynamic and often comprises integrators or
differentiators. If an algebraic feed-forward network is used to model the existing control¬
ler, dynamic information must be explicitly provided as input to the ANN model.
In general, this approach can result in controllers that are faster and/or cheaper than
traditional controllers. Using this approach, for example, Pomerleau [50] presented an
intriguing application where a neural network was used to replace a human operator, i.e.,
an existing controller.
2.3.2.4 Model-free direct control
Without an existing controller or process knowledge, controllers have to be adapted or
learn the way a human operator leams to control/operate a process for the first time. A
model-free neurocontrol design objective can be stated as
ArgMin^c{J(y* -f), tf =/(xp, tvc)},
(51)

39
where f is an ANN that is directly controlling the process inputs, and tvc are the weights
of this network. Notice that the optimization criterion J is only a function of the actual
and desired process outputs. This means that the optimization methodology employed
must be able to learn tvc without an explicit desired response or even a mathematical link¬
age to the criterion.
The key feature of this direct adaptive approach is that a process model is neither
known in advance nor explicitly developed during control design. This most common
learning algorithm for this type of control design is referred to as reinforcement learning.
The first work in this area was the “adaptive critic” algorithm proposed by Barto et al. [7].
Such an algorithm can be considered as an approximate version of dynamic programming
[73][8], later coined as Neuro-Dynamic Programming [12].
Despite its historical importance and intuitive appeal, model-free adaptive neurocon¬
trol is not appropriate for most real world applications. The plant is most likely out of con¬
trol during the learning process, and few industrial processes can tolerate the large number
of failures required to adapt the controller.
2.3.2.5 Model-reference direct control
From a practical perspective, one would prefer to let failures take place in a simulated
environment with a process model rather than in a real plant. Even if failures are not disas¬
trous they can cause substantial losses. The performance of a controller could be evaluated
based on a model for the process, rather than the process itself. The training stage of the
control design can be given by
ArgMin ^m{J(yp-f\u, tv'"))},
(52)

40
and the control design becomes
ArgMin^c{J(y* -f\u , tv'")), úc =/(*?, tvc)}. (53)
In the course of modeling the plant, the plant must be operated “normally” instead of
being driven out of control. After the modeling stage, the model can be used for controller
design. If a process model is already available, an ANN controller can be developed in a
simulation in which failures cannot cause any loss but that of computer time. A neural net¬
work controller after extensive training in the simulation can then be installed in the actual
control system.
Model-Reference direct control schemes have not only proven effective in several
studies [41 ][63], but have also already produced notable economic benefits [60]. These
approaches can be used for both off-line control and for on-line adaptation.
2.4 NOx
The Clean Air Act Amendments of 1990 require that electric utilities make significant
reductions in nitrogen oxide (NOx) emissions from their fossil-fired power plants. To
date, most efforts to reduce NOx emissions have come from expensive hardware retrofits
with less than satisfactory performance. Further complicating matters, conditions that
decrease NOx formation (lower temperature, excess fuel) result in the formation other pol¬
luting compounds, mainly carbon monoxide (CO). Similar emissions reductions are being
required in Europe through local and European Economic Community (ECC) initiatives.
Nitrogen monoxide (NO) and nitrogen dioxide (N02) are by-products of the combus¬
tion process of virtually all fossil fuels. Historically, the quantity of these inorganic com¬
pounds in the products of combustion was not sufficient to affect boiler performance; their

41
presence was largely ignored. In recent years, oxides of nitrogen have been shown to be
key constituents in the complex photochemical oxidant reaction with sunlight to form
smog. Today, the emission of N02 and NO (collectively referred to as NOx) is regulated
by the 1990 Clean Air Act Amendments and has become an important consideration in the
design of fuel firing equipment.
NOx is formed by two primary mechanisms: thermal NOx and fuel-bound NOx. Ther¬
mal NOx formation occurs only at high flame temperatures when dissociated nitrogen
from combustion air combines with oxygen atoms to produce oxides of nitrogen such as
NO and N02. The formation of thermal NOx increases exponentially with combustion
temperature and increases by a square-root relationship with the presence of oxygen in the
combustion zone. Fuel-bound NOx formation is not limited to high temperatures, but is
dependent upon the nitrogen content of the fuel. The best way to minimize NOx formation
is to reduce flame temperature, reduce excess oxygen, and/or to bum low nitrogen-con¬
taining fuels. Conditions that decrease NOx formation (lower temperature, excess fuel)
can result in incomplete combustion. These conditions result in the formation other pollut¬
ing compounds, mainly carbon monoxide (CO).
2.4.1 Reduction
The available NOx reduction technologies can be categorized into one of the follow¬
ing:
• Before Combustion: Nitrogen is extracted from the fuel. This is relatively inef¬
fective, since most of the nitrogen in the formation of NOx comes from the air
(containing N2).

42
• After Combustion: NOx is chemically reduced before leaving the stack. This pro¬
cess is also expensive, requiring hardware retrofits.
• During Combustion: Altering fuel and air flows and introducing them at different
points of the furnace can create several zones with different temperatures and
stoichiometry. These parameters significantly effect the rate of NOx formation.
The following section reviews available NOx reduction strategies and technologies for
combustion sources.
• Fuel Switching: Fuel-bound NOx formation is most effectively reduced by
switching to a fuel with lower nitrogen content. No. 6 fuel oil or another residual
fuel having a relatively high nitrogen content can be replaced with No. 2 fuel oil,
another distillate oil or natural gas (which is essentially nitrogen-free) to reduce
NOx emissions.
• Flue Gas Recirculation (FGR): Flue gas recirculation involves extraction of some
of the flue gas from the stack, and recirculation with the combustion air supplied
to the burners. The process reduces both the oxygen concentrations at the burn¬
ers and the temperature by diluting the combustion air with flue gas. CO can
become a significant problem here.
• Low NOx Burners: Installation of burners especially designed to limit NOx for¬
mation can reduce NOx emissions. Higher reduction efficiencies can be
achieved by combining a low NOx burner with FGR. Low NOx burners are
designed to reduce the peak flame temperature by inducing recirculation zones,
staging combustion zones, and reducing local oxygen concentrations.

43
• Derating: Some industrial boilers may be derated to produce a reduced quantity
of steam or hot water. Derating will decrease the flame temperature within the
unit, reducing formation of thermal NOx. Derating can be accomplished by
reducing the firing rate or by installing a permanent restriction, such as an orifice
plate, in the fuel line. Clearly this solution would have significant economic
impact on the unit.
• Steam or Water Injection: By injecting a small amount of water or steam into the
immediate vicinity of the flame, the flame temperature will be lowered and the
local oxygen concentration reduced. The result would be to decrease the forma¬
tion of thermal and fuel-bound NOx. This process generally lowers the combus¬
tion efficiency of the unit by one or two percent.
• Staged Combustion: Either air or fuel injection can be staged, creating either a
fuel-rich zone followed by an air-rich zone, or an air-rich zone followed by a
fuel-rich zone. A low NOx burner utilizing staged combustion can be installed,
or the furnace itself can be retrofitted for staged combustion.
• Fuel Rebuming: Staged combustion can be achieved through the fuel rebuming
process. A Gas Rebuming Zone (GRZ) is created above the primary combustion
zone. In the GRZ, additional natural gas is injected, creating a fuel-rich region
where hydrocarbon radicals react with NOx to form molecular nitrogen.
• Reduced Oxygen Concentration: Decreasing excess air reduces the oxygen avail¬
able in the combustion zone and lengthens the flame, resulting in a lower heat
release rate per unit flame volume. NOx emissions are reduced in an approxi-

44
mately linear fashion with decreasing excess air. However, as excess air is
reduced beyond a threshold value, combustion efficiency will decrease due to
incomplete mixing, and CO emissions will increase.
• Selective Catalytic Reduction (SCR): Selective catalytic reduction (SCR) is a
post-formation NOx control technology that uses a catalyst to facilitate a chemi¬
cal reaction between NOx and ammonia to produce nitrogen and water. An
ammonia/air or ammonia/steam mixture is injected into the exhaust gas, which
then passes through a catalyst where NOx is reduced. To optimize the reaction,
the temperature of the exhaust gas must be in a certain range when it passes
through the catalyst bed. Among its disadvantages, SCR requires additional
space for the catalyst and reactor vessel, as well as ammonia storage, distribu¬
tion, and injection system. Precise control of ammonia injection is critical. An
inadequate amount of ammonia can result in unacceptable high NOx emission
rates, while excess ammonia can lead to ammonia “slip”, or the venting of unde¬
sirable ammonia to the atmosphere.
• Selective Non-Catalytic Reduction (SNCR): Selective non-catalytic NOx reduc¬
tion involves injection of a nitrogenous agent, such as ammonia or urea, into the
flue gas. The optimum injection temperature when using ammonia is 1850
degrees F, at which 60 percent NOx removal can be approached. The optimum
temperature range is wider when using urea. Below the optimum temperature
range, ammonia is formed, and above, NOx emissions actually increase. The
success of NOx removal depends not only on the injection temperature, but also

45
on the ability of the agent to mix sufficiently with flue gas.
2.5 Fossil-Fired Power Generation
In general Canal Unit 2 is a large fossil fuel combustion engine. From an abstract per¬
spective, the combustion process takes in air and fuel, and produces energy and exhaust;
as described by:
1) Air: Fossil fuel combustion requires air, or more specifically the oxygen
contained in air. Subsystems within the plant measure, prepare and
introduce this air.
2) Fuel: Combustion also requires fuel. In the case of Canal Unit 2, the fuel
can be either #6 residual oil (leftover from the refining process) or natu¬
ral gas. Canal Unit 2 can fire oil only, gas only or a mixture of both.
Both fuels must be measured, prepared and introduced to the furnace.
3) Energy: The energy released by the oxidation of fossil fuels during com¬
bustion is used to make steam. The properties of water still make it the
best choice when converting thermal energy to work. Canal uses the
radiative and convective heat from the combustion process to transform
ultra-clean water into superheated steam. The expansion of this steam is
used to turn a turbine that turns a coil in a magnetic field, producing
electric potential. The steam, having done this work flows through
ocean water filled condensers that convert it back to super-clean water.
4) Exhaust: The gaseous products of combustion having contributed much
of their heat content to the production of steam are cleaned electrostati¬
cally and ejected into the atmosphere.
2.5.1 Process Variables
The specific process variables as they apply to the Canal generating unit are described
in more detail in the following sections. These variables are also listed in Appendix, and
will be referred to throughout this work.
2.5.1.1 Air
The air required for fossil fuel combustion is prepared and introduced in two ways.
Two large symmetrical Fans called Forced Draft fans push ambient air through a series of

46
preheaters that warm this air to between 80 and 180 degrees F. This hot oxygen rich air is
then pressed into a windbox that surrounds the furnace enclosing the burner ports.
Through an array of vents called Primary and Secondary Air Shrouds around each burner
and through secondary ports called Overfire Air Ports this pressurized air is vented into
the combustion zone. In addition to this oxygen rich air, Canal Unit 2 has the ability to
recycle exhaust gas into the combustion zone through a Gas Recirculation System.
The measurements of all this air are a function of boiler design, and fan capacities. To
increase the output of the engine, additional air must be throttled through these devices.
2.5.1.1.1 Forced draft system
The forced draft fans are 2500 horsepower, 624,000 cfm centrifugal fans with inlet
vane throttles. They are constant speed fans meaning that the fan shaft turns at a constant
speed while more or less air with more or less initial spin can be dumped into the blades
by opening or closing the vanes. If the vanes are only slightly opened the flow volume of
air available to the fans is small, and it takes less work to move it. Fan amps will be corre¬
spondingly low. If the inlet vanes are opened wider the flow is greater. Still the fan moves
at a constant speed. More work is being done, and the amperage must increase. The output
of the FD fans is derived from the boiler master signal. Forced draft output is specified
along with fuel flow by the fuel-air curve of the boiler. The fuel-air curve gives a total air
flow requirement, as well as a total fuel flow requirement for a given load.
These fans are symmetrical to the furnace like many other systems and they operate
symmetrically, through their respective ducts unless biased. Bias represents an addition or
subtraction of signal to the B side FD fan. These fans can also be trimmed to meet slightly
less or slight more than the Total Air Flow demanded by the fuel-air curve of the Boiler.

47
The FD fans are the principal air throttles of the Boiler and so have a fundamental effect
on nearly every other system.
2.5.1.1.2 Forced draft fan inlet vanes
Since the inlet vanes’ positions represent the work being done by the fan and are the
control most familiar to the operators, these tags were used to represent the FD fans.
2.5.1.1.3 02 trim
This tag represents the bias that operators set into the airflow demand predetermined
by load. Functionally this control trims the response of the FD Fan to Air Demand. This
tag gives the operators the ability to run the furnace slightly lean or rich overall.
2.5.1.1.4 Induced draft system
As mentioned in the section on the Forced Draft Fans, the function of the ID Fans is to
take whatever gasses are present in the furnace, including air that has been introduced by
the load following Forced Draft Fans, plus all products of combustion, and pull them out,
maintaining a constant under pressure in the furnace of -.5 inches of water column. The
FD Fans’ speed is kept constant while the volume of air they move is throttled with inlet
vane controls. Canal is limited by the power of these fans. Current unit maximum output is
frequently limited by the power of these fans to keep up with the increased air flows of the
recently installed low NOx shroud and overfire air system.
2.5.1.1.5 Induce draft fan inlet vanes
The induced draft fan inlet vanes are the inlet throttles to the fans, they open in
response to request for increased output and as in the case of the FD fans, represent the
work being done.

48
2.5.1.1.6 Combustion air temps
The combustion air temperature tags represent the temperature of the incoming air
after the FD Fans. The temperature of this air is a direct result of energy added to ambient
air by the Glycol Air Preheater (GAH), and the Combustion Air Preheater (CAFI). Since
density is a function of temperature, the temperature of this air can impact the combustion
process that is sensitive to the Oxygen content of air as well as the operation of other vol¬
umetric systems like the Induced Draft (ID) Fans. It also has a primary impact on exhaust
gas temperature and resultant stack gas velocity.
2.5.1.1.7 Primary air shrouds
The Primary Air (PA) Shrouds represent the circular articulating vents that surround
the individual burner orifices. These are closest to the fuel gun concentrically inside of the
Secondary Air (SA) Shrouds. They are responsible for supplying primary combustion air
to the flame front. These tags represent actuator positions.
The PA shrouds are controlled by the Burner Management System (BMS) and they
move as a group from minimum position (5% open to provide cooling air) toward open as
load increases. The signal that controls them is called the Primary Air Master Demand
(PAMD). Separate PAMD signals exist for fuel gas primary air demand and for fuel oil
primary air demand. Each specific burner effectively listens to the current fuel state. Hav¬
ing received this signal each burner’s own PA shrouds responds to the PAMD in accor¬
dance with one of two functions that are unique to it - a Burner Primary Air Shroud
Function for oil operation and a Burner Primary Air Shroud Function for gas operation.
The correct unique local shroud function is changed according to the correct master signal
depending on the fuel state of the burner. These burner and fuel specific response func-

49
tions were set up to give roughly appropriate air flow to combustion at all load points and
fuel states based on the air flow inherent to the furnace.
Aside from normal operation the PA shrouds can be biased from the fuel specific mas¬
ter signal or on an individual basis from their respective unique functions.
2.5.1.1.8 Secondary air shrouds
The secondary air shroud tags represent the broadcast actuator positions of the second,
outer concentric set of circular articulating vents that surround the individual burner ori¬
fices.
The first function of the Secondary Air (SA) Shrouds is to introduce combustion air to
the flame front following load. Their second function is to balance windbox pressure, and
therefore total airflow, against the actuation of the Overfire Air Ports and the PA shrouds.
The SA Shrouds have a master signal against which a master bias can be set. In addi¬
tion they have individual actuating functions and individual biases that can be set against
these individual functions.
2.5.1.1.9 Over fire air ports
The Overfire Air (OFA) ports are rectangular louvered ports that pass combustion air
from the Windbox to the Furnace above the top burner level. In doing this they re-oxygen-
ate the oxygen depleted flame front. The tags themselves represent the positions broadcast
from the actuators that control the articulating louvers.
The OFA ports were installed as a part of the low NOx retrofit of 1996. The Forney
low NOx burner system is designed to bum more coolly and incompletely than normal.
NOx formation has been positively linked with time exposure to higher temperatures.
After partial combustion has taken place, low in the flame front, extra oxygen rich com-

50
bustion air is introduced through the OA ports to complete the process. In this way the low
NOx burner system stages off-stoichiometric combustion to manage combustion products.
The OA port actuators receive their master signal from load. This signal can be biased.
Each actuator’s response is based on a unique function that was parametrically deter¬
mined, in concert with the Primary and Secondary Air Shrouds during installation to give
best airflow to combustion at all load points.
2.5.1.1.10 Air preheater temps
These represent the temperature of the exhaust gasses entering and leaving the ljung-
strom combustion air heat exchanger. The ljungstrom is a large (30 ft. dia.) rotating wheel,
arranged perpendicular to the gas flow. It is half enclosed by the exhaust ducts and half
enclosed by fresh air ducts. As this wheel slowly rotates, heat is absorbed by a given area
of the wheel exposed to exhaust gas. The absorbed heat is then imparted to the incoming
air while that same section traverses the fresh air duct. Elaborate seals and pressurized
sealing air keep the two gasses from mingling across this device.
The air preheater tags are somewhat redundant. The “In Temps” represent the temper¬
ature of the gas on its way in, while the “Out Temps” represent the temperature of the gas
on the way out. The heat exchange of the air preheater is a function of the device and of
the temperatures of the two gasses and is not controllable in the least. The gas temp after
the air pre heater heat exchange was a more familiar control to the operators, however our
ability to collect these signals was compromised by a failing thermocouple during a large
part of the data collection for phasel. The gas temp before the air preheater was used to
represent exit gas temp for the modeling instead.

51
2.5.1.1.11 Windbox and furnace
The windbox is an enclosed volume that surrounds the waist of the furnace and the
burner openings. Preheated, oxygen rich air is pressurized in this volume by the FD fans.
From here this air can pass only into the furnace and only through vanes that surround the
burner openings called primary and secondary sir shrouds, or through the overfire air ports
above the burners. Canal Unit 2 is a balanced draft furnace which means that air flow
through the furnace is controlled around a desired furnace pressure by both pushing and
pulling fan systems. The pushing fans are the FD fans, while the pulling fans are the
induced draft fans. The FD fans have the primary responsibility of getting the combustion
zone all the oxygen it requires. The introduction of this pressurized air is accomplished not
only by positively pressurizing the windbox but also by negatively pressurizing the fur¬
nace. With the windbox driven to a positive pressure and the furnace kept at a fixed rela¬
tive negative pressure, the velocity of combustion airflow is assured. The induced draft
fans have primary responsibility for maintaining the furnace at a negative pressure relative
to the windbox. In the course of increasing unit output the FD fans increase air flow. Their
aim is to maintain windbox pressure at +2 inches of water column while air transfer to the
furnace increases through the widening overfire ports and primary and secondary air
shrouds. The induced draft fans, trying to maintain a constant pressure of-.5 inches of
water column in the furnace despite this increasing flow of air from the windbox, also
ramp up. The opposite happens for decreasing load. When the forced draft fans decrease
their output in step with the fuel-air demand, air flow from the windbox to the furnace
decreases. In order to maintain a constant -.5 inwc in the furnace the induced draft fans

52
throttled back. Transient changes in the windbox to furnace pressure differential can also
produce automated changes in the FD and ID fan flows.
2.5.1.1.12 WindBox pressure
This tags represents the positional average windbox air pressure. It is controlled
around + 2 inwc
2.5.1.1.13 Furnace pressure
This tag represents the actual furnace air pressure.
2.5.1.2 Fuel
The Fuel required for Combustion may be either #6 Fuel Oil or Natural Gas. In both
cases the fuel is taken from storage, filtered, heated to greater or lesser degree, pressur¬
ized, and injected. In the case of #6 Fuel Oil, the temperature required to achieve a pump-
able consistency is usually around 200 degrees. Natural Gas comes from high pressure
transmission lines and once stepped down to usable pressure is warmed up to around 80
degrees F. Both fuels are then pressurized in their respective headers. It is from these
headers that burners, when they are lit, tap their fuel.
2.5.1.2.1 Burners-on/fuel
These tags represent the readings of an array of air cooled optical flame scanners
located in the furnace itself that observe the respective burner flames. Since each burner
can fire either natural gas or fuel oil. A scanner, calibrated for each fuel specific flame is
permanently assigned to each burner. Although these scanners are analog devices, their
primary function is to confirm that the flame emanating from each lit burner is of a thresh¬
old quality. If the flame they are monitoring is not of a threshold quality the scanner has

53
the will to declare a Master Fuel Trip and cut off all fuel to the furnace. This is to prevent
the introduction of unbumed fuel to the furnace. These are analog devices but because
they are calibrated with the single purpose of either positively or negatively confirming
this threshold they essentially read either 1 or 0. This specific set represents the flame
quality of its burner if that burner is on natural gas.
2.5.1.2.2 Burner cells 1-8A & 1-8B MN gas flame
These are the signals for gas flame status for each burner.
2.5.1.2.3 Burner cells 1-8A & 1-8B MN oil flame
These are the signals for oil flame status for each burner.
2.5.1.2.4 Fuel type
As the Boiler Master request increased output BTUs are requested from the Fuel Sup¬
ply Systems. As a default this request is divided evenly in proportion to burners in service,
each of which have BTU content per unit of fuel settings. The total BTUs entering the fur¬
nace via the burners in service must equal this demand.
2.5.1.2.5 Fuel oil
The fuel oil introduction system consists of a main pressure generating pump that
ramps up in output as the unit master demand requests more output in the form of BTUs.
This pump supplies an operating pressure to the fuel oil header. All oil burners once they
are lit and placed into service tap a fixed orifice from this header. Since fuel oil pressure is
fixed by the number of BTUs requested by load, and the orifice of each burner tip is a
fixed diameter if open, the number of burners in service will dramatically affect Fuel Oil

54
Pressure. Changes in the number of burners lit can vary the fuel oil pressure in the header
between 65 and 150 PSIG.
Fuel Temp Fired must be at least the temp required for pumpability, which is specific
to the viscosity of the fuel oil being used.
2.5.1.2.6 Natural gas
In a fashion similar to the fuel oil introduction system, the unit master demand
requests BTUs from the gas system. Fuel gas from the pipeline is stepped down to operat¬
ing pressure, filtered, warmed and supplied to a main gas header. All gas burners when lit
tap a fixed orifice from this header. The number of burners lit on gas can affect the actual
gas pressure indicated at the header.
2.5.1.2.7 Burner atomization
These tags represent the essential fuel oil atomizing steam parameters. Atomizing
steam is dry superheated steam extracted from the turbine or the reboiler and injected into
the oil burner tips to atomize the fuel oil as it is introduced to the combustion zone.
Burner Atomizing Steam pressure runs at a specified 20psig over fuel oil pressure.
Burner atomizing steam flow is modulated to maintain this constant difference from fuel
oil pressure while the actual temperature fluctuates somewhat at the point of extraction.
pv = nrt connects these three variables with temperature being somewhat variable, flow
being the control, and pressure being the set point.
2.5.1.2.8 Fuel oil / fuel gas flow differential
This tag represents the ratio of BTUs contributed by the fuel oil system vs. the BTUs
contributed by the fuel gas system to the total BTUs required for a given load.

55
2.5.1.2.9 Energy
During operation at Canal Unit 2 feedwater, pressurized by a large parasitic turbine
driven pump, is circulated through series of preheaters and then through the very walls of
the furnace. During this passage it is converted to steam. This steam is then collected in a
pressure vessel called a Steam Drum located at the top of the boiler where it is “dried”.
From the Steam Drum this dry saturated steam is passed through radiator like Primary
Superheater and Secondary Superheaters that hang at the top of the furnace where convec¬
tive, and radiative heat transfer occurs. From the outlet of the Secondary Superheater the
steam goes directly to the High Pressure inlet of the Turbine. Unit 2 is a single reheat
boiler which means that the exhaust from the high pressure turbine, instead of being con¬
densed, is passed back to the boiler and re superheated. This re superheated steam then
turns the Intermediate and Low Turbine Stages. Attemperating sprays inject cool feedwa¬
ter into the steam cycle between the Primary and Secondary Superheaters and also before
the Reheat Superheater. These cooling sprays dampen thermal dynamics and keep steam
temperature at the turbine roughly constant around 1000 degrees.
2.5.1.2.10 Generation
The Westinghouse turbine generator converts the expansion energy of superheated
steam to create rotational momentum in the turbine. This rotational energy is imparted to a
coil enclosed in an induced electromagnetic field. The rotation of this coil in this excited
field creates electric potential at the ends of the coil. This electric differential has roughly
560 megawatts of power with which to do work. Under normal operating conditions, and
aside from throttling effects, the output of the turbine generator is in direct relationship to
boiler output.

56
This tag represents the actual instantaneous unit output in units of power.
2.5.1.2.11 Heat rate
This is a simple calculated tag representing the sum of BTUs flowing into combustion
from oil and gas combined divided by the amount of power created. It can show the rela¬
tive efficiency of combustion-steam-power system in an energy in vs. energy out relation¬
ship. As load increases heat rate decreases due to the thermal properties of the steam loop.
2.5.1.2.12 Main steam
The main steam temperature, in concert with the throttle pressure is related via steam
tables to volume, enthalpy and entropy and describes the output state of the steam generat¬
ing system. Unit 2 is a sliding throttle unit capable of modulated steam temp output across
different throttle valve configurations. Steam output is essentially controlled by flow. As
the unit ramps up in load, more steam is generated from increased combustion. Steam
temperature is held (roughly) steady via modulation of flow through the turbine throttle
valves, which are sequentially opened. Once the unit reaches a certain level of output
(@480MW) all throttle valves are set in the fully open position and steam flow is modu¬
lated by continuing to increasing steam output through combustion throttling. At all levels
of output Steam temperature is controlled around 1000 degrees F for optimum turbine
operation
2.5.1.2.13 Temperature
These tags represent the temperature of superheated steam as it exits the secondary
superheater header and heads to the high pressure turbine throttle valves.

57
2.5.1.2.14 Attemperation spray
These represent the amount of cool feedwater that is sprayed into main steam between
the primary and secondary superheaters to control the temperature of the steam at the sec¬
ondary superheater outlet to the turbine.
U28300 represents fine control. This valve responds automatically and in analog fash¬
ion to all changes in steam temperature at the secondary superheater outlet. U28301 repre¬
sents bulk control. It responds only to changes in SSH outlet temp that are exceed preset
deadband. These coarse and fine cooling controls are combined to dampen and control
steam outlet temp against oscillations or imbalances inherent in the steam system.
All desuperheating sprays receive their volume of feedwater from total feedwater
flow.
2.5.1.2.15 Reheat steam
Exhaust from the high pressure turbine stage is cycled back to the furnace via the
reheat steam loop where it is sprayed then re-introduced to heat exchange in the reheat
superheater. Through the reheat superheater this steam is brought back up to lOOOdegF
and 580pisa upon which it is sent to the intermediate stage of the turbine. Exhaust from the
intermediate stage turbine flows to the low pressure turbine stage.
2.5.1.2.16 Temperature
This temperature represent the temperature of re superheated steam as it heads to the
intermediate turbine stage inlet.
2.5.1.2.17 Attemperation sprays
These sprays function like the superheater sprays. They inject relatively cool feedwa¬
ter into the reheat steam after it has been extracted from the turbine and before it is

58
reheated. They function to control the temperature of the steam at the outlet of the reheat
superheater. Unlike the superheat desuperheaters, these sprays do not have separate fine
and coarse control functions.
2.5.1.2.18 Furnace metal temps
These tags represent an array of thermocouples installed on the vertical legs of the
pendant superheaters. Especially in gas burning the fire side material temperature of these
heat exchangers can become problematic. Unit 2 has an especially large area of super¬
heater, which is the heat exchange closest to the fire itself. Because gas bums at a cooler
temperature than oil less radiant heat is absorbed by the waterwalls of the furnace and for
the same output of steam more heat must be passed to the steam loop through the gas
stream and the superheaters. This superheater weighted heat transfer zone in gas burning,
combined with air flow stratification that seems to be inherent to this unit, make careful
monitoring of these thermocouples necessary. Extended temps above 1100 degrees can
increase material fatigue signifigantly.
2.5.1.2.19 Secondary superheater metal temps top-bottom L-R
These represent the temperature of the firesides of selected evenly spaced legs of the
secondary superheater, which encounters hot gas second, after the primary superheater.
They are alphabetized horizontally across the superheater surface with upper representing
the trailing side and lower representing the leading side.

59
2.5.1.2.20 Primary superheater metal temps top-bottom L-R
These represent the temperature of the firesides of selected evenly spaced legs of the
primary superheater, which encounters hot gas first and is closest to the flame front. They
are alphabetized horizontally across the superheater surface.
2.5.1.2.21 Reheat superheater metal temps top-bottom L-R
These represent the temperature of the firesides of selected evenly spaced legs of the
reheat superheater, which encounters hot gas third, after the secondary superheater and
before the feedwater economizer. They are alphabetized horizontally across the super¬
heater surface with upper representing the trailing side and lower representing the leading
side.
2.5.1.2.22 Exhaust
The gasses created by combustion flow upward through the furnace gas path across
the primary and secondary superheaters, the reheat superheater, and a superheater- like
feedwater preheater called an economizer. In this pass all steam loop heat transfer occurs.
After leaving the furnace these exhaust gasses flow into a ljungstrom air heat exchanger
where heat is traded to the incoming combustion air. Under the pull of the induced draft
fans this now 350 degree gas passes through the units robust electrostatic precipitator
array, through the induced draft fans themselves and then up the stack.
The makeup of the fluegas at the point it leaves the furnace represents the overall qual¬
ity of combustion. Key parameters include how much oxygen has been left by the com¬
bustion process, and how much CO has been created. The richness or leanness of
combustion is directly evident.

60
2.5.1.2.22.1 Flue gas
2.5.1.2.22.1.1 CO
These tags represent the CO contained in exhaust gasses as measured in the side A
(U27814) and side B (U27813) furnace outlets to the exhaust ducts just after the econo¬
mizer.
These are point measures of CO in a very large duct and may not capture exact CO
content. They also display extreme side to side bias with side B showing higher CO con¬
tent. Although peculiar, this side to side bias is believed to be a real feature of the Canal
Unit 2 furnace draft. These tags are directly related to the quality of combustion and can
serve as a non delayed approximation of CO as it will be seen at the stack.
2.5.1.2.22.1.2 02
These tags represent the 02 contained in exhaust gasses as measured in the side A and
side B furnace outlets to the exhaust ducts just after the economizer. These tags are used in
modeling to represent the richness or leanness of combustion. They are impacted by and
can be used as a control reference for forced draft fan output trim on air demand. In addi¬
tion these tags are used by Canal as a part of the CEM NOx calculation.
2.5.1.2.22.1.3 Temps
The temperature of the air being forced through the boiler at Canal Unit 2 impacts and
represents many process parameters, from combustion quality, to heat transfer distribu¬
tion, to induced draft fan output. It also is control reference for the temperature and veloc¬
ity of exhaust leaving the stack.

61
2.5.1.2.22.2 Stack
The CEM (Continuous Emissions Monitoring Unit) consists of an array of extraction
gas analyzers in a computer room at the base of Canal’s 500 ft. stack. The pitots of these
analyzers sniff mixed exhaust from the top of the 18 foot wide Unit 2 flue. The specific
amounts of certain compounds measured in this gas are entered into a database. This data¬
base serves as a binding legal history of Canals environmental compliance. Each violation
of emissions limits placed on certain compounds like NOx and CO is recorded. If the unit
is in danger of breaking its allowed daily average output of these regulated pollutants,
measured from midnight to midnight, all steps must be taken to regain compliance, includ¬
ing dropping load. The cost of such a sacrifice is immense and in effect these hourly and
daily emissions limits have become control variables of primary importance.
2.5.1.2.22.2.1 CO
This tag represents the CO content of stack gas in parts per million.
It is worth noting that CO and NOx represent conflicting states of combustion as they
are currently understood and managed. To reduce NOx production, combustion is kept
cool and rich. NOx formation has been shown to positively relate to increased exposure to
combustion and increased temperature. Over fire air is used to complete this off - stoichi¬
ometric combustion. Unfortunately such rich and cool (incomplete) combustion inherently
produces increased CO.
2.5.1.2.22.2.2 NOx
This tag is calculated using a regulatory approved method and is used to represents the
pounds of NOx produced by Canal Unit 2 per million BTUs.

62
2.5.1.2.22.2.3 Temp
Stack temp is important to Canal for several reasons. Keeping stack temperature at a
certain point guarantees that no condensation of sulphur products can occur in the exhaust
ducts, precipitators, or in the stack itself. The products of sulphur condensation are acidic
and over extended periods of time can be damaging to expensive capital equipment. As
long as sulphur emissions are within limits, and they are not a problem at Canal, since
Canal uses low sulphur fuel oil, it is beneficial to push them all the way out of the stack
before they can condense. This requires sufficient stack gas temperatures and velocities.
Stack Temperature is controlled primarily by the amount of preheating that is done to
the air before it even enters the windbox. Because of the relationship of final stack gas
temperature to combustion air temperature, and the relationship of combustion air temper¬
ature to other properties of combustion, stack temperature can be an important and tricky
control point. Since the heat losses to the exhaust through the exhaust ducts, precipitators
and induced draft fans are fixed, fluegas temperature also represents stack temperature but
without gas path travel delay.

CHAPTER 3
BOILER OPTIMIZATION
The most efficient method for reducing NOx emissions is clearly during the combus¬
tion process [39]. As presented in Section 2.4 "NOx," simply changing the combustion
temperature and fuel/air distribution can dramatically affect NOx emissions. The combus¬
tion of fossil fuels inside a large-scale boiler, however, is a highly complex process; this
complexity is a direct function of the boiler size. A typical electric power boiler maintains
a "fireball" which is 3 to 5 stories tall, and there are hundreds of parameters which affect
the injection of fuel and air at different locations within the furnace.
The problem is our lack of understanding about how these combustion parameters
affect NOx formation. This multivariate optimization problem requires a technology that
can look at the process globally and determine the appropriate combination of combustion
controls.
3.1 First Principles
The concepts behind boiler optimization are relatively simple:
• If the boiler operates in an oxygen-rich environment, i.e., with unnecessary
excess air, boiler efficiency will decrease due to the loss of sensible heat up the
stack; NOx emissions will increase concurrently.
• If the boiler operates in a fuel-rich environment, i.e., with insufficient air, boiler
efficiency will decrease due to the loss of unbumed fuel. In addition, insufficient
63

64
air leads to CO formation which causes slagging and water wall corrosion, ulti¬
mately shortening boiler life.
Between these two airflow conditions there is a zone of optimum combustion. This is
shown as a dark gray band in Figure 5.
3.2 Fuel and Air Distribution
Boilers for electric power and industrial steam typically have poor distribution of fuel
and air within the furnace. This causes some regions of the firebox to be fuel-rich and
other regions to be oxygen-rich. This situation is clearly undesirable as it leads not only to
unnecessary NOx production and reduced efficiency, but reduced boiler life expectancy
due to water wall corrosion and slagging.
The variability of the fuel-air ratios at different locations throughout the boiler is rep¬
resented as a light gray band in Figure 5. This variability determines the amount of aggre¬
gate air required to ensure that all regions inside the boiler avoid fuel-rich combustion.
Figure 5: Combustion emissions characteristic versus air flow.

65
By improving the distribution of fuel and air in all parts of the firebox, it is possible to
reduce the aggregate airflow while maintaining the same safety margin. This improvement
is illustrated in Figure 6. The narrower darkly-shaded band which represents the improved
distribution of air and fuel moves closer to the zone of optimum combustion. Reducing the
aggregate airflow simultaneously increases boiler efficiency and reduces NOx emissions.
Zone of
Optimum
Fuel-rich environmen Combustion Oxygen-rich environment;
Figure 6: Effect of lower 02 on combustion emissions.
The key challenge in boiler optimization is identifying which of the many controls
affect performance and how they need to be manipulated to ensure optimal performance
as process and economic conditions change.
3.3 Boiler Tuning
Boiler manufacturers and service companies offer boiler-tuning methodologies that
use the above principles of combustion to identify a limited set of control settings which
help lower NOx and increase efficiency without the need for substantial capital expendi¬
ture. Such boiler tuning improves unit performance but does not begin to generate the sav¬
ings achievable through improved control.

66
Unfortunately, the number of control variables available to optimize performance is
too large for offline boiler tuning to predict the optimum settings. Optimum settings vary
with load, fuel quality, boiler conditions, weather, and other factors making offline tuning
difficult if not impossible.
3.4 The Role of CO
Figure 5 and 6 both show an exponential rise in CO as excess air is reduced and the
boiler approaches peak efficiency. The steepness of the CO curve depends upon the
degree of mixing of fuel and air within the furnace. Poor mixing broadens the CO curve by
creating pockets of fuel-rich and oxygen-rich combustion. Together with 02, CO levels
provide the best indication about combustion quality.
A model for CO will provide valuable information about:
• how well mixed the fuel and air are in the furnace,
• how individual setpoints can be used to improve this mixing, and
• conditions which lead to slagging and water-wall corrosion.
The CO measurement serves as a key safety constraint when optimizing the boiler. By
controlling to CO levels, the boiler can be optimized without compromising safety mar¬
gins. Improved air and fuel distribution will merely tighten the CO curve, resulting in
improved efficiency and lower NOx.

CHAPTER 4
CONTROL DESIGNS
This research investigates the applicability of neurocontrol techniques to complex pro¬
cess control problems, and develops a methodology for implementing them. Towards this
end, this work will develop several detailed neural network-based control designs and
apply them to the reduction of NOx and the maintenance of acceptable CO levels in elec¬
tric power plants. Subsequent sections implementation these control designs and use our
NOx case study to compare and contrast them. The control methodology will be presented
as follows:
1) A methodology for categorizing key process variables into groups that
are required for all control designs.
2) A methodology for formally stating the control optimization objectives
and operating constraints using the aforementioned variable definitions.
3) Performance criteria by which the various control designs will be
judged, based on these formal objectives and constraints.
4) Four formal control designs with explicitly account for state variable
dependencies.
4.1 Variable Definitions
When designing a controller for large-scale industrial processes, there are a large num¬
ber of variables to be considered. The physical processes are typically considered to have
inputs, disturbances, states and outputs. The following variable definitions are proposed
as a methodology for categorizing all process variables into subsets; these subsets will
prove useful when designing controllers in general:
67

68
1) Manipulated Variables (MVs): process inputs which have been selected
for our controller to manipulate. The MVs should be independent of one
another, i.e., manipulating one will not cause a change in any of the oth¬
ers.
2) Disturbance Variables (DVs): process inputs or disturbances that affect
the state or output of the process, but we either cannot or have chosen
not to manipulate. The DV should be independent of both each other
and the MVs.
3) Control Variables (CVs): the process state or output variables that the
controller will be designed to control. The CVs should be a function of
the MVs and DVs or there is little hope of the controller being able to
control them.
4) State Variables (SVs): process state variables, which are a function of
the MVs and/or DVs, that affect the CVs. Alternatively, the SVs may be
process output variables that have not been selected for control but need
to be considered as constraints.
Notice that the MV, DV, SV and CV definitions categorize the process logically and
not physically. These definitions divide variables based on how the controller will be con¬
figured, rather than how the physical process is configured. The MVs will always be pro¬
cess inputs, i.e., can be manipulated by operators, but the DVs can contain both process
inputs and disturbances depending on which inputs are being manipulated. Likewise, SVs
and CVs can each consist of any combination of process states and/or outputs, based on
which will ultimately be controlled.
Notation: The categorization of variables into CVs, SVs, DVs and MVs
will be used extensively throughout this work, and is conceptually consis¬
tent with the literature on optimization and control [37].
4.2 Optimization Objectives
The control objective is to lower NOx. Formally this objective needs to be stated as an
objective function for optimization. Since several of the controllers developed here are tra-

69
jectory (multi-stage) controllers, this objective function will be a function of time. Con¬
sider the single control variable NOx(t0) e 91 as the measured value of NOx at time t0.
An optimal control objective with fixed terminal time T for minimizing NOx(t) over the
interval t e (t0, t0 + T] can be given by
(54)
t = t.
In general, there will be more than one CV. If all CVs are to have equal impact on this
objective function, then two effects will have to be removed from the optimization objec¬
tive: 1) the effect of power differences between these CVs, and 2) the current value of
each CV. The following objective function extends (54) to multiple control variables
(55)
where Ncv is the number of CVs, and p( 6 91 is a priority weighting factor and
Z(x) - (x- pY)/av is the z-score statistic [66], Assuming that our controller is designed
to minimize J\ for p( > 0 the CV cvt will be minimized over the trajectory, while setting
p( < 0 will maximize the output.
Equation (55) considers the case where CVs are to be maximized or minimized. In
general, the goal is to design a controller capable of maintaining a control setpoint. A gen¬
eralized optimization objective is therefore presented as
(56)

70
where D¡ is a desirability function that can be tailored for each CV to
D"‘ax = Z(cv,(0)-Z(cv,(t 0)),
(57)
D'fn = Z(cv,.(i0))-Z(cv.( 0),or
(58)
D’p = J(Z(cVi(t)) - Z(sp(t)))2 .
(59)
4.3 Operating Constraints
Constraints will be used to ensure that the optimizer produces a feasible solution. By
feasible we mean: 1) the MV moves can be made, and 2) that when these MVs are applied
the plant will end up in a desirable state. Feasible solutions will be guaranteed by design¬
ing controllers which are able to maintain MV and SV constraints.
4.3.1 Manipulated Variable Constraints
To ensure that the MV moves can be made, the controllers will maintain simple range
constraints. A range constraint consists of the upper and lower limits that an MV will be
allowed to move. Formally the range constraint for MV mv¡ will be given by
Cmv _ r^min max,
i — 5 J 5
(60)
where C™"' is the MV’s absolute minimum and is its maximum. Controllers will be
He ‘o+T-
required to provide an optimal MV trajectory {mv (t)} t = / such that
Vi, t.
e Cfv
(61)

71
4.3.2 State Variable Constraints
Similarly, to ensure that controllers drive the plant to a desirable state, SV constraints
will also have to be addressed. Formally, controllers will be required to provide optimal
t0+ T
MV trajectories that result in SV trajectories {sv (t)}t = ,0+ i such that
SV;
^,*(0 eCf Vi,f. (62)
4.3.3 Penalty Functions
Each control design considered will employ an optimization algorithm during some
phase of its development. Some optimization algorithms are able to deal with constraints
directly, i.e., given knowledge of the constraints they can ensure a feasible solutions. Oth¬
ers, however, will have to treat constraints indirectly by addressing them with the objec¬
tive function. The most common method for addressing operating constraints in an
objective function are through the use of penalty functions [54], For example, SV con¬
straints can be stated as penalty functions of the form
Msv,, c;v) =
Í 2
, „max¿ „n
(sv¿ - C ) sv¿ > C
, „/>
(sv,- - C ) sv,- < C
V 0 else
(63)
Generalizing the penalty function to multiple SV constraints, differences in the energy
of the respective signals will once again have to be normalized out. These effects can be
compensated for using a generalized penalty function of the form
Hsvi, c;v)
\z(svi) - Z(C"‘ax))2 sv- > Cmi
(Z(5v,.)-Z(Cmi"))2 sv, < C""
(64)
V
0
else

72
Given a set of N6' SV constraints, an optimizer may satisfy these constraints by
appending their respective penalty functions to its criterion
/=/+^p’vMivi,C'v), (65)
(= 1
where p*' allows constraints to be individually prioritized. Similarly, both MV con¬
straints can be appended to the optimizer’s criterion by defining the penalty functions
{X(mv¡,C¡ )}i=|.
Note that implementing constraints with penalty functions will not guarantee that the
constraints are met precisely. If the constraints are properly prioritized relative to the opti¬
mization objectives, however, these constraints are easily maintained within the desired
level of accuracy.
4.4 Performance Criteria
For the case study, controllers will be judged based on their ability to lower NOx while
maintaining desired CO emissions. To this end, subsequent sections will measure the per¬
formance of controllers as a plant operator moves MVs according to their control laws.
Comparing controller performance, however, will prove a difficult task, since the operator
can only take the advice from one controller at a time and the plant is constantly changing
state. Although the controllers may be able to deal with non-steady-state conditions, it will
be nearly impossible to separate the process responses to the state changes versus the con¬
trol action.

73
Further complicating matters, while one-time tests will provide useful results with
which to judge the controllers, they are not the only criteria. The controllers studied will
be judged by the following criteria:
1) Ability to control NOx and CO.
2) Ability of the operators to perform the recommended MV moves
3) Flexibility with respect to changing performance objectives and operat¬
ing constraints
4) Ability to deal with changing operating states, e.g. load changes
4.5 Controller Designs
Four controller designs will be developed. The controller designs considered, fall into
the broad categories of:
1) Model-Predictive Control
2) Model-Inverse Control
3) Model-Based Direct Control
There are, however, no standard recipes for building these controllers. The field is still
immature, and neurocontrol designs presented in the literature tend to be ad hoc. This
work seeks to not only develop and test four neurocontrol designs, but also to develop a
generalized methodology for implementing control designs belonging to the above
abstract categories. Each controller must be able to deal with the MV and SV constraints,
and will be judged by the performance criteria described above.
For each of the control designs considered, there are two distinct phases in the imple¬
mentation:
1) Offline training.
2) Online control.

74
4.5.1 Steady-State Optimizer
The simplest, and most prevalent, neurocontroller in the literature is the steady-state
optimizer [43][40][31 ]. This controller belongs to the model-predictive control family.
Model-predictive control (MPC) is not new to commercial applications in the process con¬
trol industry. The advance proposed here is the application of neural network reference
models within this controls methodology.
The concept of MPC is straight forward: combine a model for the process with an opti¬
mizer to obtain real-time optimal setpoints. Model predictive controllers can be steady-
state or dynamic, depending on characteristics of their underlying process models. This
section details the design of a neural network-based steady-state MPC controller to meet
the problem specifications presented in Sections 4.2, 4.3 and 4.4.
4.5.1.1 Offline training
Training a MPC controller follows the schematic outlined in Figure 7. Notice that
there are actually two reference models being trained: one SV model and one CV model.
The details for how to train these models will be covered in Chapter 6. Notice, however,
that the criterion j" is a model training criterion to be presented in Chapter 6, and not the
control performance objective f presented in Section 4.4 "Performance Criteria."
The model definitions required by the steady-state optimizer are:
1) Steady-State SV Model: sv = ssSVModel(mv, dv)
2) Steady-State CV Model: cv = ssCVModel(mv, dv, sv)

75
Figure 7: Offline training and retuning configuration for steady-state optimizer.
The reason to have a CV model is obvious, it will provide the reference model that the
optimizer uses to figure out its optimal MV setpoints. The motivation for having a SV
model, however, is somewhat less apparent. The problem is that changes made to the MVs
by the optimizer will not only change the CVs, but also the SVs. The optimizer will have
to consider the effect that MVs will have on the SVs, if it is to accurately predict their
effect on CVs. Note that the CV model has an input space that consist of MVs, DVs and
SVs.
4.5.1.2 Online control
The online control configuration is illustrated in Figure 8. Here an optimizer calculates
ArgMin_±*{f} using the SV and CV reference models developed during model train-
mv
ing. The optimizer starts with the current value of the MVs mv* = mv, uses the SV
. \ j|{ <
model to estimate the current SVs sv , which are then used, along with the current value

76
of the DVs dv , to estimate the current CVs cv* . The optimizer then iteratively updates
• • • t t
its estimate for the optimal MVs mv to minimize its objective function j .
Figure 8: Online control configuration for steady-state optimizer.
Both direct and descent-based optimization can be used for MPC. If the number of
MVs is small, then direct optimizer provides an efficient alternative. As the number of
MVs grows, however, direct optimization quickly becomes impractical. Descent-based
optimization is possible because the SV and CV models are capable, via backpropagation,
of calculating the gradient of f with respect to their inputs, i.e., their inputs sensitivities
given the sensitivities at their outputs. In this manner, the optimizer calculates the CV sen¬
sitivities df /dev* , from which the CV Model is able to calculate SV sensitivities
df /dsv* and partial MV sensitivities df'/dmv* , from which the SV model calculates
the remaining partial MV sensitivities df' /dmv* , and finally the optimizer is able to
update its optimal MV estimate using the MV gradient of
df _ dfv +dfv
dmv* dmv* dmv*
(66)

77
This is really just the backpropagation of backpropagations, a.k.a. more fun with the
chain rule.
Several optimization methods were tested for the optimizer, along with various tech¬
niques for dealing with the constraints. The most effective combination identified was to
use the unconstrained conjugate gradients method in combination with an objective func¬
tion which included the SV and MV constraint penalty functions
Vv
V"v
f - ± Z pí’Z(«v,*> + ± E pTHsv,*, CD + -L, E PrM-v,*, Cf), (67)
aT;
i= 1
NSV ; =
N"‘v,
i = 1
where the MV, SV, DV and CV variable sets, along with their corresponding constraints,
are defined in Section 6.6.3 "Final Variable Sets," and all priorities have been set to 1. The
details of the conjugate gradients method will be presented in Section 6.5 "Learning Algo¬
rithm."
Penalty function can negatively impact the performance of a descent-based optimizer
by adding complexity to the performance surface having little to do with the underlying
problem. This is particularly true when the constrained variables lie outside of their con¬
strained values. For the SV constraints, there is no choice but to use penalty functions for
constraints. For MVs, however, there are alternatives, because the MVs always start at
their current values which are always within the constraints. Hence, there is little to no
overhead to using MV constraints for our online optimizer. It terms of the performance
surface, the constraints can be thought of as placing a guardrail on both sides of our cur¬
rent position in weight-space along our path, while having little impact on the local topog¬
raphy of the road.

78
4.5.2 Steady-State Model-Inverse Controller
The next controller design belongs to the model-inverse control (MIC) family. Con¬
ceptually, model-inverse control is straightforward: train a model to predict the MVs from
the current and known DVs, SVs and CVs, then, given a desired CV setpoint, this model
can be used directly to obtain the required MVs. Implementing a MIC controller is also
straightforward and can work reasonably well, given that the relationship between MVs
and CVs is in fact invertible. This sections details the design of a neural network-based
MIC controller, designed to meet the problem specifications presented in Sections 4.2, 4.3
and 4.4.
4.5.2.1 Offline training
Training a MIC controller follows the schematic outlined in Figure 9. Once again,
notice that the MV model is being implemented by separate inverse-SV (ISV) and
inverse-MV (IMV) models. Once again, the details for how to train these models will be
covered in Chapter 6. The model definitions required by the steady-state model-inverse
controller are:
1) Steady-State ISV Model: sv = ssISVModel(cv, dv)
2) Steady-State IMV Model: mv = ssIMVModel(cv, sv, dv)

79
Figure 9: Offline training and retuning configuration for model-inverse controller.
Analogous to our MPC controller, two models have been developed which when com¬
bined can invert the process. The reason to have a IMV model is obvious, it provides the
inverse-model that the controller uses to figure out optimal MV setpoints. The problem is
that not all CV-SV combinations are feasible. Given a specified CV target, the ISV model
estimates the corresponding SVs which are presented to the IMV model.
4.5.2.2 Online control
The online control configuration is illustrated in Figure 10. If a known target existed
for the CVs, the online control implementation would actually be quite trivial. One com¬
plexity is that the exact value for the lowest achievable NOx from the controller for a
given set of conditions is not known. Another complication with MIC is how to deal with
constraints. If one applies a target CV to the input of the inverse-model, it will predict a set
of inputs which it believes would have achieved this target. The problem is that the model
does not understand the MV or SV constraints, and if one of the inputs it predicts falls out-

80
side these constraints the controller can not provide the required setpoints. This is analo¬
gous to the problem faced with SV or CV constraints for MPC. The implementation
outlined in Figure 10, uses an optimizer in order to overcome both of these issues. Clearly,
CV constraints are straightforward.
Figure 10: Online control configuration for model-inverse controller.
The MIC controller uses an optimizer to calculate ArgMin^{f) using the ISV and
CV
IMV reference models developed during model training. The optimizer starts with the cur¬
rent value of the CVs cv* = cv, uses the ISV model to estimate the current SVs sv* ,
which are then used, along with the current value of the DVs dv , by the IMV model to
—^
estimate the current MVs mv . The optimizer then iteratively updates its estimate for the
optimal MVs mv* , to minimize its objective function f.
Once again both direct and descent-based optimization can be used for MIC, and once
again a conjugate gradients-based optimizer was selected. Descent-based optimization is
possible because the ISV and IMV models are capable, via backpropagation, of calculat¬
ing the gradient of f with respect to their inputs, i.e., their inputs sensitivities given the

81
sensitivities at their outputs. In this manner, the optimizer calculates the MV sensitivities
df Idmv* , from which the IMV Model is able to calculate SV sensitivities dfIdsv* and
partial CV sensitivities df‘v/dcv* , from which the ISV model calculates the remaining
partial CV sensitivities df*' /dev* , and finally the optimizer is able to update it optimal
CV estimate using the CV gradient of
df = 8fsv ar
3^* a~^*
ocv ocv ocv
(68)
The optimizer’s objective function, which includes the SV and MV constraint penalty
functions, is the same objective function used by our steady-state optimizer. The only dif¬
ference is how the sensitivities flow through the system, as outlined above.
4.5.3 Dynamic Model-Predictive Controller
The steady-state optimizer considered above is a member of the MPC family. The vast
majority of MPC applications use models which are first-principles based [37], Since it is
not possible to build an accurate first-principles model of NOx, a new steady-state opti¬
mizer for MPC using neural network models was developed. The vast majority of MPC
applications are dynamic, however. The steady-state optimizer only considers the effect
that MV changes will have on the unit in steady-state conditions.
This section develops a dynamic neural network-based MPC controller. The main dif¬
ferences between this controller and our steady-state optimizer is that it:
1) Understands the dynamics of the process.
2) Provides a trajectory of MV setpoints designed to optimize the path of
the unit into the future, rather than a optimal steady-state position. In
other words, the controller not only considers where your going but how
you’ll get there.

82
The concept behind this controller’s operation is identical to that of the steady-state
optimizer: combine a model for the process with an optimizer to obtain real-time optimal
setpoints. The only difference is that the models are now dynamic, and the optimal set-
points become optimal setpoint trajectories.
This sections details the design of a neural network-based dynamic MPC controller to
meet the problem specifications presented in Sections 4.2, 4.3 and 4.4.
4.5.3.1 Offline training
Training a dynamic MPC controller follows a similar schematic as outlined in Figure
7, with the inclusion of each variables explicit dependence on time t, as illustrated in Fig¬
ure 11. Here the SV and CV models are performing single-stage prediction with respect to
the MVs and DVs; notice that the CV model uses the current value of the SVs sv(t + 1).
The reasons for this configuration will become clear when we consider the online control
implementation in the next section.
Figure 11: Offline training and retuning configuration for steady-state optimizer.

83
Refer to Chapter 6 for details on training the dynamic SV and CV reference models
used by the dynamic MPC controller. For now we simply state the model definitions
required by the dynamic MPC controller:
1) Dynamic SV Model: sv(t+ 1) = dSVModel(mv(t), dv(t))
2) Dynamic CV Model:
cv(t+l) = dCVModel(mv(t), dv(t), sv(t + 1))
4.5.3.2 Online control
The online control configuration follows a similar configuration to the steady-state
optimizer presented in Figure 8. Here a dynamic optimizer is required, however. The opti¬
mizer calculates ArgMin > {./(t)} using the dynamic SV and CV models developed
mv*(t)
during model training. The steady-state optimizer used an application of the chain rule for
ordered partial derivatives, which has been coined “backpropagation” [40], From the per¬
spective of the chain rule, our new optimizer is identical and only the criterion changes.
From the perspective of the literature, this algorithm has been coined “backpropagation
through time” [40][70].
The optimizer starts with the current value of the MVs mv*(t) = mv(t); uses the SV
^ )|(
model to estimate the resulting SVs sv (t + 1); which are then used, along with the cur¬
rent value of the DVs dv(t), to estimate the resulting CVs cv*(t + 1) . Notice that each
estimate can rely on both present and past values of the inputs. The optimizer will then
repeat this process over the time interval t e (tQ, t0 + T] to produce the MV, SV and CV
^ t0+T— 1 - _! t0+T x* t0+T
trajectories {mv (0}/ = /„ > {5V (0}/ = f0+i and icv (0}/ = /„+ l > respectively.

84
The objective function which included the SV and MV constraint penalty functions
can now be calculated
At„) =
in + Tr Ncv
1
N!v
N"
-1 p>i*w)+-Tv I p?v ww. o+x Prvx(«v,-( o.
yv/f, ^ , ,
i = 1
(69)
where the MV, SV, DV and CV variable sets, along with their corresponding constraints,
are defined in Section 6.6.3 "Final Variable Sets," and all priorities have been set to 1.
The optimizer then iteratively updates its estimate for the optimal MV trajectories,
>.* t0+T-\
{mv = tg , to minimize its objective function J (t) . Each step in the iteration per¬
forms the following, starting with t = tQ + T- 1 and iterating down to t = t0 : first, the
optimizer calculates the CV sensitivities dJ°(t)/dev* (t), from which the CV Model is
able to calculate SV sensitivities df \t)!dsv* (t) and partial MV sensitivities
dfv(t)/dmv*(t - 1), from which the SV model calculates the remaining partial MV sen¬
sitivities dfv(t)ldmv (t— 1), and finally the MV sensitivity at time t— 1 can be calcu¬
lated as
df(t) = df\t) df\t)
dmv*(t- 1) dmv*(J— 1) dmv*(t-\)
(70)
Once the backward pass is complete, the optimizer is now able to update it’s optimal
ta+T~.
MV trajectory estimate {mv (t)}t = to using the MV gradient trajectory
sao rr-
.dmv (0-
t = ta
(71)

85
Notice that the sensitivities at time t depend on the sensitivities in the future. This is
because the models variables at time t depend on the variables in the past, i.e., the models
are dynamic. Hence the term “backpropagation through time.”
This entire optimization cycle is run at each time step tQ. The optimizer derives the
next T- 1 MV moves, and the first MV setpoint is applied to the unit mv (/0). At this
point the entire process is repeated.
4.5.4 Model-Reference Adaptive Controller
The final controller design considered belongs the model-reference adaptive control
family (MRAC). Like the dynamic MPC controller, the MRAC controller understands
process dynamics and provides a trajectory of MV setpoints which optimize both where
you are going and how you get there. The fundamental difference between these two con¬
trollers is how this optimal trajectory is derived. The MPC design utilized an online opti¬
mizer to calculate this trajectory, while the MRAC design develops a neural-network
based controller which is able to calculate the optimal trajectory directly. Hence this is our
first direct controller, i.e., calculates MV setpoints directly.
Notice that the MIC design would have provided a direct online controller, if it wasn’t
for:
1) the lack of a known target NOx level, and
2) the requirements for MV and SV constraints.
The MRAC design is able to overcome both of these hurdles by building knowledge of
the best achievable NOx level and by building all of the constraints directly into the con¬
troller. The main advantage to the MRAC design is online response time. There is no opti-

86
mization to run, one simply presents the controller with the current, and past, state of the
process, and it generates a MV setpoint as quickly as a neural network can think. These
benefits do not come for free, however. The main drawbacks to the MRAC design are:
1) Extensive offline training and retuning requirements.
2) Inflexible online configuration, with respect to changing optimization
objectives and operating constraints.
This sections details the design of a neural network-based dynamic MRAC controller
to meet the problem specifications presented in Sections 4.2, 4.3 and 4.4.
4.5.4.1 Offline training
Training a MRAC controller requires two stages. The first stage is identical to training
and returning the dynamic MPC controller. Here, dynamic SV and CV models are devel¬
oped using the same steps outlined in Figure 11. The second stage uses these models to
train the controller with offline data, as illustrated in Figure 12. The offline training is sim¬
ilar to the online optimization which is performed for the MPC design, except this optimi¬
zation is performed across the training dataset rather than online.
Figure 12: Offline training and retuning configuration for model reference controller.

87
Once again a dynamic optimizer is required, and the objective function is given by
CL
(69). To train the controller, the optimizer calculates ArgMin^CL{f (t)}, where Jp are
the weights of the control law neural network.
Training the control law (CL) model starts with the actual values for the DVs, SVs and
CVs, and random initial weights for its CL model Jp . Starting at time t0 = T¡+ 1,
where T¡ is the first sample in the training dataset; training uses the CL model to estimate
^ j|( f \ jj(
the resulting MVs mv (t); which are then used to estimate the resulting SVs sv (t + 1)
—^
and CVs cv (t + 1) . This process is repeated over the time interval t e (t0, tQ + T] to
t0+T-\ v :|. t0+T
produce the MV, SV and CV trajectories {mv (t)}i = io , {sv (0}i = i0+i an(l
‘o+T
{cv (t)}t = to+\, respectively.
The training algorithm then iteratively updates its estimate for the optimal CL model
weights, JPCL, to minimize its objective function f \t). Each step in the iteration per¬
forms the following, starting with t = tQ + T and iterating down to t - t0 + 1: First, the
training algorithm calculates the CV sensitivities df (t)/dcv* (t), from which the CV
Model is able to calculate SV sensitivities df (t)/dsv*(t) and partial MV sensitivities
dfv(t)/dmv*(t - 1), from which the SV model calculates the remaining partial MV sen-

88
sitivities dfv(t)/dmv*(t - 1), and finally the MV sensitivity at time t - 1 can be calcu¬
lated as
df(t) = df\t) + df\t) (7T
dmv*(t- 1) dmv*(t- 1) dmv*(t—\)
The MV sensitivities are finally passed to the CL model which backpropagates them
to derive its weight gradients df /dW , which the training algorithm is able to use to
update its control law’s weight estimate.
The training algorithm then increments tQ and repeats the entire process, until the
training algorithm has converged. When t0 = Tj— T, tQ is reset to tQ = 7j. + 1. The rea¬
son for training the CL model in increments of T is because the SV and CV have a limited
prediction horizon, the time before their estimates are no longer valid. By resetting the
state of these models to the actual state of the unit after T samples, we are able train the
CL model within the prediction horizon of the SV and CV models.
4.5.4.2 Online control
The online control configuration for the MRAC design is straight forward, as illus¬
trated in Figure 13. Simply supply the controller with the current SVs, DVs and CVs, and
it outputs the next MV setpoint. This setpoints contains knowledge about the optimal
achievable NOx, SV constraints, MV constraints and the trajectory through which it will
drive the process into the future.

89
{sv(t), dv(t)}
Control
mv*(t)
Plant
cv(t)
f-^
Law
Figure 13: Online control configuration for model reference controller.
Clearly, the controller is only as good as its underlying reference models. In addition,
considerable care must be taken to ensure that the training data contains regions of the
input space where the SV and MV constraints have been exercised. The design is easily
augmented with limiters to guarantee that MV constraints are maintained. However, there
is little that can be done to guarantee that the SV constraints are maintained.

CHAPTER 5
DATA PREPARATION
Given the detailed control designs just presented, the next step is to implement the
actual controllers by developing the required reference models. Both reference model and
controller implementations require a significant amount of process data. Data collection is
the most important aspect of any modeling or optimization project. There is a common
saying “junk in, junk out,” this study was relentless in reenforcing this lesson. Applying
the most sophisticated modeling and/or optimal control algorithms in the world will not
make up for problems with data preparation.
With the advanced distributed control systems (DDS) and supervisory control and data
acquisition (SC AD A) systems readily available in today’s process plants, the relative
quantity and quality of available data is overwhelming. Much of the statistics and model¬
ing literature has been dedicated to the problems faced when drawing inferences from
small sample spaces. Modem processing plants are anything but data limited. The relevant
problems are just the opposite, how to draw meaningful inferences from a massive sample
space.
The following section presents solutions for the most significant challenges faced in
preparing data for modeling and optimization. Much of what is presented in this section
was learned the hard way, during modeling and optimization.
90

91
5.1 Data Management
The power plant treated in this work collects and stores tens of thousands of variables
from various sensors and actuators throughout the plant. These variables were collected by
the DCS and forwarded to a data historian called PI, marketed by OSI, Incorporated.
Notation: The author will reserve the term “variable” to represent process
states with a specific physical interpretation, which are considered relevant
for modeling. The DCS will collect information from many sensors or
actuators each of which represent a single process variable. The specific
points collected by the DCS will be referred to as “tags.”
The DCS works with exception-based sampling, a non-uniform sampling scheme.
Each tag is given an absolute deviation above which an exception is raised. Raised excep¬
tions are forwarded to the process control algorithms within the DCS, as well as to PI. The
PI data historian then applies a time stamp and stores the exception. The time resolution of
a sample can be trusted within about ±2 seconds, and the quantization error is approxi¬
mately equal to the deviation settings for each tag in the DCS.
This exception-based data acquisition scheme allows the data historian to store an
impressive amount of data. Uniformly sampled trends are provided through simple func¬
tion calls to Pi’s API. Any subset of the 10,000+ tags can be easily recalled for arbitrary
time ranges over the last year of continuous plant operation. Initial concerns over the data
quality given this non-uniform sampling scheme were quickly dismissed. However, sig¬
nificant tuning of the tag deviations was required however.

92
5.2 Variable Selection
For the case study of boiler optimization, it is clear that we are looking for process
variables that are related to the measurement of NOx or CO, and impact the combustion
process with respect to these measurements. With the massive number of variables to
choose from in the database, the first line of defense is to use our first-principles knowl¬
edge of the process.
Interviews were conducted with engineering and operations personnel from the plant.
They were asked to:
1) Identify all tags that represent the variables NOx and CO.
2) Identify all variables that have any affect on combustion parameters like
emissions, fuel and air flows, temperatures or pressures inside the
boiler.
3) Rank these variables with respect to their effect on NOx and CO as
essential, secondary or minimal.
4) Identify all of the tags that represent or directly impact the listed vari¬
ables.
5) Classify each of these tags as either:
•Setpoint: can be manipulated by operators via the DCS.
•Tunable Parameter (TP): can be manually manipulated by “engi¬
neers with wrenches”.
•Disturbance: can not be manipulated but has an affect on com¬
bustion.
•State: represents a particular state of combustion which will have
an affect on emissions, can be a function of setpoints and/or dis-

93
turbances but cannot be manipulated directly.
6) Identify operating constraints and concerns that require monitoring
when manipulating any of the setpoints.
This process, and subsequent iterations, produced 76 essential, 100 secondary and 300
minimal tags for Canal Electric Generating Station. The essential tags are listed in Appen¬
dix. The operators and engineers identified CO as their overwhelming operating con¬
straint. The process of variable selection shall be continued in Section 6.2 "Model
Definitions."
5.3 Validation
In order to get a handle on the quality of data being collected by the data historian, for
each of the essential tags above, one month of data were extracted between 1/1/1998 and
2/1/1998. The sampling rate was set to Ts = 5 seconds. One of the most useful data vali¬
dation tools was to simply trend each tag over various intervals. The eye is able to spot
most data integrity issues that are commonly missed by statistical indicators.
The standard descriptive statistics of mean, variance, max and standard error were cal¬
culated for each tag over the datasets. These statistics were compared against first-princi¬
ples knowledge to look for the data integrity issues considered below.
5.3.1 Quantization and Clipping
The most common data integrity problem encountered with industrial data historians
is having invalid range settings. Range settings for event based data acquisition systems
are analogous to sampling rate for uniform sampling. There are three important settings,
the data minimum, data maximum and absolute deviation. When either the data minimum

94
settings are too high or the data maximum settings are too low, the archived data is
clipped. In addition, if the deviation setting is too large, quantization errors can signifi¬
cantly corrupt the archived data. The best way to detect these errors is simple visual
inspection, and the only remedy is to correct the settings.
5.3.2 Missing Data
Along with each sample, the historian provided status information indicating any
errors that were detected by the DCS when collecting the sample. Considering any error as
missing, the dataset contained 10.46% missing data. Since one of the goals of this project
is to evaluate the performance of both static and dynamic models, all missing values will
need to be accounted for despite the fact that there numbers are quite low.
60.00%
-
-
-
-
-
-
-
-
-
-
-
-
â–  Good
â–¡ Missing
40.00%
30.00%
20.00%
10.00%
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
K4 vS4 -fC4 nS4 .K4 K4
-A ^ ^ ^ ^ /
Figure 14: Daily % missing across February dataset.
Figures 14 illustrates the daily sum of missing values across the dataset. One can eas¬
ily identify a region of four consecutive days where the plant was missing data. This
region corresponds to a unit shutdown. Note that removing large blocks of time from the

95
data will not be an issue for dynamic modeling, as long as the valid data falls into a rela¬
tively small number of blocks with a large number of samples that are free of missing val¬
ues. Removing this region from our analysis, the missing percentages become less than
1%.
Since the remaining errors are highly sporadic and down sampling of the data is
required, there exist a solution for dealing with the remaining errors. The collected data
was sampled at Ts = 5 seconds. Down-sampling to Ts = 60 seconds will be applied by
averaging 12 consecutive samples and decimating. If 6 or fewer of the samples are in
error, then the remaining will be averages and the status of the decimated sample’s status
will be set to “good”. If more than 6 samples are in error, then the decimated sample’s sta¬
tus will be set to “missing”. This procedure removed all remaining errors from this
dataset.
Although the above procedure was able to clean all of the data for this dataset, it is still
possible that future datasets may still contain missing values. For all other cases, missing
data will be replaced with interpolated data between the nearest surrounding valid sam¬
ples.
5.3.3 Outliers
Treating each tag as a random variable, the tags were standardized to a Z-Score by
subtracting the mean and dividing by their standard deviation
(73)

96
The Z-score removes all effects of offset and measurement scale. They can approach
positive and negative infinity. Table 1 shows the probability of the absolute value of a Z-
score exceeding some limit for normally distributed variables.
Table 1: Probability of Z-Score exceeding value.
|Z-Score|
P(Exceeding)
1.28
0.2
1.64
0.1
1.96
0.05
2.58
0.01
Calculating the probabilities of a Z-score value exceeding a for our datasets as
N
P(Z, > a) = - y \
tffi lo
\Z¡(Xj)\ > a
else
(74)
where N is the total number of samples in the dataset. Tags with P(Z¿ > 1.64) > 0.20 we
considered candidates for filtering or smoothing operations, and tags with
P(Z¡ > 2.58) > 0.02 were considered candidates for outlier removal.
Filtering operations were most commonly applied to tags with sensor cleaning and cal¬
ibration spikes. Sensor calibration can occur as often as hourly for sensors exposed to the
fluegas. Cleaning and calibration will corrupt the data with short-duration spikes within
the range of normal data for the tag. Lowpass filtering was used to remove these effects
when encountered.
Tags with outliers had large spikes, well outside the range of normal data. Outliers
were most often caused by glitches during data acquisition. Errors in the data acquisition
system characteristically caused large instantaneous changes in the tag’s value for a short

97
period; after this period the tag’s value would instantaneously return to its true value. Out¬
liers were removed by replacing the missing data with linearly interpolated data.
5.4 Time Constants
There are many delays or dead-times inherent to the process. The natural response
time of most MV actuators, however, is less than 1 second. Their effect on combustion is
felt within seconds. Much of this will take place faster than it can resolved. If its effect can
be measured within our sampling resolution, then the dynamic models should have no
trouble extracting the temporal relationships.
There are two temporal lags that will give the modeling effort trouble. First, some set-
point actuators are driven with PID loops which have been dampened to prevent operators
from over reacting. The steady-state settling time of such loops can be as long as 90 min¬
utes. This situation has been compensated for by using a simple first-order low pass filter
to dampen the actuator setpoint signal to match the actuators response characteristics.
Second, some sensors, particularly the continuous emissions monitors (CEMs), can
have significant extraction times. The Canal Electric Generating Station CEMs measuring
NOx and CO add a 8 to 10 minute dead-time between the gases formation in the boiler
until they are recorded. This situation has been taken care of by shifting these tags within
the dataset, such that setpoints and process outputs are aligned.
5.5 Normalization
In support of modeling, all tags were normalized. Variables applied to a neural net¬
work should fall within the neuron’s activation limits. All of the neural networks consid¬
ered in this study utilize a tank activation function, therefore all variables were normalized

to fall within the range [-0.9,0.9]. Assuming normal distributions, the tag Z-score’s were
used for normalization such that only 1% of the data would fall outside the neuron’s acti¬
vation limits, according to
x
-2.58,
(75)
where r =
2.58-(-2.58)
0.9-(-0.9)
2.867 is the ratio between acceptable Z-score range and the
neuron’s activation range. Denormalization was then preformed according to

CHAPTER 6
MODELING
The modeling objectives for this work are inherently tied to the control designs and the
case study presented in the proceeding sections. To this end, the objective of this section is
to determine the best model architecture for each of the model definitions required by the
various control designs considered. Candidate model architectures will be judged based on
their ability to predict process dynamics. The following architectures will be considered:
1) Auto-Regressive Moving Average Model (ARMA)
2) Multi-layer Perceptron (MLP)
3) Time-Delay Neural Network (TDNN)
4) Gamma Neural Network (GNN)
5) Nonlinear State-Space Model (NLSS)
Notice that the ARMA model has been included to provide a benchmark and to vali¬
date the application of nonlinear control strategies for the case study application.
6.1 Methodology
The methodology for developing the “best” models will be as follows:
1) Model Definitions: The ultimate goal is to find the best possible models
for NOx and CO. The control designs presented require that these pro¬
cesses are represented using specific model definitions. Detailed speci¬
fications for the models are developed, i.e., identifying the specific
process inputs and outputs to be used by the models.
2) Datasets: A 3 month dataset will be generated using the methodology
presented in Chapter 5, and divided into disjoint training, cross-valida¬
tion and testing regions.
99

100
3) Learning Algorithm: The learning algorithm used for model training
combines the Polak-Ribiere algorithm [49] with a line search as pre¬
sented by Brent [14]. This algorithm is presented along with the details
describing its application.
4) Performance Criteria: The criteria for selecting the “best” model are
presented.
5) Variable Pruning: A MLP is constructed, and starting from the input
sets determined from first-principles knowledge in Section 5.2 "Vari¬
able Selection" the input sets are pruned to the smallest possible set of
relevant variables.
6) Architecture Selection: Optimal parameters (e.g. number of hidden lay¬
ers, processing elements and memory taps) will be individually deter¬
mined for each combination of architecture and process output using a
direct search methodology.
7) Analysis: The results will be analyzed to find the best steady-state and
dynamic models for each model definition.
6.2 Model Definitions
In support of the control designs presented in Chapter 4, models will be developed
according to the following model definitions:
1) Steady-State SV Model: sv = ssSVModel(mv, dv)
2) Steady-State CV Model: cv = ssCVModel(mv, dv, sv)
3) Steady-State ISV Model: sv = ssISVModel(cv, dv)
4) Steady-State IMV Model: mv = ssIMVModel(cv, sv, dv)
5) Dynamic SV Model: sv(í+ 1) = dSVModel(mv(t), dv(t))
6) Dynamic CV Model:
cv(t + 1) = dCVModel(mv(t), dv(t), sv(t + 1))
where the vectors mv, dv, sv and cv represent the steady state values of the MV,
DV, SV and CV variable sets, respectively; mv(t), dv(t) , sv(t) and cv(t) represent

101
their respective values at time t; and by cv = ssCVModel{mv, dv, sv) we mean that
ssCVModel is a model which takes on vectors mv, dv and sv as inputs, and produces
the vector cv as an output.
Notation: The term model definition will be used to describe the input/out¬
put space of a model along with whether it is steady-state or dynamic,
where model refers to a particular realization of a predictor which imple¬
ments the model definition. There will be many models developed which
implement each of the above model definitions.
6.2.1 Variable Definitions
The model definitions were based on definitions for the MV, DV, SV and CV variable
sets. The tag list in Appendix is categorized according to the variable definitions, which
shall then be used to implement the model definitions.
6.2.1.1 Control variables
The case study considers a single control variable, NOx. While CO is a process output
and maintaining appropriate levels of CO is also an objective, CO shall be considered a
constrained SV. Notice that NOx has been marked as a CV in the Essential Tag List under
the field labeled “Type.”
6.2.1.2 Manipulated variables
The MVs are defined as the variables that we want the controller to manipulate. There
are many inputs that could be manipulated, but there is a cost associated with the number
of manipulated inputs. These costs include:
1) generalization costs associated with the “curse of dimensionality,”

102
2) computational costs associated with the optimization during both mod¬
eling and control, and
3) operational costs associated with getting the operators to implement the
MV setpoints.
The goal is for the controller to manipulate those variables which have the greatest
impact on combustion, but the number of inputs should be restricted. Beginning with a
wide list of all potential MV candidates developed from first-principles knowledge of the
process, variables with the least impact on the SVs and CVs are pruned. The pruning
methodology will be presented in Section 6.6 "Variable Pruning." The complete list of
potential MVs is presented in the Essential Tag List.
6.2.1.3 Disturbance variables
The disturbance variables are defined as combustion variables that have an affect on
combustion and are independent of all other MVs and DVs. Variables which are functions
of MVs or other DVs will be considered as SVs. The associated costs with MVs all apply
to DV except for the operator manipulation costs. In addition to true process disturbances,
the DVs will contain MVs that we have chosen not to manipulate. The initial DVs are
listed in the Essential Tag List.
6.2.1.4 State variables
As mentioned above, CO will be considered a state variable. This will allow the con¬
troller to constrain its allowable levels. This will not, however, be our only SV. The initial
set of SVs considered for modeling is presented in the Essential Tag List.
6.2.1.5 Variable representation
You probable noticed that many of the variables chosen to represent initial MVs, DVs,
SVs or CVs represent the same underlying process variable. The process in Section 5.2

103
"Variable Selection" resulted in subsets of tags that represent or impact the same physical
variable, but they each represent the variable in a slightly different way.
For example consider the physical variable of gross airflow, the total amount of air
entering the boiler. It is clear from first-principles that this variable has a significant
impact on NOx (see Chapter 3). The forced draft (FD) fans deliver gross airflow to the
combustion process (see Section 2.5 "Fossil-Fired Power Generation"). The output of the
FD fans is derived from the boiler master signal. Forced draft output is specified along
with fuel flow by the fuel-air curve of the boiler. The DCS has five different representative
tags for the variable of gross airflow:
1) Fan vane position: The FD fans are constant speed fans, meaning that
the fan shaft turns at a constant speed while more or less air with more
or less initial spin can be dumped into the blades by opening or closing
the vanes.
2) Fan amps: If the inlet vanes are opened wider the flow is greater, and
more work is being done and the amperage must increase.
3) FD fan demand: The fuel-air curve gives a total air flow requirement,
which is characterized within the DCS as a demand signal for the FD
fan controller.
4) 02 trim: Prior to presenting the FD fan demand signal to the FD fan
controller, the operator is provided with a trim signal that can shift this
demand +-10% of its range. This trim allows the operators to add or
remove gross air at their discretion.
5) FD fan setpoint: This DCS signal, which is simply the addition of the
FD fan demand with the 02 trim, is presented the FD fan control logic
where a PID control loop maintains desired airflow.
Notice that all of these representations are in our essential tag list marked as MVs.
Clearly, these variables are not all independent. In addition to “Type” you will notice a
field in the Essential Tag List called “Group.” This field will be used to identify tags
which represent the same physical process variables. The following sections present a

104
methodology for reducing the MVs, DVs and SVs down to a minimal set which have the
greatest impact on combustion. All of the variable from a group can be removed from con¬
sideration, but at most one variable from each group will be allowed in our final variable
sets.
6.3 Datasets
Section 5.3 presented the sampling methodology for retrieving process data from the
data historian. This methodology was applied to produce a contiguous time series of 3
months of process data from 1/18/98 to 4/18/98 while the unit was operating continuously.
This data also included parametric testing of key MVs thought to affect the process out¬
puts of interest.
The 3 months of available data had to be divided into datasets for training, cross-vali¬
dation and testing of each of the models. Due to the temporal nature of the data and the
dynamic nature of the models being considered, the datasets would each have to be contig¬
uous in time yet disjoint from one another. In order to accomplish this while trying to keep
both the cross-validation and testing data as close as possible to the training data, the data
was divided into 3 contiguous and disjoint time regions as follows: the first 2 weeks were
used for cross-validation, weeks 2 thru 10 were used for training, and the remaining 2
weeks were used as a blind test set.
Each time region was then sampled according to the method outlined in Section 5.3 to
produce training, cross-validation and testing datasets. The cross-validation and testing
regions were only sampled once, producing only one cross-validation and one testing
dataset common to all models considered. The training region, however, was sampled

105
D = 30 times, producing 30 training datasets as 30 observations of the underlying pro¬
cess. These 30 observations were sampled by offsetting the start of each sampling by 2
seconds from the previous, where ±2 seconds is the approximate resolution of the data
historian.
6.4 Performance Criteria
The objective of this section is to develop the “best” possible models based on the
model definitions presented in Section 6.2. To this end, many models will be developed
for each of the six model definitions; these models will then compete for being the “best”
for a particular model definition. The “best” models will then be used in Chapter 7.
This section will formally present the criteria by which the models will be judged.
Clearly, these criteria should be closely related to the objective function used to train the
models.
6.4.1 Objective Function
All of the model definitions considered are for predictors, i.e., they solve the general¬
ize nonlinear regression problem. Model training will therefore fall into the category of
supervised learning, where there is a known desired response for the model d(t) e 91 .
Although there are many objective function which can be applied to this problem, the
ordinary mean-squared-error objective function is by far the most common and successful.
The MSE objective function will be given by
J =
¿1N t= j! = i
(77)

106
where y(t) e 91 is the output of the model, and T is the temporal length of the training
dataset. Note that both steady-state and dynamic models will be developed, therefore the
objective function has been written in a form that is applicable for both. When the model
is steady-state, T can be considered the number of samples in the training dataset.
The subsequent sections will be training multiple models for a given model definition,
and then comparing them to see which performed the best. The MSE could serve as a suit¬
able candidate for our “best” metric. There are a few problems with comparing models
based on MSE, however:
1) It is difficult to derive meaning from the MSE value associated with
testing an individual model.
2) It cannot be used to compare the performance between individual out¬
puts, because the MSE is significantly affected by the power of the indi¬
vidual output variables.
3) It cannot be used to compare the performance of models tested on sepa¬
rate datasets, since the power of each variable will be highly variable
across datasets.
Since all of the models will be tested against a common dataset, the third problem will
not be a factor here. The most compelling reason not to use MSE as a performance crite¬
rion, is simply one of interpretation. It is very difficult to tell how well a model is perform¬
ing relative to the process. We will instead introduce two related performance criteria
which will not only serve as a metric with which to compare models against one-another,
but whose values possess simple physical interpretation.
6.4.2 Normalized Mean-Squared Error
The first metric is a close relative to the MSE, called the normalized mean-squared-
error (NMSE). This metric normalizes the MSE relative to what is sometimes called the

107
“trivial predictor.” This predictor is simply an estimate for the process statistical mean,
and the NMSE is given by
T
NMSEy: = X
t = 1
(4(0-MO)2
(!b-T/(0)2
(78)
where pf is the / -th element of the vector ft^, which is the statistical mean of the desired
response. Notice that the NMSE is based on individual output variables which allows us to
compare the model performance for specific model outputs.
The main advantages to the NMSE over the MSE are:
1) Differences in the power of individual outputs is normalized out.
2) Simple interpretation can be used to provide a feel for model perfor¬
mance based on the NMSE value alone. The best performance a model
can have would be to match the desired response precisely, which
would result in a NMSE = 0. Similarly, a NMSE = 1 indicates that
the model is doing no better than predicting the process mean. While
this is clearly not the worst performance a model can have, it is certainly
cause for speculation.
6.4.3 Correlation
Both the MSE and NMSE metrics represent the model’s error with respect to the
desired response. The error of a model describes how well it matches the actual value of
the desired response. It is very possible for the model to have a large error, and still con¬
tain valuable information about the process. This is particularly true when using models
for optimization and control applications.
Consider the example illustrated in Figure 15. It is clear that the NMSE and MSE
between variable d and model m, will be lower than the respective errors between d and
m2 â–  It is also clear that m2 has captured more of the dynamic information of d.

108
This example illustrates the need for an additional metric. We are building models
which are not going to be used for their ability to forecast the actual value of a variable,
but rather to explain the cause-and-effect relationships between the model’s input and out¬
put variables. For such applications, correlation provides a better metric. The correlation
of a model can be calculated as follows
where
(79)
T
?y = V). (80)
t= l
9
A R~ = 0 suggest that there is no linear relationship between the model’s output and
y¡
2
the desired response, while R~ = 1 suggests that they are identical (in a linear sense).
Both NMSE and R provide useful metric with which to judge model performance. In
fact, used together they provide direct insight into the bias-variance dilemma of model
development. These will be the primary metric used to asses model performance in this
section.

109
6.5 Learning Algorithm
The learning algorithm, or algorithms for a neural-network-based control application
must be able to deal with the following configurations:
1) Training: where the initial parameters of models and/or controllers are
determined for the first time.
2) Retuning: where the model and/or controller parameters are adjusted
online to allow the system to compensate for changes in its operating
environment.
3) Optimization: where the next MV setpoint, or setpoint trajectory must
be determined.
The proper choice of a learning algorithm needs to consider the specific requirements
for these three situations. The first decision is whether the learning algorithm should be
incremental or batch. It is the incremental learning schemes that are usually seen in close
relationship with online adaptation as seen in retuning and optimization. Indeed, the field
of adaptive controls applies incremental schemes almost exclusively.
By contrast, batch learning schemes are usually considered to be committed to offline,
nonadaptive operation, as found in training. However, it is equally possible to apply batch
learning in adaptive configurations running parallel to the plant in operation [31]. In fact,
there are many reasons to consider batch schemes when dealing with complex nonlinear
control systems, most importantly batch schemes are required for:
1) second-order approximation methods requiring line searches. This con¬
cerns particularly the conjugate gradient methods and Powell’s algo¬
rithm. These algorithms are not able to keep the search directions
conjugate without line searches.
2) most global optimization methods. Except for simulated annealing and
evolutionary algorithms, global optimization algorithms require the cost
function to be evaluated at various points of the state space with the
results compared.

110
The computational complexity lost by not using second-order approximations is by
itself reason enough to restrict our learning to batch schemes. For first-order methods,
there are no estimates of convergence even for exact quadratic functions. In other words,
they can converge arbitrarily slowly. Second-order methods can be compared by the num¬
ber of cost function evaluations that are necessary to reach the minimum of a quadratic
function. Hryces [31] demonstrates how the conjugate gradients algorithm can be applied
to speed up both the backpropagation and backpropagation through time algorithms.
These accelerations use only n(K+ 1) cost function evaluations, where n is the number
of free parameters and K is the number of evaluations required by the line search.
It should be noted that the first-order gradient descent is not the last resort for incre¬
mental learning. The Kalman training algorithm proposed by Singhal and Wu [57] and
extended and applied to several neurocontrol problems by Puskorius and Feldkamp
[51 ][52] exploits second-order information by building a parameter covariance matrix, an
analogy to the Hessian matrix in variable metric local optimization methods. Even with
sophisticated methods such as the Kalman training algorithms, incremental learning has
many practical limitations. For the identification of strongly nonlinear plants of realistic
size and difficulty, like those considered in this work, these methods can require tens of
thousands of sampled measurements [52], Furthermore, they do no support global optimi¬
zation.
The robustness of a neurocontroller solutions will be predominately a function of its
performance during training, retuning and optimization. Consider the consequences of:
1) Showing up and trying to win operator confidence with poorly trained
models.

Ill
2) Having models which had worked properly being placed in closed-loop
control of a power plant, which suddenly stop working following a
retraining.
3) A controller not able to see an optimal control trajectory because of the
non-convex nature to the performance surface which exist between two
distinct operating modes of the unit.
Since the learning schemes described above will be applied to highly complex nonlin¬
ear systems, and thus highly non-convex performance surfaces, global optimization will
be considered a requirement. The notation ArgMin^{J} shall be used to refer to the glo¬
bal optimization of cost functional J with respect to the variables . The global optimiza¬
tion algorithms used throughout the rest of this work are detailed next.
6.5.1 Global Optimization Algorithm
An extremum of a completely arbitrary function of real-valued arguments is not com¬
putable because it requires a combinatorial search in infinite spaces. Global optimization
approaches are typically forces to make explicit or implicit assumptions about the function
or are satisfied with trying to allocate the available resources so that the probability of
finding the optimum is as high as possible.
If an exhaustive search is not feasible, it is logical to try to cover the cost function
domain as far as possible with the given resources and then to use some local optimization
method. Then it is sufficient if every interesting local basin of attraction is hit by a starting
point. For n-dimensional domains, the trivial coverage by all combinations of only two
values per dimension leads to generating 2n starting vectors. These correspond the cor¬
ners of a hypercube. Yet sparser coverage is by 2n vectors corresponding to starting
points in the centers of the hypercube sides.

112
For training and retuning learning tasks, these methods are still far too computation¬
ally burdensome. For these configurations we will have to be satisfied with simply gener¬
ating IVs starting vectors uniformly. For optimization learning tasks, however, there are
far fewer free parameters, one for each MV. For this configuration, global optimization
shall be determined by running 2N1"' local optimizations from starting vectors x with
elements
1 i = 2/—1
-1 i - 2j •
0 otherwise
(81)
6.5.2 Local Optimization Algorithm
The conjugate gradients algorithms used for all local optimization is given by:
I: Given a set of weights We, determine a search direction using the
Polk-Ribere algorithm presented in Section 3
II: Optimize the objective function J( W + IW^,) with respect to the
sca¬
lar variable /, using the line search algorithm presented in Section
2.1.2.5
III: Update the weights according to JV(
e + 1
iWl
We + .,re
IV: Repeat until either e > e"'ax or J( W ) - J( JPe + /) < A f"
where e is the index of iteration, e'ax is the maximum number of epochs, and Aj"“" is
the minimum performance improvement for early stopping.

113
6.6 Variable Pruning
Appendix identified the initial tag sets to be used as MVs, DVs and SVs. The goal of
this section is to develop a methodology for pruning these tag sets into the smallest set of
variables required to produce accurate models for the model definitions presented in Sec¬
tion 6.2. At this point, we have exhausted our first-principles understanding of the process,
which was used to create the Essential Tag List. The variable pruning methodology will
therefore consist of training multiple models with different combinations of inputs, and
selecting the best models based on the performance criteria above. The methodology is an
educated direct search approach.
There are a few rules and assumptions that will be used to limit the search:
1) Only one variable from each group will be allowed for each type.
2) Operators do not want to manipulate more than 8
3) Variables for the best steady-state model will be the best variables for
dynamic modeling.
4) Variables for the best MLP(15,5) model will be the best variables for all
model architectures considered.
Here, MLP(15,5) is defined as a MLP with 2 hidden layers, with 15 processing ele¬
ments in the first layer and 5 in the second. These assumptions are inherently flawed.
Given limited computational resources, however, they are reasonable assumptions.
The variable pruning methodology will consist of two algorithms. Both algorithms use
a MLP(15,5) using the steady-state CV model definitions with one modification; the CO
variable has been moved from the SV set to the CV set, i.e., from the models input to its
output. This modification was made to force the pruning algorithms to consider each vari¬
ables effect on both NOx and CO.

114
6.6.1 Group Pruning Algorithm
The goal of the group pruning algorithm is to identify which variable groups in the
Essential Tag List have the greatest impact on the performance of the model. The group
pruning algorithm was implemented as follows:
I: Train the MLP(15,5) 1^“"' times for each unique group in the Essen¬
tial Tag List
II: For each group model, select the “best group” model as that with the
lowest cross-validation MSE
III: Calculate the correlation of each “best group” model across the blind
testing dataset, and define the “removal group” model as the “best
group” model with the highest correlation
IV: Remove all variables from the Essential Tag List that are in the group
associated with the “removal group” model
V: Repeat this steps I-V until there are only Ng'oups unique groups
remaining
This algorithm was run with N1u"s = 10 and ]\fg'°,ips = 10. Notice that there are a
total of 15 unique group models (ignoring NOx and CO), resulting in 150 model trainings
for the first iteration. This results in 15 “best group” models. The total number of model
trainings required by this algorithm for all iterations is therefore
10(15 + 14+ 13 + 12+ 11) = 650. (82)
The average training run requires 30 minutes on a 300MHZ intel based workstation,
requiring approximately 14 dedicated days of CPU time. It is easy to see the limitations to
direct search methodologies. There is a lot of work recently in the literature applying
genetic algorithms to discrete optimization problems like variable and model architecture
selection. The author believes that these techniques have a great deal of merit, and will
result in significant improvements to model building methodologies. This is still a

115
research topic, however, and not the research topic chosen for this work. The “bruit force”
method has been selected because of its robustness to parameterization.
The computational complexities of this algorithm and the algorithms discussed in sub¬
sequent sections are managed using a distributed computing environment. Distributing the
training of a neural network is yet another research project. Rather that distributing the
training of a single training run, the multiple training runs were distributed across a net¬
work of computers. The author had evening access to more than 50 intel based worksta¬
tions connected through a local area network. These machines were used in parallel to run
this and subsequent algorithms. This algorithm, for example, was run on a single Satur¬
day, when the machines were not in use.
The results of this algorithm are provided in Figure 16. The algorithm was able to
remove 3 groups from the model without degrading its performance. In fact, there is an

increase in the generalization ability of the model. The 3 groups removed are listed in
Table 2 in the order they were removed.
116
Table 2: Types in order they were removed.
Order
Group
1
Fuel
2
SAS
3
PAS
There could be concern that the PAS group was removed using this algorithm. There is
an unbalanced number of variables in this group. It is not surprising that removing these
variables might increase the models ability to generalize. Given our restrictions on the
number of MVs, however, the PASs are not a good candidate for control. They were
included primarily to see if the models required them as DVs. Since they can be removed
with only minimal impact on modeling results, there is little reason at this time to investi¬
gate their impact further. They are a good candidate for future expansion of our control¬
lers, however.
6.6.2 Representation Pruning Algorithm
The goal of the representation pruning algorithm is to identify the best variable repre¬
sentation within each remaining group. The representation pruning algorithm was imple¬
mented as follows:
I: Select the first group in the reduced Essential Tag List
II: If there is only one unique representation in the group then select the
next group and goto step II
III: For each unique representation in the selected group, train the
MLP(15,5) 10 times with the variables for this representation as the
only input variables from the selected group; The resulting models will
be called “representation” models

117
IV: For each “representation” model, the training result with the lowest
cross-validation MSE is selected as the “best representation” model
V: Calculate the correlation of each “best representation” model across
the blind testing dataset, and remove all variables associated with
every “representation model” except for the “best representation”
model with the highest correlation
VI: Select the next group and goto step II
Notice that there are 6 remaining groups with more than one unique representation,
containing a total of 12 unique representations. This algorithm, therefore, requires 120
model trainings. The results after pruning the variable representations for each of the 4
groups are illustrated in Figure 17. As illustrated in this figure, each group was pruned
down to a single variable representation without loss of fidelity.
Figure 17: Representation pruning algorithm results.
6.6.3 Final Variable Sets
The final variable selections after the group and representation pruning algorithms are
provided in Table 3. These selections are of interest from a first-principles perspective.
Notice that the MVs contain the primary controls over gross air, i.e., the FD fans, and

118
recall from Chapter 3 that excess gross airflow is a primary cause of NOx formation. The
MVs also contain over-fire air (OFA), and gas recirculation (GR) air, which recall from
Section 2.5 "Fossil-Fired Power Generation" that the primary reason for installing the
OFA and GR subsystems is to better control NOx.
Table 3: Final variable selections after pruning.
Manipulated Variables
Disturbance Variables
State Variables
Control Variables
1/3 OFA Damper Pcs
Ambient Air Press
Sec Air Temp Side A
CEMNOx
2/3 OFA Damper Pcs
Ambient Air Temp
Sec Air Temp Side B
FD Fan 2A Inlet Vane
Bnr Atm Stm Press
OEM CO
FD Fan 2B Inlet Vane
Bnr Atm Stm Temp
Generated MW
GR Fan 2A Inlet Dmpr Pcs
OEM Barometric Pressure
Windbox Pressure
GR Fan 2B Inlet Qnpr Pcs
Cond Back Pres - Side A
GR Fan Hppr Dmpr A Pcs
Cond Back Pres - Side B
GR Fan Hppr Dmpr B Pcs
Fuel Gas Flew Indication
Fuel Oil Flow Indication
Fuel Temp Fired
Furnace Pressure
There are also some surprises. Consider, for instance, burner atomizing steam pressure
and temperature. During the initial operator interviews it was thought that the atomizing
steam would have a nominal affect on combustion, but their impact on the models devel¬
oped has been significant. We latter verified this impact by manipulating these variables at
the plant.
6.7 Architecture Selection
The final variable selections, identified through the pruning algorithms in the last sec¬
tion, will be considered the best variables for all model architectures. The author recog¬
nizes that the best variables could be a function of the model architecture. The
computational complexity involved in variable pruning, however, makes finding the
topology specific variables impractical.

119
Their are a number of architecture specific parameters that will require selection for
each architecture, however. Identifying the best architecture parameters will employ a
similar methodology to that used for variable pruning, i.e., using a direct search of the fea¬
sible parameter space. The following presents a detailed design for the model architectures
investigated in the study, along with the results used to identify their best parameteriza¬
tion.
6.7.1 ARMA Model
The ARMA model considered in this study is given by
N"+]^T Nm
p(t) = Y an Y i + (83)
n = 1 m = 0
^ Nu N
where ü = {mv, dv} e 9? "are the model inputs, an g 91 y, Vn are the auto-regressive
—^ N
coefficients, bm e 91 Vm are the moving average coefficients, Nn is the number of
N
auto-regressive taps, Nm is the number of moving average taps, and t e 91 v is a constant
vector.
6.7.1.1 Parameter selection
Given the best set of MVs and DVs from variable pruning and only a single CV, there
are only two parameters to be identified for the ARMA model, Nn and Nm. The algorithm
used to determine N was as follows:
I: Set Nn = 0
II: Train ARMA(Nn, 1)
III: Increment Nn by 1 and repeat II until Nn = 10

120
IV: Calculate the correlation of each model in {ARMA(Nn, 1 )}VA, across
*
the blind testing dataset, and assign Nn, the optimum number auto¬
regressive taps according to the model with the highest correlation
Figure 18: Results of auto-regressive taps search algorithm for the NOx ARMA Model.
Figure 20 illustrates the results of this algorithm for the dCVModel. For the dCV-
Model definition, ARMA(5, 1) was chosen as the model with the highest correlation, tak¬
ing into consideration our desire to minimize the number of taps. This algorithm was
repeated for the dSVModel definition, and the combined results are summarized in Table
6. Notice that only the dynamic model definitions have been considered here, since the
ARMA is a dynamic model.

121
Table 4: Results of auto-regressive tap search algorithm for all dynamic models.
Model
PEs
Train NMSE
Train R
dCVModel
5
0.903
0.194
dSVModel
4
0.932
0.178
Given the optimum number of auto-regressive taps, Nn, for each model definition, a
*
similar algorithm was used to determine the optimal number of moving average taps, Nm .
This algorithm can formally be stated as follows:
I: Set Nm = 1
II: Train the MLP(Nn, NJ
III: Increment Nm by 1 and repeat II until Nm = 10
*
IV: Calculate the correlation of each model in {ARMA(Nn, Nm)}yNm
*
across the blind testing dataset, and assign Nm , the optimum number
of moving average taps, according to the model with the highest corre¬
lation
Figure 19: Results of the moving average tap search algorithm for NOx ARMA Model.

122
Figure 21 illustrates the results of this algorithm for the dCVModel. For the dCV-
Model definition, ARMA(5, 3) produced the model with the highest correlation. Once
again, the algorithm was repeated for each of the dynamic model definitions, and the
results are summarized in Table 7. The models summarized in this table will be taken as
the best ARMA models for their corresponding model definitions.
Table 5: Results of moving average tap search algorithm for all dynamic models.
Model
PEs
Train NMSE
Train R
dCVModel
3
0.888
0.225
dSVModel
4
0.902
0.219
6.7.2 Multi-layer Perceptron
The multi-layer perceptron (MLP) architecture considered in this study will consist of
2 hidden layers with tanh processing elements (PEs), and an output layer with a linear PE.
Formally, the architecture is given by
y
= #,a(#/'2a(#/,1Ó + S''1) + bhl) + b\
(84)
v Nu sb 1 Nh x N
where u = {mv, dv} e 91 " are the model inputs, W e 91 11 " is the matrix of
â–º /it N,
weights for the first hidden layer, b e 91 ' are the bias values for the first hidden layer,
- £V> Al2 X All •
h2 N„-
€ 5H 'is the matrix of weights for the second hidden layer, b e 5H are the
bias values for the second hidden layer, e 91 y 1,2 is the matrix of weights for the
M
second hidden layer, b e 91 are the bias values for the second hidden layer, Nu is the
total number of MVs and DVs in the input layer, Nh] is the number of PEs in the first hid-

123
den layer, Nh2 is the number of PEs in the second hidden layer, N is the number of PEs
in the output layer, and a is the tanh logistic function.
6.7.2.1 Parameter selection
Given the best set of MVs and DVs from variable pruning and only a single CV, there
are only two parameters to be identified for the MLP, Nh, and Nhl â–  The algorithm used to
determine ./V/, ] was as follows:
I: Set Nhl = 5
II: Train the MLP(Nhx, 5) 10 times
III: Select MLP*(Nhx, 5) as the model with the lowest cross-validation
MSE from the 10 training results
IV: Increment Nhx by 2 and repeat II until Nhx = 25
V: Calculate the correlation of each model in {MLP*(Nhx, 5)}VA,-
*
across the blind testing dataset, and assign Nh ¡, the optimum number
of hidden PEs in the first layer, according to the model with the high¬
est correlation

124
1.000
0.900
0.800
0.700
0.600
0.500
0.400
0.300
0.200
0.100
0.000
5
7
9
11
13
15
17
19
21
23
25
—a—Train R
0.589
0.643
0.666
0.757
0.793
0.817
0.875
0.876
0.894
0.884
0.887
—■—Test R
0.475
0.512
0.563
0.635
0.668
0.750
0.774
0.793
0.786
0.791
0.752
PEs
Figure 20: Results of hidden layer #1 PE search algorithm for the NOx MLP Model.
Figure 20 illustrates the results of this algorithm for the ssCVModel. For the ssCV-
Model definition, MLP( 19, 5) produced the model with the highest correlation on the
blind test set. This algorithm was repeated for each of the steady-state model definitions,
and the results are summarized in Table 6. Notice that only the steady-steady model defi¬
nitions have been considered here, since the MLP is a static model.
Table 6: Results of hidden layer #1 PE search algorithm for all steady-state models.
Model
PEs
Train NMSE
Train R
Test NMSE
Test R
ssCVModel
19
0.130
0.876
0.182
0.793
ssSVModel
15
0.252
0.766
0.331
0.712
ssISVModel
17
0.087
0.892
0.143
0.844
ssIMVModel
23
0.293
0.698
0.354
0.637
*
Given the optimum number of hidden processing elements in the first layer, Nh ,, for
each model definition, a similar algorithm was used to determine the optimal number of
*
hidden PEs in the second hidden layer, Nh2 . This algorithm can formally be stated as fol¬
lows:

125
I: Set Nh2 = 2
II: Train the MLP(Nhx, Nhl) 10 times
*
III: Select MLP*(Nhx, Nhl) as the model with the lowest cross-validation
MSE from the 10 training results
IV: Increment Nhl by 1 and repeat II until Nh2 =10
*
V: Calculate the correlation of each model in {MLP* (Nh j, Nh2)} vtv,!2
*
across the blind testing dataset, and assign Nhl, the optimum number
of hidden PEs in the second layer, according to the model with the
highest correlation
Figure 21: Results of the hidden layer #2 PE search algorithm for NOx MLP Model.
Figure 21 illustrates the results of this algorithm for the ssCVModel. For the ssCV-
Model definition, MLP* (19, 7) produced the model with the highest correlation on the
blind test set. Once again, the algorithm was repeated for each of the model definitions,
and the results are summarized in Table 7. The models summarized in this table will be
taken as the best MLP models for their corresponding model definitions.

126
Table 7: Results of hidden layer #2 PE search algorithm for all steady-state models.
Model
PEs
Train NMSE
Train R
Test NMSE
Test R
ssCVModel
7
0.135
0.895
0.298
0.801
ssSVModel
4
0.265
0.786
0.401
0.723
ssISVModel
3
0.219
0.791
0.365
0.732
ssIMVModel
9
0.296
0.705
0.336
0.641
6.7.3 Time-Delayed Neural Network
As presented in Chapter 2, the most commonly applied time-delayed neural network
(TDNN) is simply a MLP with a tapped-delay-line (TDL) preprocessor at its input. This
architecture can be formally presented as follows
p = JpyG(Jph2a(jthlTDL(ñ) + S7'1) + b1'2) + by, (85)
Nu Nu x Nt
where TDL'.yi “ —» 91 is the TDL mapping, and NT is number of taps.
6.7.3.1 Parameter selection
The TDNN architecture, therefore, adds a single additional parameter to the MLP
architecture, NT. Clearly, the MLP parameters A^, and Nhl will be function of Nr. Once
again, however, NT will also be function of these parameters, and this study will simplify
this problem by assuming that the best parameter for the MLP will also be optimal for the
TDNN.
*
Given the optimum number of hidden processing elements for the MLP, Nh, and
*
N/i2 , for each model definition, the algorithm used to determine NT was implemented as
follows:
I: Set NT= 2
II: Train the TDNN(Nh j, Nhl, NT) 10 times

127
III: Select TDNN*(Nh], N/l2, NT) as the model with the lowest cross-val¬
idation MSE from the 10 training results
IV: Increment NT by 2 and repeat II until Nr = 20
V: Calculate the correlation of each model in
* *
{TDNN* (Nh!, N/a, AV)}wvr across the blind testing dataset, and
*
assign Nt, the optimum number of taps, according to the model with
the highest correlation
Figure 22: Results of tap search algorithm for NOx TDNN Model.
Figure 22 illustrates the results of this algorithm. For the dCVModel definition,
TDNN* (19, 7, 8) produced the model with the highest correlation on the blind test set.
The algorithm was repeated for each of the dynamic model definitions, and the results are
summarized in Table 8. The models summarized in this table will be taken as the best
TDNN models for their corresponding model definitions.

128
Table 8: Results of tap search algorithm for all dynamic models.
Model
Taps
Train NMSE
Test NMSE
Train R
Test R
dCVModel
8
0.067
0.124
0.928
0.878
dSVModel
4
0.116
0.207
0.889
0.796
6.7.4 Gamma Neural Network
The Gamma Neural Network (GNN) is similar to the TDNN considered above, with
two fundamental differences: 1) the TDL feed-forward memory mechanism is replaced
with an HR Gamma Filter (GF), and 2) the memory is embedded within the hidden layers
of the architecture. This study will consider an architecture containing two GFs, one in the
input layer, and the second in the first hidden layer. This architecture can be formally pre¬
sented as follows
p = JPyoClPh2Gh(cy(TP,''Gu(ú) + bhX)) + bh2) + by, (86)
where Gu\\R^" -> iR^'1 X Nah is the hidden
layer GF, and these memories have NGu and NGh memory taps, respectively.
6.7.4.1 Parameter selection
The GNN, therefore, adds two additional parameter to the MLP architecture, NGu and
Nch . We again simplify the parameter selection problem by assuming that the best
parameter for the MLP will also be optimal for the GNN.
*
Given the optimum number of hidden PEs for the MLP for each model definition, Nh ¡
*
and N,i2 , the algorithm used to determine NCu was implemented as follows:
I: Set NGu = 2

129
II: Train a GM(iV/;], Nh2, NGu, 4) 10 times
* *
III: Select GM*(Nhx, Nh2, NCu, 4) as the model with the lowest cross-
validation MSE from the 10 training results
IV: Increment NGu by 1 and repeat II until NCu =10
V: Calculate the correlation of each model in
* *
{GM*(N¡ñ, N/i2, NGu, 4)}VCw across the blind testing dataset, and
*
assign NGu , the optimum number of taps, according to the model with
the highest correlation
1.000
0.900
A A * A A A A A *
0.800
0.000
0
1
2
3
4
5
6
7
8
9
10
—a— Train R
0.893
0.916
0.933
0.936
0.931
0.930
0.936
0.934
0.936
0.930
0.940
â–  TestR
0.804
0.832
0.834
0.846
0.835
0.835
0.831
0.830
0.819
0.824
0.819
PEs
Figure 23: Results of taps search algorithm for NOx GNN model.
Figure 23 illustrates the results of this algorithm. For the dCVModel definition,
GM*( 19, 7, 3, 4) produced the model with the highest correlation on the blind test set.
This algorithm was repeated for each of the dynamic model definitions, and the results are
summarized in Table 9.

130
Table 9: Results of taps search algorithm for all dynamic models.
Model
Taps
Train NMSE
Test NMSE
Train R
Test R
dCVModel
3
0.065
0.156
0.936
0.846
dSVModel
5
0.129
0.202
0.875
0.795
Given the optimum number of memory taps in the input layer, NCu, the optimal num¬
ber of memory taps for the first hidden layer were determined as follows:
I: Set Nch = 2
II: Train the GM(Nhl, Nhl, NCu, Nch) 10 times
* * *
III: Select GM*(Nhx, N/l2, NGu, NGh) as the model with the lowest cross-
validation MSE from the 10 training results
IV: Increment NCh by 1 and repeat II until NGh =10
V: Calculate the correlation of each model in
* * *
{GM* (Nh j, N/i2, NGu, NGh)} v/vc/1 across the blind testing dataset, and
*
assign NGh, the optimum number of memory taps in the first layer,
according to the model with the highest correlation
Figure 24: Results of hidden taps search algorithm for NOx GNN model.

131
Figure 24 illustrates the results of this algorithm for the dCVModel. For the dCV-
* *
Model definition, GM*(Nhl, Nh2, 3, 4) produced the model with the highest correlation
on the blind test set. Once again, the algorithm was repeated for each of the model defini¬
tions, and the results are summarized in Table 7. The models summarized in this table will
be taken as the best GNN models for their corresponding model definitions.
Table 10: Results of hidden taps search algorithm for all dynamic models.
Model
Taps
Train NMSE
Test NMSE
Train R
Test R
dCVModel
4
0.034
0.086
0.960
0.908
dSVModel
5
0.045
0.095
0.952
0.903
6.7.5 Nonlinear State-Space Model
As presented in Chapter 2, the nonlinear state-space (NLSS) model actually consists of
two separate models; a state evolution model and a output observation model. Here we
consider the case where both of these models are MLPs, each with a single hidden layer.
Formally, the NLSS model considered for this study is given by
*(0 = JpxG(JphxK0 + ^hx) + ^
(87)
j>(0 = Tfa(^,yKt) + bhy) + by,
(88)
where r(t) = {u(t),x(t- 1)} e 91 1 and s(t) = {u(t), x(t)} e 91 " Nx are state
hx ^N,r x (N„ + Nr) ±hx ^Nhr x , ±x „jVr
representation vectors; IFeiR ,SeiR , ir e iR and b e 91
are the weights of the state evolution model; W e 9?
V _ mNhyX(Nu+Nx) e ^Nhy
#6^X^and %ye*t y are the weights of the output observation model; N is the
total number of MVs and DVs in the input layer; Nhx is the number of hidden PEs in the

132
state evolution model; Nhy is the number of hidden PEs in the output observation; N is
the number of CVs in the output, and ct is the tanh logistic function.
There are therefore 3 parameters which need to be determined for the NLSS model: 1)
the number of hidden states Nx, 2) the number of hidden PEs in the state evolution model
Nhx, and 3) the number of hidden PEs in the output observation model Nhy. Once again,
these parameters will be determined using an exhaustive search methodology. We begin
by fixing the number of hidden PEs in both models to 4, and determine the number of hid¬
den states. Defining NLSS(NX, Nhx, Nhy) as a NLSS model as described above, the num¬
ber of hidden state can be determined as follows:
I: Set Nx = 2
II: Train the NLSS(NX, 4, 4) 10 times
III: Select NLSS*(NX, 4, 4) as the model with the lowest cross-validation
MSE from the 10 training results
IV: Increment Nx by 1 and repeat II until Nx = 12
V: Calculate the correlation of each model in {NLSS* (Nx, 4, 4)} VN
*
across the blind testing dataset, and assign N , the optimum number of
hidden states, according to the model with the highest correlation

133
1.000
^
0.800
0.700
0.600
0.300
0.000
2
3
4
5
6
7
8
9
10
11
12
a Train R
0.911
0.943
0.906
0.911
0.931
0.922
0.945
0.948
0.955
0.921
0.949
â–  Test R
0.704
0.738
0.783
0.772
0.806
0.773
0.781
0.777
0.725
0.752
0.730
States
Figure 25: Results of hidden states search algorithm for NOx NLSS model.
Figure 25 illustrates the results of this algorithm. For the dCVModel definition,
NLSS*(6, 4, 4) produced the model with the highest correlation on the blind test set. The
algorithm was repeated for each dynamic model definition, and the results are summarized
in Table 11.
Table 11: Results of hidden states search algorithm for all dynamic models.
Model
States
Train NMSE
Test NMSE
Train R
Test R
dCVModel
6
0.090
0.221
0.906
0.783
dSVModel
9
0.120
0.300
0.879
0.695
*
Given the optimal number of hidden states Nx, the optimal number of hidden PEs in
the state evolution model can be determined as follows:
I: Set Nhx = 2
*
II: Train the NLSS(NX, Nhx, 4) 10 times

134
III: Select NLSS*(NX, Nhx, 4) as the model with the lowest cross-valida¬
tion MSE from the 10 training results
IV: Increment Nhx by 1 and repeat II until Nln =12
V: Calculate the correlation of each model in {NLSS* (Nx, Nhx, 4)} Vv,lt
*
across the blind testing dataset, and assign Nhx according to the model
with the highest correlation
1.000
Z—-—'* '—* *
0.800
2
3
4
5
6
7
8
9
10
11
12
a Train R
0.871
0.898
0.894
0.901
0.931
0.958
0.977
0.988
0.975
0.966
0.984
—•—Test R
0.685
0.716
0.701
0.763
0.808
0.848
0.827
0.868
0.838
0.809
0.804
State Hidden PEs
Figure 26: Results of state hidden PEs search algorithm for NOx NLSS model.
Figure 26 illustrates the results of this algorithm. For the dCVModel definition,
NLSS*(6, 9, 4) produced the model with the highest correlation on the blind test set. The
algorithm was repeated for each of the model definitions, and the results are summarized
in Table 12.
Table 12: Results of state hidden PEs search algorithm for all dynamic models.
Model
States PEs
Train NMSE
Test NMSE
Train R
Test R
dCVModel
9
0.017
0.129
0.988
0.868
dSVModel
11
0.047
0.191
0.950
0.806

135
Finally, the optimal number of hidden states Nx and hidden PEs in the state evolution
*
model Nhx, the optimal number of hidden PEs in the output observation model can be
determined as follows:
I: Set Nhy = 2
II: Train the NLSS(N*X, Nhx, Nby) 10 times
III: Select NLSS*(NX, Nhx, Nhy) as the model with the lowest cross-vali¬
dation MSE from the 10 training results
IV: Increment Nh by 1 and repeat II until Nh =12
V: Calculate the correlation of each model in
* *
{NLSS*(NX, Nhx, across the blind testing dataset, and
*
assign Nhy according to the model with the highest correlation
1.000
A A * * A * A A A
' " " "
'
2
3
4
5
6
7
8
9
10
11
12
—a—Train R
0.896
0.939
0.959
0.965
0.976
0.975
0.963
0.974
0.971
0.963
0.966
—■—Test R
0.735
0.845
0.884
0.896
0.891
0.895
0.854
0.845
0.773
0.785
0.770
State Output PEs
Figure 27: Results of output hidden PEs search algorithm for NOx NLSS model.
Figure 27 illustrates the results of this algorithm. For the dCVModel definition,
NLSS*(6, 9, 5) produced the model with the highest correlation on the blind test set. The

136
algorithm was repeated for each of the model definitions, and the results are summarized
in Table 13.
Table 13: Results of output hidden PEs search algorithm for all dynamic models.
Model
Output PEs
Train NMSE
Test NMSE
Train R
Test R
dCVModel
5
0.030
0.094
0.965
0.896
dSVModel
4
0.070
0.135
0.926
0.864
6.8 Analysis
Figure 28 summarizes the final modeling results after architecture selection. Recall
that the objective of this section was to find the most accurate model for the six model def¬
initions required to implement our four control designs. The best models identified for
each definition are:
1) Steady-State SV Model: MLP*( 15, 4)
2) Steady-State CV Model: MLP*{ 19, 7)
3) Steady-State ISV Model: MLP*(\1, 3)
4) Steady-State IMV Model: MLP*(23, 9)
5) Dynamic SV Model: GM*( 15, 4, 3, 4)
6)Dynamic CV Model: GM*( 19, 7, 5, 5)

Figure 28: Best models for all model definitions by architecture.
The next chapter will implement each control design using these models.

CHAPTER 7
CONTROLLER IMPLEMENTATIONS
This section will implement the four control designs presented in Chapter 4, using the
“best” reference models developed in Chapter 6, and identify the “best” control design for
meeting the objectives outlined in Section 4.4 "Performance Criteria." The “best” control¬
ler will be determined as follows:
1) Using dynamic models as simulators for the plant, controller perfor¬
mance is quantified offline.
2) Allowing the controllers to manipulate actual plant values, controller
performance is quantified online.
7.1 Offline Quantification
To quantify the performance of the four control designs offline, a common dataset and
plant simulator need to be selected. The dataset should not include data that was used dur¬
ing training of the controllers or their underlying reference models. Recall that a blind 1-
week dataset was set aside during modeling for testing and that none of the models have
ever seen this dataset during training. In addition, the dynamic process models chosen for
offline quantification should not be models that were used as reference models for the any
of the control designs. Since the best models developed in Chapter 6 were used as refer¬
ence models, i.e., the models with the lowest cross-validation MSE from 10 separate train¬
ings, the offline quantification will use the second best models as the dynamic process
models.
138

139
Offline quantification will apply each controller to the dynamic process models across
the test dataset, and calculate the average NOx reduction along with the average CO pro¬
duction above the maximum CO constraint. Formally, the average NOx percent reduction
will be reported according to
A NOx = —— ¿ NOx*(t)-NOx(t), (89)
Tf~ T't = t,
where T¡ is the start of the test dataset, Tj- is its end, NOx(t) is the value of NOx at time t
predicted by the dynamic process models in response to the actual MV setpoints mv(t),
and NOx*(t) is the predicted value of NOx in response to the controllers optimal MV set-
points inv (t) .
Similarly, the average CO above the maximum constraint will be reported according
to
rco]
i Tf r
Tf- T ¿
J ‘t=T:
CO*(t)-CO
v 0
CO*(t) > CO
else
(90)
where CO*(t) is the value of CO at time t predicted by the dynamic process models in
response to the controllers optimal MV setpoints mv*(t), and CO"'ax is the max limit set
for CO. All of the results presented are for a CO maximum of 500ppm.
The optimal MV, SV and CV trajectories across the test dataset produced by each con¬
troller are estimated by:

140
I: Set optimal MV, SV and CV trajectories to their actual values
(i) = mv(t)
sv*(i) = sv(t)
cv* (t) = cv(t)
(91)
II: For t = T¡ to Tj— 1
^ â– T'
i: Get the optimal MVs from the controller, mv (t)
N 5¡€
ii: Use the SV model to calculate the resulting SVs, sv (t + 1)
- â–  A ^
iii: Use the CV model to calculate the resulting CVs, cv (t + 1)
iv: Copy forward the MVs to carry forward the process state to initialize the
next step
(92)
V: Next t
The results of applying this algorithm to each controller are illustrated below. Figure
29 presents the average NOx reduction ANOx, while Figure 30 illustrates the average CO
above its max f CO~\. These are not particularly encouraging results. While each control¬
ler did manage to reduce NOx, the reductions were quite small. Furthermore, the control¬
lers seemed to have even less effect on the CO above 500ppm. The key question at this
stage is to determine whether these results represent:
1) all of the potential NOx reductions inherent to the process,
2) a problem with the control design, or
3) a problem with the reference models for the underlying process.

141
0.000
-0.002
-0.004
-0.006
-0.008
-0.010
-0.012
-0.014
-0.016
-0.018
Baseline
Steady-State
MPC
Steady-State
MIC
Dynamic MPC
Dynamic
MRAC
â–¡ Avg NOx Reduction
0.000
-0.013
-0.016
-0.009
-0.010
Figure 29: Average NOx reduction over testing dataset.
To help answer this question, a second run of the offline quantification algorithm was
run. This time, however, the test models were replaced with the reference models used to
develop each controller. These results are presented in Figures 31 and 32.
Figure 30: Average CO above max over testing dataset.

142
Figure 31: Average NOx reduction over testing dataset using train and test models.
Clearly, there is a problem. The following observation can be made:
1) the controllers appear to be working fine, and
2) the reference models and the test models are providing inconsistent
knowledge about the process.
Avg CO Above Max
Figure 32: Average CO above max over testing dataset using train and test models.

143
Chapter 8 addresses these issues in more detail, but for now the quantification of the
current controllers continues.
7.2 Online Quantification
Online quantification will be restricted to measuring the controllers ability to affect the
steady-state performance of the unit. Quantifying online performance will require running
online experiments, where MVs are moved and the resulting CVs measured. Each experi¬
ment will have to be of relatively short duration, since the longer an experiment takes the
less likely it is that steady-state conditions will be maintained. In order to compare the per¬
formance of different controllers, which will invariable have to perform their actions
under different conditions, baseline conditions will be established prior to each experi¬
ment. These baseline conditions will be stated in terms of MVs since we have no direct
control over DVs. To account for differences between the DVs between individual experi¬
ments, the MVs will be returned to the baseline conditions prior to each experiment.
Each online experiment will follow the following protocol: The unit operator is asked
to bring the unit to steady-state conditions, i.e., holding all MV setpoints constant. After
the unit has reached steady-state, measurements are taken to establish baseline conditions,
2k.bciS6 ^base
mv and cv . The controller is then queried for MV setpoints, mv (t), which will
be applied by the unit operator.
New setpoints will be repeatedly queried and applied until the unit has once again
returned to steady-state. The frequency with which an operator carries out this process,
will be at their discretion. If the operator feels that individual MV setpoints can not be
made, then these setpoints are constrained and the controller is queried for another set of

144
MVs. When none of the MV setpoints can be made, or the controller has no new advice,
i.e., it has saturated against its constraints, then the unit is at steady-state and the experi¬
ment is terminated.
^cxp
The CVs are once again measured to establish experiment conditions, cv . The dif¬
ference between the experiment and baseline conditions will be called the experiment
delta, A^p. The operators are then asked to return the MV setpoints back to their baseline
cv
±t)ClS6
conditions mv , and a third measurement is taken to establish validation conditions
±vcil
cv . The difference between the experiment and validation conditions will be called the
validation delta, A^f /. Notice that if the validation conditions match the baseline condi-
CV
tions, then the experiment and validation deltas will have identical magnitude and sign.
Controllers are then compared based on their ability to affect the CVs relative to these
baseline values across multiple experiments. Key to the accuracy of these experiments
will be how these measurements are reported. The next sections outline the measurement
methodology applied during offline quantification.
7.2.1 Measurement Methodology
One of the most critical aspect to quantifying the affect that different control strategies
have on a real process, will be in determining the significance of our results relative to
process noise and changing steady-state conditions. The following outlines the statistical
measurement methodology followed in this study.

145
—±base
Each measurement, e.g. cv
, taken of the units steady-state condition for a variable
at time ta, e.g. cv(ta) , will implement the following:
I: Collect T observations, {cv(t)}t°= to
II: Calculate the sample mean
tn+T
\ Z ^(0
III: Calculate the sample standard deviation
(93)
L+T
,base
- £ (cv(0-A )
(94)
t = tn
Each delta calculated, e.g. A^p, of changes to the units steady-state condition will
implement the following:
—±base —*■val
I: Given measurements for cv and cv
II: Calculate the lower bound of a (1 - a) 100% confidence interval for
the difference between two sample means according to
(f£ase-av_?')-— hb"e + ^' [66],
vrcv 'ey T V cv cv L J
(95)
7.2.2 Results
Given the results obtained in Section 7.1 "Offline Quantification," there seems little
hope of conducting successful online experiments. While online tests might be useful in
confirming the offline results, there are considerable costs associated with conducting
them. In addition to the time and resource requirements, there is the invaluable capital of
buy-in from operations and engineering to be considered. Most of these operators and

146
engineers have spent 5 to 20 years learning how to drive and maintain the unit. They are
the experts. Bringing in a new technology that is going to show them how to better operate
their unit, needs to be managed carefully. The operators and engineers have seen many
technologies come and go and have never seen a technology able to model unit emissions
much less control them.
With all of this said, 10 experiments were conducted with the steady-state optimizer.
This was the first controller implemented, and these experiments were conducted before
all of the above results we analyzed. The results were about as encouraging as they were in
our offline analysis, and are presented in Figures 33 and 34.
0.015
0.010
0.005
0.000
-0.005
-0.010
-0.015
-0.020
-0.025
-0.030
1
2
3
4
5
6
7
8
9
10
—♦— Experiment
0.005
-0.006
0.002
-0.022
-0.023
-0.008
0.010
0.003
-0.027
-0.003
—•— Validation
-0.006
-0.029
-0.016
-0.022
-0.019
-0.027
-0.003
-0.032
-0.015
-0.030
- A - Average
-0.001
-0.018
-0.007
-0.022
-0.021
-0.018
0.003
-0.014
-0.021
-0.017
Experiment
Figure 33: Change in NOx for steady-state controller experiments.
Figures 33 shows the measured NOx change between the baseline and experiment,
along with the corresponding change between the experiment and validation regions. We
can see that there seems to be a slight decrease in NOx. The average NOx percent change

147
for all experiments was 4.8%. The changes were so small with respect to the steady-state
NOx variance, however, that few of the experiments proved significant.
Figures 34 shows the corresponding CO changes. Here, its hard to see any trend. The
average CO percent change for all experiments was 2.7%. Once again, these changes are
so small that few of the experiments proved significant.
Figure 34: Final CO level for steady-state controller experiments.

CHAPTER 8
PARAMETERIZATION PROBLEM
The work thus far has resulted in some unexpected results. Applying accepted neural
network modeling techniques has produced several models for a process that contain dras¬
tically different “knowledge” about the process. Here, knowledge refers to the cause-and-
effect relationships between MVs and CVs, which is what each of the control designs
depend upon. Contrast this with our modeling results from Section 6.8 "Analysis," where
these same models demonstrated consistent and robust “knowledge” of the process, where
knowledge was considered to be the models ability to predict or forecast a blind dataset.
Also notice, that this same blind test dataset was used to perform the offline quantification
of the controllers.
This problem is analogous to the parameterization problem in classical adaptive con¬
trol theory. An adaptive control system centers around the idea that a process is described
as a mathematical function with parameters. It might therefore be expected that the way
parameters are estimated is essential to the success of an adaptive controller. It is useful to
view parameter estimation in the broader context of system identification. The key ele¬
ments of system identification are the selection of model architecture, experiment design,
parameter estimation, and validation. Since system identification is executed automati¬
cally in adaptive systems, it is essential to have a good understanding of all aspects of the
problem. The elements of system identification are known to be fundamental issues in
adaptive control theory.
148

149
Even though this is a problem from classical controls, there are no classical solutions
which apply to neurocontrol designs. The primary reason for this is the use of neural net¬
works as process models. Neural networks are fundamentally different than linear or first-
principles-based process models in that they are nonparametric. The individual parameters
have no physical interpretation, i.e., they are meaningless coefficients in a black box. We
are therefore seeking a method to validate the parameterization of a nonparametric system.
In classical adaptive control theory, parameterization is a design-time issue that is typi¬
cally dealt with analytically. We can expect that the solution for neurocontrol theory will,
like most other aspects of neural-network-based system design, have to be dealt with
empirically.
8.1 Search for a Validation Metric
Let’s begin our search for a methodology to validate the parameterization of neural
network models by reviewing the metrics used to assert the performance of these models.
Figure 35 presents three of these metrics for the steady-state CV MLP model. The primary
metric of model quality used in this study has been correlation. The correlation between
the actual unit NOx emission and the MLP’s prediction of NOx was 0.801. This metric
indicates that the MLP model understands a considerable amount about the variation in
NOx production. Furthermore, notice that the MLP produced this prediction by observing
the MVs, DVs and SVs only, implying that the MLP understands a considerable amount
about the relationship between these variables and NOx formation.
The normalized mean-squared error (NMSE) was also used to asses the worthiness of
each model. Recall that a NMSE greater or equal to 1 indicates that the model is perform¬
ing no better than a trivial prediction, which simply predicts the statistical mean. Here the

150
MLP predicted the blind test dataset with a NMSE of 0.298, indicating that the model is
doing considerably better than simply predicting the mean.
ssCVModel MLP
Figure 35: Summary of validation metrics for MLP CV model.
The correlation and NMSE results provided in Figure 35 are for the model with the
lowest cross-validation MSE from a set of 10 models trained from different random initial
conditions. Also recall that the offline quantification of the controllers was performed
using the models with the second lowest cross-validation MSE. Furthermore, remember
that we determined that the controllers preformed very well when their reference models
were used to quantify their performance. We are therefore looking for a validation metric
which would provide some insight into the difference between these two models. The cor¬
relation and NMSE results presented in Figure 35 cannot differentiate between these two
models, because they relate to only one of these models.
It is clear that the validation metric that we are seeking should provide an indication
about how a particular model compares with the ensemble of possible models which could

151
be developed for the process. For nonlinear models trained with a gradient descent learn¬
ing algorithm, this must include some understanding about the structure of the perfor¬
mance surface. Figure 36 presents the correlation and NMSE over the blind test dataset for
each of the 10 training runs of the steady-state CV MLP model. Recall that each training
run starts from random initial conditions. With the exception of two training runs, which
appear to have been trapped in local minima, there is relatively consistent performance
across the 10 sample models from the ensemble of possible models. Furthermore, there is
virtually no difference between the best model and the second best model. Clearly, the
correlation and NMSE metrics are not going to provide insight into why two parameter-
izations for the same controller provide significantly different results.
0.900
0.800
0.700
0.000
|L
0.769 | 0.763 0.791 | 0.447 0.727
0.711 0.801 I 0.483 0.780
â–¡ NMSE | 0.331 | 0.407 0.304 0.544 j 0.420 ; 0.371 0.298 | 0.575 : 0.313
0.760
0.340
Figure 36: NMSE and R for all 10 training results for MLP CV model.
In addition to correlation and NMSE, Figure 35 presented a metric labeled “std err/
var.” The realization that we are looking for a metric capable of differentiating between a
single model and the ensemble of possible models, leads naturally to considering standard

152
errors. Recall that Tibshirani [61] proposed the “bootstrap” method for estimating the
standard errors of a MLP’s predictions. Having estimates for the standard errors will pro¬
vide insight into variability in the predictions between models drawn from the ensemble.
The “std err/var” column in Figure 35 presents the average standard error calculated using
the 10 model trainings and the “bootstrap” method across the test dataset, divided by the
measured NOx variance across this dataset. The reason for dividing by the variance is to
provide an intuitive feel for the reported value of this metric. A standard error equal to this
variance indicates almost no confidence in the model predictions, while a standard error
significantly smaller than the variance indicates confidence that the models understand
more that the natural variation of the variable. The “std err/var” metric value of 0.358
reported in Figure 35 fails to differentiate our MLP models.
The problem with all of these metrics is that they assess the models ability to predict
future outputs of the process. The models are able to make accurate predictions, because
although the test dataset is blind it must contain similar process relationships, cause-and-
effect, to the training dataset. When our control algorithm is run across this dataset, it is
changing these relationships. There are three likely mechanisms by which this could hap¬
pen:
1) The MVs and DVs are not independent. When the controller moves a
MV it assumes that the other MVs and DVs remain constant.
2) The SV models are not accurately modeling the impact of MV moves.
3) The MVs are highly correlated. Thus changing a MV in the test dataset
breaks our assumption that the correlation relationships within the train¬
ing and test datasets are the same.
It is unlikely that case 1) is the cause of our troubles. For if the MVs and DVs are not
independent, then there is no way for any of the models to have determined this. Given

153
that all of the models have been developed from the same MV and DV variable sets. If in
fact these variables are not independent, then the impact of this on our control designs
would be the same as if these variables are correlated which will be considered in case 3).
If there is a problem with the SV models, this would definitely have an impact on the
results of our controllers. When examining the validation metric for the steady-state CV
model above, we were providing it with the actual SVs as input rather than the predicted
SVs from the SV model. Recall that the later configuration is how the controller utilizes
these models. We have seen, however, that the correlation and NMSE metrics for the SV
models demonstrate that these models are able to predict the SVs over the blind test
dataset. To verify that this is not the cause of our problems, Figure 37 presents the results
for the combined SV/CV configuration for the same metrics considered in Figure 35. Here
although using the SV/CV combined model did reduce the fidelity of our ability to predict
the CVs, the degradation in performance does not justify the poor performance of our con¬
troller.
ssSVModel / ssCVModel
Figure 37: Summary of validation metrics for combined SV/CV model.

154
Case 3) is a valid cause for our parameterization problem. When modeling an indus¬
trial process, correlated variables are going to be a fact of life. Unfortunately, the available
tools in the literature for dealing with such situations are very limited. The following sec¬
tions will demonstrate that correlation has a tremendous impact on the design of model-
based controllers.
8.2 Correlation Paradox
Correlated variables are a natural phenomenon, as two variables are correlated when
they are related through some physical laws or process. In fact, the mission statement for
the empirical investigator is to infer, or learn, these physical laws from observations of
process data. Hence, correlation is a double-edged sword. Without it learning is not possi¬
ble as there is nothing to infer, while unanticipated correlation makes it difficult, if not
impossible, to understand what has been learned or inferred.
Industrial process plants, through centralized control, systematically correlate their
process variables. The plant’s distributed control system (DCS) maintains a large number
(thousands to tens of thousands) of feed-forward and feedback control loops, from a rela¬
tively small number (tens to hundreds) of operator setpoints. The DCS is designed to batch
control over as many subsystems as possible, from the smallest possible number of opera¬
tor setpoints. If it were feasible the operator would only have a single setpoint called
demand. As a rule, most variables within an industrial plant will be highly correlated to
plant demand.
The effects of correlation have been well document for modeling applications like sys¬
tem identification [56] and regression [53], The most significant attribute that these appli¬
cations have in common is that they deal with systems that are either linear, or have a

155
relatively simple parametric nonlinear form. This feature allows an investigator to apply
analytically assess the significance of model parameters. In fact, the correlation effects are
automatically accounted for in significance testing. Thus the validation required to assess
the parameterization of a model is accomplished through significance testing.
To illustrate the effect of correlation, consider the physical processing plant that pro¬
duces output y from inputs u x and u2. An investigator building a model for this plant is
presented with data for y, u ¡ and u2, without any information about the physical system.
The investigator hypothesizes that the data was generated according the regression model
y = Pj«i + P2w2 + rl’ (96)
where P, and P2 are unknown coefficients, and r| is a zero-mean uncorrelated distur¬
bance term. Applying the method of least squares (LS), Ramanathan [53] shows that the
corresponding normal equations are given by
PlZ"l + PlZWlW2 = 2>!
(97)
Pll>l“2 + Pll>2 = YayUl
(98)
with solutions
= Yyu\Yu2-'Zyu2Yu\u2
XmÍZm2-(X"im2)2
p = Yyu2TuZ\-Yyu\Yu\u2
5>iXM2~(I>1W2)2
(99)
(100)

156
and variances
Var(P,) = f- (101)
2"l(* ~r)
2
Var( p2) = f- (102)
Xm2(! ~r)
2
Cov(\3„ p2) = -- 7"a '' , (103)
vZMiZw2(1~r)
2
where a“ = Far(r|) and r is the correlation coefficient between ux and u2.
Suppose ux and u2 are highly correlated, r is near ±1. It is evident from equations
(101) and (102) that the variances, and hence the standard errors of p, and p? will be
2
very large when r is close to 1. A large variance means poor precision and a low student
t-statistic, which results in insignificance. In addition, we can see from equation (103) that
the covariance between the regression coefficients will be very large, in absolute value. If
the estimates are correlated, each coefficient is capturing part of the effect of the other
variable and hence it is difficult to obtain the separate effects of w, and u2 on y. In other
words, we cannot hold u2 constant and increase u, alone, because u2 being correlated
with u j, will also change as a result. Or vise-versa, since correlation is not cause.
Ramanathan [53] offers the following properties of models derived from correlated
input variables:
1) If two or more explanatory variables in the multiple input (MI) model
are exactly linearly related, then the model cannot be estimated.

157
2) If some explanatory variables are nearly linearly related, then OLS and
MLE and hence are unbiased, efficient, and consistent.
3) The effect of correlation among input variables is to increase the stan¬
dard errors of the parameters and reduce the t-statistics, thus making
these parameters less significant (and possibly even insignificant). The
tests of hypotheses are, however, valid.
4) The covariance between the parameters of a pair of highly correlated
variables will be very high, in absolute value, thus making it difficult to
interpret individual coefficients.
5) Correlation may not affect the forecasting performance of a model and
may possibly even improve it.
8.2.1 Effects of Correlation on Neural Network Modeling
The neural network literature is not without reference to the issues of correlated inputs,
but the number of references are disproportionately low. These issues have not received
the attention within the neural network community that they have in related modeling
communities like system identification. The easy answer to why is that neural networks
are highly nonlinear structures making investigations into relevant statistics difficult if not
impossible. A more thorough understanding, however, lies in the way in which neural net¬
works have been applied to date. The author offers the following observations:
1) Descent based learning will always arrive at a solutions, regardless of
the degree of correlation within the input space. The solution, however,
is rarely unique nor globally optimal.
2) The vast majority of applications for neural networks, rely purely on
their ability to forecast. For this reason, standard errors for a networks
predictions has been a recent topic of research.
3) If the correlation relationships in the input space of all testing datasets
are identical to those in the training dataset, then the correlated inputs
will not degrade the network’s predictions. This situation, which will
often be the case for prediction applications, is unlikely in control appli¬
cations since the controller will independently move these inputs.

158
4) The network coefficients, its weights, have no parametric interpretation.
Their has been little reason, even if we knew how to, to calculate their
standard errors and significance.
Because of the way in which neural networks have been applied, the vast majority of
work done on the issue of correlated inputs deals almost entirely with its effect on the
dynamics of learning. Correlation in the input space drastically reduces the rate of conver¬
gence for descent-based learning algorithms [22], To counter this effect, the most com¬
mon neural network solution to correlation is to projection the input spaces to a lower
dimensional subspace with orthogonal bases. Most of these projection operators are linear
and based on the energy or information content of the variables. The most common such
projection is principle component analysis. The result of using such a projection is to pre¬
cede the neural network with a simple matrix multiply. Thus the actual inputs to the neural
network are completely, or nearly completely uncorrelated features and thus accelerate the
convergence of the learning algorithm.
Although very useful, these techniques do not provide a solution to the parameteriza¬
tion problem. These techniques can be used to improve the consistency of the neural net¬
work portion of the model, but do not improve the consistency of the combined models
containing the transform stage with the neural network. This point may be confusing now,
but the following sections should help to clarify it.
In summary, neural networks have found an applications niche where this robust pre¬
dictor has demonstrated the ability to out forecast more traditional methods. One of the
problems with moving neural networks from an academic interest to an accepted model¬
ing methodology is the lack of standardized reporting. The vast majority of neural network
applications to date apply the inferences of a model without knowledge of their statistical

159
significance. Recent work in standard errors for these predictions have taken an important
step towards solving part of this problem [21].
8.2.2 Effects of Multicollinearity on Model-Based Control
Each of the controllers presented in Chapter 4 belong to the general family of model-
based control, i.e., they each explicitly use a reference model during offline training or
online control. The process knowledge provided by the model in each of these control
designs is
dev
dmv
(104)
where the vectors cv and mv are the CVs and MVs, respectively. Recall our example in
Section 8.2 "Correlation Paradox," where the model was given by
y = Mi + P2M2 + rl-
(105)
If this model was used as a reference model for a model-based control scheme,
cv = {y} and mv = {ux,u2} could be considered to be the CVs and MVs, respec¬
tively. Hence the process knowledge provided by our model is the set {p ¡, P2}. But as we
have seen when the MVs and u2 are highly correlated, the standard errors for P, and
P? will be very large. A situation that will quickly render our process knowledge insignif¬
icant. Clearly, this will have a tremendous impact on the performance and robustness of
our controller.

160
8.3 Validation Metric
Neural networks do not have the same convenient interpretation for their coefficients.
They do, however, infer relationships between the MVs and CVs, which is used directly
by the controllers. The method of extracting this relationship from a neural network
model, is commonly referred to as sensitivity analysis. The MV sensitivities dcv/dmv can
be calculated directly using backpropagation. These sensitivities will be a function of the
unit’s operating state {mv, dv}. Recall that we are trying to get a feel for how these sensi¬
tivities vary across the ensemble of possible models. Figure 38 illustrates the MV sensitiv¬
ities for a typical operating state across the 10 steady-state MLP models, representing a
sampling from this ensemble.
Figure 38: Sensitivity results for all 10 training results for NOx CV model.
For the first time, we can clearly see our parameterization problem! Each model is pro¬
viding significantly different process knowledge to the controller. More explicitly, our
best and second-best models with respect to NMSE, R and SE, represented by training

161
runs number 3 and 7, look as if they are modeling two completely different processes.
Alas there’s hope.
Figure 38 provides the average sensitivity reported from the ensemble of models,
along with the corresponding 95% confidence intervals. This result demonstrates that we
are not able to assert the directional sensitivity with any degree of confidence. It is easy to
see why the controllers had so much trouble.
Figure 39: NOx CV model sensitivity with 95% confidence intervals.
8.4 Revised Representation Pruning Algorithm
Recall that the goal of the representation pruning algorithm is to identify the best vari¬
able representations within each group. In Chapter 6, the “best” variable representation
was determined by selecting the representation which produced a model with the highest
correlation on a blind test dataset. In hindsight, it is clear that the definition of "best" must
be application-specific. If the models were being applied to a forecasting application, then
the representation pruning algorithm would have suited our purposes. For control applica-

162
tions, however, it is clear that the definition of "best" will have to be modified. Note that
the robustness of the predictors produced by this algorithm should still be in doubt. If the
forecasts are always made for a dataset with identical correlations characteristics to the
training dataset, then the model’s forecast would be fine. If, however, the correlation is
less physical, then there is still cause for concern.
8.4.1 Input Sensitivity Standard Errors
Recall the “bootstrap” approach to calculating the prediction standard errors for a neu¬
ral network presented in Section 2.2.4.0.1 "Bootstrap methods." The validation metric to
differentiate models to be used for our model-based control designs will similarly need to
evaluate the standard error of the MV sensitivities. The following algorithm has been
developed to approximate the MV standard errors using a “bootstrap” methodology:
I: Generate N8 datasets, each one of size N° drawn with replacement
II: For each bootstrap dataset b e [ 1, iv ], find
ArgMin$h{J(db-p(ub, Tv¿))}
(106)
III: Estimate the standard error of the i th input sensitivity as
(107)
where
(108)

163
Conceptually, this algorithm simply calculates the variance of the input sensitivities
across multiple models, all trained using different random initial conditions and indepen¬
dently sampled “bootstrap” datasets. Clearly, the variance of the sensitivity calculations
presented in Figure 38 will be high. If an algorithm can be found that is capable of reduc¬
ing these standard errors, the offline controller quantification results should improve.
8.4.2 Algorithm
The revised representation pruning algorithm uses the standard error estimates for the
model’s input sensitivity across a dataset, as presented in Section 8.4.1 "Input Sensitivity
Standard Errors," to determine “best” as follows:
I: Perform the group selection pruning algorithm to the Essential Tag
List in the appendix
II: Select the first MV group in the reduced Essential Tag List
III: If there is only one unique “representation” in the group then select the
next MV group and goto step III
IV: For each unique representation in the selected group, train a
MLP(15,5) 30 times with the variables for this representation as the
only input variables from the selected group; the 30 models resulting
from each training will be called “representation” model set
V: Calculate the standard error for the input sensitivities for each MV,
VI: Calculate the average normalized input sensitivity standard error for
each “representation” model set according to
(109)

164
VII: Select the best “representation” model set as the set with the lowest
input sensitivity normalized average error, and remove all variables
associated with every “representation” model set except for the best
“representation” model set
VIITSelect the next group and goto step III
Figure 40: Results of revised representation pruning algorithm.
The results after running the revised pruning algorithm for each of the 4 groups are
illustrated in Figure 40. Notice that unlike the initial pruning algorithm, this algorithm sig¬
nificantly degraded the fidelity of the model as it pruned each representation group. The
variable selections after the revised representation pruning algorithms are provided in
Table 14.

165
Table 14: Final variable selections after revised pruning.
Manipulated Variables
Disturbance Variables
State Variables
Control Variables
1/3 OFA Damper Bias
Ambient Air Press
Sec AirTemp Side A
CEM NOx
2/3 OFA Damper Bias
Ambient Air Ternp
Sec Air Temp Side B
FD Fan Bias
Bnr Atm Stm Press
CEM CO
" 1
02 Trim
Bnr Atm Stm Tern p
Generated M W
GR Fan 2A Inlet Dmpr Bias
CEM Barometric Pressure
W indbox Pressure
GR Fan 2B Inlet Dmpr Bias
Cond Back Pres - Side A
GR Fan HpprDmprA Bias
Cond Back Pres - Side B
GR Fan HpprDmprB Bias
Fuel Gas Flow Indication
Fuel Oil Flow Indication
Fuel Temp Fired
Furnace Pressure
8.4.3 Variable Representation
The results of the revised representation pruning algorithm are very interesting, and
provide some lessons about variable representation choices. For each MV group, the
revised algorithm chose the operator bias, rather than the positions selected by the initial
algorithm. Recall the example given in Section 6.2.1.5 "Variable representation," where 5
tags were identified that represented the single process variable of gross airflow. As we
have seen, this many-to-one mapping between tags and variables is very common in
industrial plants. The tags selected for variable representation should have the following
characteristics:
• Uncorrelated: The tag or tags chosen to represent the variable should be uncorre¬
lated with each other and with tags chosen to represent the other input variables.
• Representative: The tags chosen should provide a complete representation of the
dynamics of the variable, i.e., they should be as correlated as possible to the pro¬
cess variable.
• Dynamic: The tags chosen should contain as much of the real variation in the
process variable as possible. If the variable is changing for any reason, even if
the variation if not intended or not desirable, this information should be repre-

166
sented in the chosen tags.
From the discussion above, it is easy to see that all 5 tags, except FD fan trim, will be
highly correlated to plant demand. As noted previously, most tags will be correlated with
plant demand. This would make these four tags a poor choice for representing FD fan air¬
flow. FD fan trim, on the other hand, is not correlated to demand. The trim is a bias tag
that only moves when the operator wishes to alter the fuel-air mixture, represented by FD
fan demand, that was designed into the DCS. The bias is also appealing from an optimiza¬
tion perspective, since it provides a way to tune control over the process without having to
control the process.
Turning our attention to the second and third criteria for variable representation above.
The FD fan trim is certainly correlated to gross airflow, since moving it will alter the FD
fan setpoint. Dynamic, however, the FD fan trim is not. The only time the trim moves is
when the operator touches it. For the unit considered in this study, the trim was rarely
touched1. In addition, the trim does not contain any of the natural or unintended variability
in airflow, e.g. the slack in the PID controller. This variability is real, in the sense that air¬
flow actually changed, and has an impact on the combustion process. Although not
intended, this variation provides rich data for learning.
So we can either choose a representation rich with dynamic information about the pro¬
cess but highly correlated to demand, or a representation completely uncorrelated to
demand that is rarely moved. Pruning the variable representations based on the fidelity of
the resulting model will select tags rich in dynamic information regardless of correlation:
1. This is an interesting fact, since FD fan trim turned out to be one of the signifi¬
cant levers over optimizing NOx.

167
pruning based on standard errors of the knowledge extracted about the cause-and-effect
relationships will prioritize independent tags. The answer to this dilemma identified in this
work requires two solutions: 1) with sufficient parametric testing, structured movement of
the MVs, the models will be able to extract enough knowledge about the process for the
controller to function properly; and 2) when the controllers are connected to the plant, they
will move the MVs and model retuning will continue to extract more knowledge about the
process.
Figure 41 illustrates the MV sensitivities for the same operating state illustrated in Fig¬
ure 38, across the first 10 of the 30 revised steady-state MLP models.
1.200
1.000
0.800
0.600
0.400
0.200
0.000
-0.200
-0.400
-0.600
-0.800
Figure 41: Sensitivity results for all 10 training results for revised NOx CV model.
There is a visible difference in the significance, i.e., our confidence, in these models as
compared to original models. These models have extracted consistent process knowledge
with respect to the relationships between the MVs and CVs. A comparison between the

168
standard error calculations between these models and the initial models is presented in
Figure 42. This result clearly validates our observations about Figure 41.
1.000
0.800
0.600
0.400
0.200
0.000
-0.200
-0.400
-0.600
1/3 OFA
Damper
Bias
2/3 OFA
Damper
Bias
FD Fan
Bias
02 Trim
GR Fan
2A Inlet
Dmpr Bias
GR Fan
2B Inlet
Dmpr Bias
GR Fan
Hppr Dmpr
A Bias
GR Fan
Hppr Dmpr
B Bias
â–¡ Before
0.229
-0.398
0.084
0.824
-0.030
-0.276
0.123
0.126
â–  After
0.053
-0.025
0.031
-0.071
-0.014
-0.018
-0.040
0.092
Figure 42: Revised NOx CV model sensitivity with 95% confidence intervals
8.5 Modeling
Figure 43 presents the results of training each of the best model architectures from
Chapter 6 with the new variable definitions presented in Table 14. Here, each model archi¬
tecture was trained 10 times, the training result with the lowest cross-validation error was
selected, and the selected result was tested against the blind test set. Notice that this pro¬
cess is equivalent to the training process for the previous models.
Comparing Figure 43 with Figure 28, the new models have lost some fidelity with
respect to their ability to predict a blind test dataset. The degradation in fidelity, however,
is quite small when compared with the increase in confidence with respect to the MV sen¬
sitivities that the new models have (Figure 42).

169
0.000
MLP
TDNN
GM
NLSS
â–¡ CV Model
0.753
0.818
0.880
0.851
â–  SV Model
0.709
0.733
0.845
0.820
â–¡ ISV Model
0.687
â–¡ IMV Model
0.607
Figure 43: Best revised models for all model definitions by architecture
8.6 Control Implementation
The assumption is that models with higher confidence in the MV sensitivities will pro¬
vide better reference models for our model-based control designs. Each control design will
now be implemented with these new models and their performance re-quantified.
8.6.1 Offline Quantification
The offline performance for the new controller will follow the same methodology pre¬
sented in Section 7.1 "Offline Quantification." Once again, the dynamic process models
chosen for offline quantification will not be the reference models used to implement any
of the control designs. In particular, the offline quantification will use the second best
model, with respect to cross-validation MSE, as the dynamic process models.
Once again, offline quantification will apply each controller to the dynamic process
models across the test dataset and calculate the average NOx reduction along with the
amount of CO production above the maximum CO constraint. The NOx and CO results
are presented in Figures 44 and 44, respectively.

170
Avg NOx Reduction
Baseline
Steady-State
M PC
Steady-S tate
MIC
Dynamic MPC
Dynamic MRAC
â–¡ Before
0
-0.01339
-0.01556
-0.009
-0.01035
â–  After
0
-0.111
-0.054
-0.116
-0.107
Figure 44: Average NOx reduction over testing dataset using old and revised models.
Avg CO Above Max
140
120
100
80
60
40
20
0
â–¡ Before
â–  After
Figure 45: Average CO above max over testing dataset using old and revised models.
Recall that it was at this point in developing the controllers the first time that the prob¬
lem first surfaced. Clearly, the new controllers are performing much better than the origi¬
nal ones, whose results have been included in Figures 44 and 45 for reference.

171
8.6.2 Online Quantification
The offline quantification first identified a problem with the control designs, and now
indicates that it has been solved. The proof, however, is always in the pudding, or at least
in the response of the actual plant. It stands to reason, that if we were not able to develop
models with consistent knowledge of the process, then there is little hope that the resulting
controllers would be able to improve the process. Consistent models of the process, how¬
ever, do not necessarily imply that the resulting controllers will be able to improve the
process.
To quantify the ability of the controllers to improve the process, the methodology out¬
lined in Section 7.2 "Online Quantification" is applied. Once again, 10 experiments were
conducted with the steady-state optimizer. The results this time, however, are very encour¬
aging.
Figure 46 shows the measured NOx change between the baseline and experiment,
along with the corresponding change between the experiment and validation regions.
There is a significant decrease in NOx. The average NOx percent change for all experi¬
ments is -24.83%.
Figure 47 shows the corresponding CO changes. Here, the controller was able to have
a significant impact on CO. The average CO for both baseline and experiment is approxi¬
mately 500ppm, but the variance of the baseline data is 121.576ppm while the experiment
standard deviation is 13.473.

172
Figure 46: Change in NOx for revised steady-state controller experiments.
800
700
600
500
400
300
200
100
0
1
2
3
4
5
6
7
8
9
10
* Baseline
570
559
301
583
656
508
566
423
653
349
♦ Experiment
482
511
494
482
509
513
507
483
497
513
—•—Validation
487
583
212
640
648
470
534
438
715
407
Experiment
Figure 47: Final CO level for revised steady-state controller experiments.
This offline quantification procedure was repeated for each of the four control designs,
and the NOx and CO results are summarized in Figures 48 and 49. Included in this figures
are the estimated results that each controller thought it would achieve across the same

173
dataset, this was accomplished by using a simulator for the plant as described in Section
7.1 "Offline Quantification."
Figure 48: Average percent NOx reduction for 10 online experiments.
There are some interesting observations about these results:
1) The performance of the MIC controller was substantially poorer than
the other control designs in both offline and online analysis.
2) There online performance of the steady-state optimizer was better than
either of the dynamic controllers.
3) The dynamic controllers performed significantly better in offline analy¬
sis, that they did for online analysis.
The poor performance of the MIC controller can be explained by its treatment of con¬
straints. Recall that the MV constraints were treated as penalty functions for the online
optimizer, and that the CO maximum constraint was addressed by placing CO in the CV
set and fixing it to 500ppm. This leaves only a single degree of freedom for the online
optimizer, which is the single CV of NOx. The problem is that when the online optimizer
starts moving NOx to meet its objective function, it will have to stop as soon as one of the

174
MV constraints exercise their penalty functions. There are no other degree of freedom for
the optimizer to explore, other than to simply stop moving NOx.
Notice that although penalty function are also employed in the steady-state optimizer,
this problem does not exist. Here, the online optimizer has N1"' degrees of freedom, the
number of MVs, so that when a MV constraint’s penalty function is exercised it can sim¬
ply stop moving that MV. Furthermore, when a SV or CV constraint’s penalty function is
exercised, the optimizer can explore other combinations of MVs to maintain this con¬
straint while still trying to lower NOx.
Figure 49: Average percent CO reduction above 500ppm for 10 online experiments.
The observation that the steady-state optimizer out-performed the dynamic controllers
is most likely a function of the online test procedure. Recall that the controllers were
deployed online in an advisory mode. Here, the controllers optimal MV setpoints are for¬
warded to an operator, who is then responsible for manipulating the actual plant setpoints.
One has to remember that the operators are responsible for the unit, which is worth several

175
hundred million dollars. Operators operate the unit based on years of experience and
dogma. Sorting out which is which is not an easy task, and invariably the operator will and
should make the final call about whether the requested setpoints are safe. The result was
that MV setpoints were not made on regular intervals, and what the optimizer recom¬
mended as simultaneous MV moves might have been implemented one at a time over a
period as long as 20 minutes. Furthermore, recall that the measurement methodology, pre¬
sented in Section 7.2.1 "Measurement Methodology," required operating the unit under
steady-state conditions. Clearly, any dynamic advantage that these controllers might have
had was probably lost.
The observation that "the dynamic controllers performed significantly better in offline
analysis, that they did for online analysis" can also be chalked up to the online test proce¬
dure. In fact, this result reinforces this conclusion. The fact that the offline analysis dem¬
onstrated a significant benefit to using dynamic controllers, provides incentive to continue
work toward a closed-loop configuration.

CHAPTER 9
CONCLUSION
This project developed four neurocontrollers for the complex industrial process of
NOx formation. All the these neurocontrollers demonstrated benefit in an applications
area where traditional control designs have proven ineffective [39]. The first conclusion of
this study is that neurocontrollers are able to deal with highly complex industrial process
applications where more traditional methods have not been successful.
The objectives of this study went beyond demonstrating that a neurocontroller could
be developed, however. This work demonstrated neurocontrol designs that
1) are straightforward to implement,
2) account for dependent internal process states, and
3) can deal with correlated process variables.
9.1 Contributions
9.1.1 Application-Based Neurocontrol Implementation Methodology
A neurocontrol implementation methodology was developed whereby a process engi¬
neer with reasonable knowledge about the process variables can develop advanced neuro¬
controllers. The process engineer was able to simply classify process variables as MVs,
DVs, SVs and CVs, and the methodology was able to automate the development of each
stage of the controller. A set of supporting algorithms was proposed, implemented and
validated.
176

177
9.1.2 State-Space Neurocontrol Designs
Several neurocontrol designs were implemented ranging from purely input/output to
full state-space. Consistent with the literature on neural network modeling [31], the full
state-space architecture was difficult to train and was thus not the best performer. The best
performer, however, was a partial state-space model. This work demonstrated that avail¬
able process knowledge can be used to create partial state-space models which can signif¬
icantly improve the overall controller performance.
9.1.3 Methods for Dealing with Correlation
The primary limitations of neurocontrol identified during this study are directly related
to problems that arise when modeling with correlated input variables. In fact, the over¬
whelming conclusion of this work is that correlation in the input space is the single most
important factor when designing a inductive model-based control system. Decisions over
static vs. dynamic modeling, single-stage vs. multi-stage optimization, model topology,
controller type and every other seemingly monumental decision will prove irrelevant if
correlation issues are not properly addressed.
A new metric and methodology for variable selection were proposed, developed and
validated as a viable solution to correlation issues for industrial applications. This solu¬
tions is a data mining approach for applications where many representations are available
of the same underlying process variables. This approach will not work for applications
where multiple choices for a process variable are not available, and the author proposes
using this new metric as an objective for the learning rule as an extension to this work.

178
9.1.4 Accurate Combustion Models
Neural networks have found an applications niche, where this robust predictor has
demonstrated the ability to out forecast more traditional methods. One of the problems
with moving neural networks from an academic interest to an accepted modeling method¬
ology is the lack of standardized reporting. The vast majority of neural network applica¬
tions to date, apply the inferences of a model without knowledge of their statistical
significance. The lack of standardized reporting has not clouded the success of neural net¬
work because most of these applications rely solely on the model’s ability to forecast,
without regard for what relationships the model has inferred from the underlying process.
As clearly demonstrated in this work, the same cannot be said for neurocontrol, how¬
ever. The results presented here raise significant questions about the metrics being used to
evaluate model performance in the field of neural networks. Metrics which evaluate the
predictions made by a model do not imply that the model has learning the fundamental
cause-and-effect relationships within the process. A new metric was proposed and demon¬
strated to overcome this limitation.
9.1.5 Novel Combustion Controller
An online NOx optimizer was developed and evaluated. This optimizer demonstrated
a consistent 45% reduction in the overall NOx emissions from the power plant.
9.2 Afterword
The benefits demonstrated in this study at the Canal Electric generating station have
been maintained for more than two years. Since these initial results, neurocontrollers have

179
been deployed on 12 generating units throughout the United States. In addition to being
installed on more units, the neurocontrollers have been extended with
1) closed-loop implementations,
2) a software program that automates the neurocontrol design methodolo¬
gies,
3) online automation of the variable and architecture pruning methodolo¬
gies using evolutionary computing techniques, and
9.3 Future Direction
The author offers three important directions for extending this work:
1) development of learning algorithms that explicitly account for MV sen¬
sitivity standard errors,
2) algorithms to support automatic determination of variable type, and
3) online adaptation of the model-reference adaptive controller using reen¬
forcement learning strategies.

APPENDIX
Table 15: Essensial tag list.
Descriptor
Type
Unit 2 1/3 OFA Master Bias
MV
Unit 2 1/3 OFA Master Out
MV
Unit 2 2/3 OFA Master Bias
MV
Unit 2 2/3 OFA Master Out
MV
APFI 2A Air In Temp
MV
APH 2B Air In Temp
MV
Fuel Temp Fired
MV
Boiler Efficiency (abbr. Losses)
CV
Bnr Atm Stm Press
MV
Exit 02 B-Side
SV
Exit 02 A-Side
SV
Flue Gas CO B-Side
SV
Flue Gas CO A-Side
SV
Generated MW
SV
PAS3A-1 Pos
MV
PAS2B-1 Pos
MV
PAS2A-1 Pos
MV
PAS 1B-1 Pos
MV
PAS 1A-1 Pos
MV
PAS6B-1 Pos
MV
PAS 6A-1 Pos
MV
PAS5B-1 Pos
MV
PAS 5A-1 Pos
MV
PAS4B-1 Pos
MV
PAS 4A-1 Pos
MV
PAS3B-1 Pos
MV
PAS8B-1 Pos
MV
PAS 8A-1 Pos
MV
PAS7B-1 Pos
MV
PAS 7A-1 Pos
MV
U2 PAS Mast Gas Bias
MV
U2CEM CO
SV
Fuel Oil Flow
MV
FUEL GAS PRES INDICATION
SV
FUEL GAS FLOW INDICATION
SV
IDF2B INL VANE
MV
IDF 2AINL VANE
MV
Descriptor
Type
PAS 3A-1 Bias
MV
PAS2B-1 Bias
MV
PAS 2A-1 Bias
MV
PAS 1 B-1 Bias
MV
PAS 1 A-1 Bias
MV
PAS 6B-1 Bias
MV
PAS 6A-1 Bias
MV
PAS5B-1 Bias
MV
PAS 5A-1 Bias
MV
PAS4B-1 Bias
MV
PAS 4A-1 Bias
MV
PAS 3B-1 Bias
MV
PAS 8B-1 Bias
MV
PAS 8A-1 Bias
MV
PAS7B-1 Bias
MV
PAS 7A-1 Bias
MV
Heat Rate
CV
GR Fan Hppr Dmpr B Pos
MV
GR Fan 2B Inlet Dmpr Pos
MV
GR Fan 2A Inlet Dmpr Pos
MV
Relative Efficiency
CV
Windbox Pressure
SV
CEM Barometric Pressure
SV
Unit 2 FD Fan Bias
MV
Unit 2 Heat Rate
CV
CEM Nox
CV
Boiler 02 Trim Unit 2
MV
CEM Opacity Unit 2
CV
CEM Stack Temp
SV
AIR PREHTR 2A GAS OUT TEMP
SV
AIR PREHTR 2B GAS OUT TEMP
SV
Unit 2 SAS Master
SV
Unit 2 SAS Master Bias
SV
Unit 2 PA Shroud Mast 0
MV
Unit 2 PAS Master Bias
MV
U2 PA Shroud Mast Gas
MV
180

REFERENCES
[1] Anderson, C. W., Learning to control an inverted pendulum using neural networks,
IEEE Control Systems Magazine, pages 31-37, 1989.
[2] Astrom, K. J., & McAvoy, T. J., Intelligent control: an overview and evaluation,
Handbook of Intelligent Control, Van Nostrand Reihhold, New York, pages 3-34,
1992.
[3] Astrom, K. J., & Wittenmark, B., Adaptive Control. Addison-Wesley, Reading, MA,
1989.
[4] Archetti, F., A sampling technique for global optimization, Towards Global Optimi¬
zation. Amsterdam, 1975.
[5] Baker, W. L., & Farrell, J. A., An introduction to connectionist learning control sys¬
tems, Handbook of Intelligent Control, Van Nostrand Reihhold, New York, pages
35-64, 1992.
[6] Barnes, C., Applications of neural networks to process control and modeling, Inter¬
national Conference on Artificial Neural Networks ’91. Espoo, pages 321-326, 1991.
[7] Barto, A.G., Sutton, R.S., & Anderson, C., Neuron like elements that can solve diffi¬
cult learning control problems, IEEE Transactions on Systems, Man, & Cybernetics.
13:835-846, 1983.
[8] Barto, A.G., Sutton, R.S., & Watkins, C.J.C.H., Sequential decision problems and
neural networks, Advances in Neural Information Processing Systems 2. 1990.
[9] Baxt, W., & White, H., Bootstrapping confidence Intervals for Clinical Input Vari¬
able Effects in a Network Trained to Identify the Presence of Acute Myocardial Inf¬
arction, Tech. Rep.. 1994.
[10] Bello, M.G., Enhanced training algorithms, and integrated training/architecture
selection for multilayer perceptron networks, IEEE Transactions on Neural Net¬
works 3, pages 864-875, 1992.
[11] Bellman, R. E., Dynamic Programming. Princeton University Press, Princeton, NJ,
1957.
[12] Bertsekas, D.P., & Tsitsiklis, J.N., Neuro-Dynamic Programming. Athena Scientific,
Belmont, MA, 1996.
181

182
[13] Bishop, C., Neural Networks for Pattern Recognition. Clarendon Press, Oxford,
1995.
[14] Brent, R., Algorithms for Minimization without Derivatives. Prentice Hall, Engle-
wook Cliffs, NJ, 1973.
[15] Bryson, A.E., & Ho, Y., Applied Optimal Control. Hemishpere Publishing Corpora¬
tion, New York, 1975.
[16] Campolucci, P., Piazza, F., & Uncini, A., Online Algorithms for Neural Networks
with HR Synapses, IEEE Int. Conf. on Neural Networks ICNN-95. Perth, Australia,
1995.
[17] de Vries, B., & Principe, J.C., A Theory for Neural Networks with Time Delays,
Advances in Neural Information Processing Systems 3. Morgan Kaufman, pages
162-168, 1991.
[18] Feldkamp, L.A., & Puskorius, G.V., Training controllers for robustness: multistream
DEKF, Proceedings of IEEE International Conference on Neural Networks. Orlando,
FL, pages 2377-2382, 1994.
[19] Feldkamp, L.A., & Puskorius, G.V., Decoupled Kalman Trainings of neural and
fuzzy controllers for automotive systems, Proceedings of IEEE VTS Workshop on
Fuzzy and Neural System, and Vehicle Applications. Tokyo, 1991.
[20] Fukanaga, K., Statistical Pattern Recognition 2nd Edition. Harcourt Brace Jovanov-
ich, Cambridge, Massachusetts, 1990.
[21] Ge, R.P., A filled function method for finding global minimizer of a function of sev¬
eral variables, Math Program, pages 191-204, 1990.
[22] Halkjaer, S., & Winther, O., The effect of correlated input data on the dynamics of
learning, Advances in Neural Information Processing Systems 9. pages 169-175,
1997.
[23] Hardle, W., Applied Nonparametric Regression. Cambridge University Press, New
York, 1990.
[24] Haykin, S., Neural Networks: A Comprehensive Foundation. Macmillan College
Publishing Company, 1994.
[25] Haykin, S., Adaptive Filter Theory. 3rd edition, Prentice-Hall, pages 562-588, 1996.
[26] Hecht-Neilsen, R., Neurocomputing. Addison-Wesley Publishing Company, pages
168, 1990.
[27]Heskes, T., Practical confidence and prediction intervals, Advances in Neural Infor¬
mation Processing Systems 9. pages 176-182, 1997.

183
[28] Hardle, W., Applied Nonparametric Regression. Cambridge University Press, 1991.
[29] Hinton, G.E., Learning translation invariant recognition in massively parallel net¬
works, Proceedings PARLE Conference on Parallel Architectures and Languages
Europe, pages 1-13, 1987.
[30] Hrycej, T., Model-based training methods for neural controllers, Proceedings of the
International Conference of Artificial Neural Networks. Brighton, 1992.
[31] Hrycej, T., Neurocontrol. John Wiley & Sons, Inc., New York, 1997.
[32] Jang, J.-S.R., Sun, C.-T., & Mizutani, E., Neuro-Fuzzy and Soft Computing: A Com¬
putational Approach to Learning and Machine Intelligence. Prentice Hall, pages 523-
533, 1997.
[33] Kailath, T., Linear Systems. Prentice-Hall, Inc., Englewood Cliffs, NJ, 1980.
[34] Kalman, R.E., On the general theory of control systems, Proceedings of the 1 st Inter¬
national Congress on Automatic Control. Moscow, pages 481-492, 1960.
[35] Kalman, R.E., Design of self-optimizing control systems, Trans ASME, pages 468-
478, 1958.
[36] Koza, J.R., Genetic Programming. The MIT Press, Cambridge, MA, 1996.
[37] Levine, W.S., The Control Handbook. CRC Press, 1996.
[38] Levy, A.V., & Montalvo, A., The tunneling algorithm for the global minimization of
functions, SIAM J Sci Stat Comp, pages 15-19, 1985.
[39] Liptak, B.G., Optimization of Unit Operations. Chilton Book Company, Radnor,
Pennsylvania, 1987.
[40] Miller, W.T., & Werbos, P.J., Neural Networks for Control. The MIT Press, Cam¬
bridge, MA, 1992.
[41] Nguyen, D., & Widrow, B., The truck backer-upper: An example of self-learning in
neural network, Neural Networks for Control, pages 287-300, 1990.
[42] Nix, D., & Weigend, A., Estimating the mean and variance of the target probability
distribution, Proceedings of the IJCNN. pages 55-60, 1994.
[43] Omidvar, O., & Elliott, D.L., Neural Systems for Control. Academic Press, San
Diego, CA, 1997.
[44]Papoulis, A., Probability. Random Variables, and Stochastic Processes. 3rd ed.,
McGraw-Hill, New York, 1991.

184
[45] Pearlmutter, B., Gradient Calculations for Dynamic Recurrent Neural Networks: A
Survey. IEEE Transactions on Neural Networks. Vol. 6, No 3, pages 1212-1228,
1995.
[46] Perrone, M.P., General averaging results for convex optimization, Proceedings 1993
Connectionist Models Summer School, pages 364-371, 1994.
[47] Perrone, M.P., & Cooper, L.N., When networks disagree: ensemble methods for
hybrid neural networks, Artificial Neural Networks for Speech and Vision, pages
126-142, 1993.
[48] Poggio, T., & Girosi, F., Regularization algorithms for learning that are equivalent to
multilayer networks, Science 247. pages 978-982, 1990.
[49] Polak, E., Computational Methods in Optimization. Academic Press, New York,
1971.
[50] Pomerleau, D.A., Neural network-based vision processing for autonomous robot
guidance, Applications of Artificial Neural Networks II. volume 1469, pages 121-
128, SPIE, Bellingham, Washington, 1991.
[51] Puskorius, G.V., & Feldkamp, L.A., Decoupled extended Kalman filter training of
feedforward layered networks, Proceedings of the Joint Conference on Neural Net¬
works ’91. Seattle, pages 771-777, 1991.
[52] Puskorius, G.V., & Feldkamp, L.A., Neurocontrol of nonlinear dynamical systems
with Kalman filter trained recurrent networks, IEEE Transactions on Neural Net¬
works. Seattle, pages 279-297, 1994.
[53] Ramanathan, R., Introductory Econometrics. Harcourt Brace Jovanovich, Publishers,
1989.
[54] Rao, S.S., Engineering Optimization: Theory and Practice. John Wiley & Sons, Inc.,
New York, 1996.
[55] Rumelhart, D., Hinton, G., & Williams, R., Learning internal representations by
error backpropagation, Parallel Distributed Processing: Explorations in the Micro¬
structure of Cognition, pages 1: 322-328, MIT Press, MA, 1977.
[56] Scharf, L.L., Statistical Signal Processing: Detection. Estimation, and Time Series
Analysis. Addison-Wesley Publishing Company, New York, 1991.
[57] Singhal, S., & Wu, L., Training multilayer perceptrons with the extended Kalman
algorithm, Advances in Neural Information Processing Systems, Morgan Kaufmann,
San Mateo, CA, 133-140, 1989.

185
[58] Sofge, D.A., & White, D. A., Neural network based process optimization and con¬
trol, Proceedings of the IEEE Conference on Decision & Control, pages 3270-3276,
1990.
[59] Soulie, F.F., & Gallinari, P., Industrial Applications of Neural Networks. World Sci¬
entific, Singapore, 1998.
[60] Staib, W.E., The Intelligent Arc Furnace: Neural networks revolutionize steelmak¬
ing, Proceedings of the World Congress on Neural Networks, pages 1:466-469, 1993.
[61] Tibshirani, R., A comparison of some error estimates for neural network models,
Neural Computation. 8:152-163, 1996.
[62] Tikhanov, A.N., & Arsenin, V.Y., Solutions of Ill-Posed Problems, V.H. Winston,
Washington, DC, 1977.
[63] Troudet, T., Towards practical control designs using neural computation, Proceed¬
ings of the International Conference on Neural Networks, volume 2, pages 675-681,
1991.
[64] Tsypkin, Y.Z., Foundations of the Theory of Learning Systems. Nauka, Moskva,
1970.
[65] Vasiliev, O.V., Optimization Methods. World Federation Publishers Company,
Atlanta, GE, 1996.
[66] Walpole, R.E., Probability and Statistics for Engineers and Scientists. Macmillan
Publishing Company, 1985.
[67] Weigend, A.S., & Gershenfeld, N.A., Time Series Prediction: Forecasting the Future
and Understanding the Past Addison Wesley Publishing Company, 1994.
[68] Wendemuth, A., Opper, M., & Kinzel, W., The effect of correlations in neural net¬
works, J. Phvs. A. pages 26:3165, 1993.
[69] White, D.A., & Sofge, D.A., Handbook of Intelligent Control: Neural. Fuzzy, and
Adaptive Approaches. Van Nostrand Reinhold, New York, 1992.
[70] Williams, R.J., & Zisper, D., A Learning Algorithm for Continually Running Fully
Recurrent Neural Networks, Neural Computation. Vol. 1, pages 270-280, 1989.
[71] Williams, R.J., & Peng, J., An Efficient Gradient-Based Algorithm for On-Line
Training of Recurrent Network Trajectories, Neural Computation. Vol. 2, pages 490-
501, 1990.
[72]Werbos, P., Beyond regression: new tools for prediction and analysis in the behav¬
ioral sciences. Ph. D. Thesis, Harvard University, Cambridge, MA.

186
[73] Werbos, P.J., Advanced forecasting methods for global crisis warning and models of
intelligence, General Systems Yearbook, pages 22:25-38, 1977.

BIOGRAPHICAL SKETCH
Wesley Curtis Lefebvre was bom in Athens, Georgia, August 28, 1965. Curt grad¬
uated from Grandville High School in Grandville, MI, and continued his education at
Michigan State University. During his freshman year at Michigan State, Curt accepted a
cooperative education position with International Business Machines (IBM) in Owego,
NY. Curt continued his education, splitting each year evenly between Michigan State and
IBM. After several years Curt left Michigan State to finish his undergraduate education at
the State University of New York at Binghamton year round while maintaining full-time
employment with IBM.
After receiving a bachelor's degree in electrical engineering, Curt accepted a posi¬
tion at IBM’s research center in Yorktown Heights, NY where he was first exposed to the
field of neural networks. Curt decided to pursue graduate studies in neural networks, and
chose to attend the University of Florida in order to study under Dr. Jose Principe, one of
the first investigators of dynamic neural network theory. He graduated with a master of
engineering in electrical engineering, and his master's thesis was titled "An Object-Ori¬
ented Implementation of Artificial Neural Networks."
Curt began working on his doctoral degree while founding NeuroDimension, Inc.,
a software company based on the work from his master’s thesis. The case study for his
doctoral work led to the founding of a second company in the Boston area called NeuCo,
Inc. Curt is currently the President and CEO of NeuCo where he will continue to work in
the field of neural networks and neurocontrol after receiving his Ph.D. in May 2000.
187




PAGE 1

1(85$/ 1(7:25.%$6(' &21752/ '(6,*16 )25 &203/(; ,1'8675,$/ 352&(66 $33/,&$7,216 %\ :(6/(< &857,6 /()(%95( $ ',66(57$7,21 35(6(17(' 72 7+( *5$'8$7( 6&+22/ 2) 7+( 81,9(56,7< 2) )/25,'$ ,1 3$57,$/ )8/),//0(17 2) 7+( 5(48,5(0(176 )25 7+( '(*5(( 2) '2&725 2) 3+,/2623+< 81,9(56,7< 2) )/25,'$

PAGE 2

$&.12:/('*0(176 :KHQ ORRNLQJ IRU D JUDGXDWH VFKRRO ZDV WROG WR SLFN DQ DGYLVRU QRW D VFKRRO 0\ FKRLFH ZDV 'U -RVH 3ULQFLSH DQG KLV PHQWRULQJ DQG IULHQGVKLS WKURXJK D PDVWHUfV 3K' DQG WKH VWDUWXS RI WZR FRPSDQLHV KDV SURYHQ WKLV DGYLVH WR EH VRXQG ZRXOG DOVR OLNH WR WKDQN 'U *OHQ -RKQVRQ IURP ,%0
PAGE 3

7$%/( 2) &217(176 SDJH $&.12:/('*0(176 LL /,67 2) 7$%/(6 Y /,67 2) ),*85(6 YL &+$37(56 ,1752'8&7,21 3URSRVHG :RUN 'RFXPHQW 2UJDQL]DWLRQ /,7(5$785( 5(9,(: 2SWLPL]DWLRQ 1HXUDO 1HWZRUNV &RQWURO 7KHRU\ 12[ )RVVLO)LUHG 3RZHU *HQHUDWLRQ %2,/(5 237,0,=$7,21 )LUVW 3ULQFLSOHV )XHO DQG $LU 'LVWULEXWLRQ %RLOHU 7XQLQJ 7KH 5ROH RI &2 &21752/ '(6,*16 9DULDEOH 'HILQLWLRQV 2SWLPL]DWLRQ 2EMHFWLYHV 2SHUDWLQJ &RQVWUDLQWV 3HUIRUPDQFH &ULWHULD &RQWUROOHU 'HVLJQV LLL

PAGE 4

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

PAGE 5

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

PAGE 6

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b PLVVLQJ DFURVV )HEUXDU\ GDWDVHW )LJXUH ([DPSOH RI YDULDEOHV ZLWK ODUJH 106( EXW KLJK 5 )LJXUH 5HVXOWV RI W\SH SUXQLQJ DOJRULWKP )LJXUH 5HSUHVHQWDWLRQ SUXQLQJ DOJRULWKP UHVXOWV )LJXUH 5HVXOWV RI DXWRUHJUHVVLYH WDSV VHDUFK DOJRULWKP IRU 12[ $50$ 0RGHO )LJXUH 5HVXOWV RI PRYLQJ DYHUDJH WDS VHDUFK DOJRULWKP IRU 12[ $50$ 0RGHO )LJXUH 5HVXOWV RI KLGGHQ OD\HU 3( VHDUFK DOJRULWKP IRU WKH 12[ 0/3 0RGHO YL

PAGE 7

)LJXUH 5HVXOWV RI WKH KLGGHQ OD\HU 3( VHDUFK DOJRULWKP IRU 12[ 0/3 0RGHO )LJXUH 5HVXOWV RI WDS VHDUFK DOJRULWKP IRU 12[ 7'11 0RGHO )LJXUH 5HVXOWV RI WDSV VHDUFK DOJRULWKP IRU 12[ *11 PRGHO )LJXUH 5HVXOWV RI KLGGHQ WDSV VHDUFK DOJRULWKP IRU 12[ *11 PRGHO )LJXUH 5HVXOWV RI KLGGHQ VWDWHV VHDUFK DOJRULWKP IRU 12[ 1/66 PRGHO )LJXUH 5HVXOWV RI VWDWH KLGGHQ 3(V VHDUFK DOJRULWKP IRU 12[ 1/66 PRGHO )LJXUH 5HVXOWV RI RXWSXW KLGGHQ 3(V VHDUFK DOJRULWKP IRU 12[ 1/66 PRGHO )LJXUH %HVW PRGHOV IRU DOO PRGHO GHILQLWLRQV E\ DUFKLWHFWXUH )LJXUH $YHUDJH 12[ UHGXFWLRQ RYHU WHVWLQJ GDWDVHW )LJXUH $YHUDJH &2 DERYH PD[ RYHU WHVWLQJ GDWDVHW )LJXUH $YHUDJH 12[ UHGXFWLRQ RYHU WHVWLQJ GDWDVHW XVLQJ WUDLQ DQG WHVW PRGHOV )LJXUH $YHUDJH &2 DERYH PD[ RYHU WHVWLQJ GDWDVHW XVLQJ WUDLQ DQG WHVW PRGHOV )LJXUH &KDQJH LQ 12[ IRU VWHDG\VWDWH FRQWUROOHU H[SHULPHQWV )LJXUH )LQDO &2 OHYHO IRU VWHDG\VWDWH FRQWUROOHU H[SHULPHQWV )LJXUH 6XPPDU\ RI YDOLGDWLRQ PHWULFV IRU 0/3 &9 PRGHO )LJXUH 106( DQG 5 IRU DOO WUDLQLQJ UHVXOWV IRU 0/3 &9 PRGHO )LJXUH 6XPPDU\ RI YDOLGDWLRQ PHWULFV IRU FRPELQHG 69&9 PRGHO )LJXUH 6HQVLWLYLW\ UHVXOWV IRU DOO WUDLQLQJ UHVXOWV IRU 12[ &9 PRGHO )LJXUH 12[ &9 PRGHO VHQVLWLYLW\ ZLWK b FRQILGHQFH LQWHUYDOV )LJXUH 5HVXOWV RI UHYLVHG UHSUHVHQWDWLRQ SUXQLQJ DOJRULWKP )LJXUH 6HQVLWLYLW\ UHVXOWV IRU DOO WUDLQLQJ UHVXOWV IRU UHYLVHG 12[ &9 PRGHO YLL

PAGE 8

)LJXUH 5HYLVHG 12[ &9 PRGHO VHQVLWLYLW\ ZLWK b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

PAGE 9

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n ZRUNV KDYH EHHQ SUHGRPLQDQWO\ DSSOLHG WR IRUHFDVWLQJ RU PRGHOLQJ DSSOLFDWLRQV %DVHG RQ WKHLU VXFFHVV LQ VXFK DSSOLFDWLRQV WKHUH KDV EHHQ VLJQLILFDQW LQWHUHVW LQ XVLQJ QHXUDO QHWn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

PAGE 10

ZRUOG DSSOLFDWLRQV )LUVW WKH FRQWUROODEOH SDUDPHWHUV RYHU PRVW LQGXVWULDO SURFHVVHV DUH KLJKO\ FRUUHODWHG RIWHQ QRW IRU SK\VLFDO UHDVRQV EXW EHFDXVH RI RXU SURFHVV FRQWURO VWUDWHn JLHV 6HFRQG LQWHUPHGLDWH SURFHVV VWDWHV WKDW DIIHFW WKH SURFHVV RXWSXW ZKLFK DUH DOVR DIIHFWHG E\ WKH FRQWUROODEOH SDUDPHWHUV KDYH D VLJQLILFDQW LPSDFW RQ FRQWUROOHU SHUIRUn PDQFH :KHQ WKH FRQWUROOHU FKDQJHV WKH FRQWUROODEOH SDUDPHWHUV WKH LPSDFW WKDW WKLV KDV RQ WKH SURFHVV VWDWHV ZKLFK ZLOO LQ WXUQ DIIHFW WKH SURFHVV RXWSXW LV QRW DFFRXQWHG IRU LQ PRVW QHXURFRQWURO GHVLJQV LQ WKH OLWHUDWXUH 7KLV GLVVHUWDWLRQ DGYDQFHV WKH ILHOG RI QHXURFRQWURO E\ SURYLGLQJ WKH IROORZLQJ VROXWLRQV ILUVW WKH XVH RI VWDWLVWLFDO VLJQLILFDQFH WHVWLQJ RQ WKH ORFDO OLQHDUL]HG UHODWLRQn VKLSV H[WUDFWHG IURP QRQOLQHDU QHXUDO QHWZRUN PRGHOV WR DYRLG SUREOHPV ZLWK FRUUHODWHG FRQWUROODEOH SDUDPHWHUV VHFRQG DXJPHQWLQJ QHXURFRQWURO GHVLJQV WR LQFRUSRUDWH GHSHQn GHQW VWDWH PRGHOV 7KHVH HQKDQFHPHQWV KDYH EHHQ DSSOLHG WR IRXU GLVWLQFW QHXURFRQWURO DUFKLWHFWXUHV 7KH QHZ FRQWURO DUFKLWHFWXUHV KDYH EHHQ DSSOLHG WR WKH QRYHO DSSOLFDWLRQ RI FRQWUROOLQJ 12[ HPLVVLRQ IURP DQ RLO DQG JDVILUHG HOHFWULF SRZHU SODQW [

PAGE 11

&+$37(5 ,1752'8&7,21 1HXUDO QHWZRUNV KDYH IRXQG DQ DSSOLFDWLRQV QLFKH DV UREXVW SUHGLFWRUV WKDW KDYH GHPn RQVWUDWHG WKH DELOLW\ WR RXWIRUHFDVW PRUH WUDGLWLRQDO PHWKRGV LQ FRPSOH[ UHDOZRUOG DSSOLn FDWLRQV 7KH YDVW PDMRULW\ RI QHXUDO QHWZRUN DSSOLFDWLRQV WR GDWH UHO\ VROHO\ RQ WKH PRGHOfV DELOLW\ WR IRUHFDVW ZLWKRXW UHJDUG IRU ZKDW WKH PRGHO KDV OHDUQHG DERXW WKH UHODn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n URFRQWURO >@>@> @ >@ EXW WKH OLWHUDWXUH FRQWDLQV D GLVSURSRUWLRQDWHO\ ORZ QXPEHU RI SDSHUV SUHVHQWLQJ UHDOZRUOG QHXURFRQWURO DSSOLFDWLRQV 7KH YDVW PDMRULW\ RI WKHVH SDSHUV KDYH GHYHORSHG FRQWUROOHUV IRU IDEULFDWHG VLPXODWHG RU ODERUDWRU\ FRQWUROOHG SURFHVVHV 3URSRVHG :RUN 7KH JRDO RI WKLV SURMHFW LV WR GHYHORS UREXVW QHXURFRQWURO GHVLJQ PHWKRGRORJLHV IRU FRPSOH[ LQGXVWULDO SURFHVV FRQWURO DSSOLFDWLRQV ,QGXVWULDO FRQWURO DSSOLFDWLRQV DUH FKDUDFn WHUL]HG E\ QRQOLQHDU QRLV\ QRQ*DXVVLDQ KLJKO\ FRUUHODWHG DQG QRQVWDWLRQDU\ SURFHVVHV 7KLV LV DQ DSSOLFDWLRQV DUHD ZKHUH FODVVLFDO FRQWURO GHVLJQV KDYH SURYHQ LQHIIHFWLYH

PAGE 12

&DVH 6WXG\ $V D FDVH VWXG\ WKLV ZRUN GHYHORSV RQOLQH DGYLVRU\ QHXURFRQWUROOHUV GHVLJQHG WR PLQn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n ]DWLRQ SUREOHP UHTXLUHV D WHFKQRORJ\ WKDW FDQ ORRN DW WKH SURFHVV JOREDOO\ DQG GHWHUPLQH WKH DSSURSULDWH FRPELQDWLRQ RI FRPEXVWLRQ FRQWUROV 7KH QHXURFRQWUROOHUV GHYHORSHG LQ WKLV VWXG\ ZLOO IRUZDUG FRQWURO VHWSRLQWV WR RSHUDn WRUV WKURXJK WKH SODQWnV H[LVWLQJ GLVWULEXWHG FRQWURO V\VWHP '&6f 7KH QHXURFRQWUROOHUV ZLOO EH UHTXLUHG WR SURYLGH VHWSRLQWV WKDW PLQLPL]H 12[ HPLVVLRQV ZKLOH PDLQWDLQLQJ XQLW RSHUDWLQJ FRQVWUDLQWV $ GHPRQVWUDWLRQ V\VWHP KDV EHHQ FRPSOHWHG DW &DQDO (OHFWULFfV 0: WDQJHQWLDOO\ ILUHG 8QLW &KDUOHV 5LYHU $VVRFLDWHV &5$f DQG &RPPRQZHDOWK (QHUJ\ MRLQWO\ IXQGHG WKLV VWXG\ 2EMHFWLYHV 7KH DXWKRU EHOLHYHV WKDW QHXURFRQWURO VWUDWHJLHV KDYH QRW EHHQ PRUH VXFFHVVIXO LQ UHDO ZRUOG DSSOLFDWLRQV EHFDXVH WKH\ f DUH GLIILFXOW WR LPSOHPHQW f IDLO WR DFFRXQW IRU GHSHQGHQW LQWHUQDO SURFHVV VWDWHV DQG f KDYH GLIILFXOW\ GHDOLQJ ZLWK FRUUHODWHG SURFHVV YDULDEOHV 7KH VFLHQWLILF REMHFWLYHV IRU WKLV ZRUN DUH WR GHYHORS

PAGE 13

f DQ DSSOLFDWLRQEDVHG QHXURFRQWURO LPSOHPHQWDWLRQ PHWKRGRORJ\ f VWDWHVSDFH QHXURFRQWURO DUFKLWHFWXUHV f PHWKRGV IRU GHDOLQJ ZLWK FRUUHODWHG GDWD f DFFXUDWH FRPEXVWLRQ PRGHOV DQG f D QRYHO FRPEXVWLRQ FRQWUROOHU $SSOLFDWLRQ%DVHG 1HXURFRQWURO ,PSOHPHQWDWLRQ 0HWKRGRORJ\ 0RVW QHXURFRQWURO LPSOHPHQWDWLRQV LQ WKH OLWHUDWXUH KDYH EHHQ DG KRF DQG DSSOLFDWLRQ VSHFLILF >@ ,Q PDQ\ FDVHV WKH SURFHVV KDV EHHQ VLPXODWHG DQG WKXV NQRZQ FRPSOHWHO\ 0RGHP FRQWURO WKHRU\ KDV LQWURGXFHG PDQ\ VRSKLVWLFDWHG FRQWURO GHVLJQV EXW WKH IDFW UHPDLQV WKDW DSSUR[LPDWHO\ b RI WKH LQGXVWULDO FRQWURO DSSOLFDWLRQV DSSO\ D VLPSOH SURn SRUWLRQDO LQWHJUDO GLIIHUHQWLDO 3,'f FRQWUROOHU >@ 7KH 3,' FRQWUROOHU RQO\ UHTXLUHV WKH SURFHVV HQJLQHHU WR VSHFLI\ UHDVRQDEOH NQRZOn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n WUROOHU ZLOO EH LPSOHPHQWHG LQ &KDSWHU EDVHG RQ WKLV ODEHOLQJ RI WKH SURFHVV YDULDEOHV ZLWKRXW DGGLWLRQDO SURFHVV NQRZOHGJH 7KLV ZLOO DOORZ SURFHVV HQJLQHHUV ZLWK UHDVRQDEOH SURFHVV NQRZOHGJH DQG ZLWKRXW DQ\ NQRZOHGJH RI QHXUDO QHWZRUN RU QHXURFRQWURO WKHRU\ WR VXFFHVVIXOO\ GHSOR\ VXFK WHFKQRORJLHV 1RWLFH WKDW DFKLHYLQJ VXFK D VWUDLJKWIRUZDUG

PAGE 14

LPSOHPHQWDWLRQ PHWKRGRORJ\ GRHV QRW LPSO\ WKDW WKH GHWDLOV EHKLQG WKH QHXURFRQWUROOHU LPSOHPHQWDWLRQ DUH HDV\ LW VLPSO\ UHTXLUHV WKDW WKH\ FDQ EH DXWRPDWHG 6WDWH6SDFH 1HXURFRQWURO 'HVLJQV 1RQOLQHDU VWDWHVSDFH QHXUDO QHWZRUN DUFKLWHFWXUHV RIIHU WKH JUHDWHVW PRGHOLQJ SRWHQn WLDO EXW WKH GLIILFXOWLHV LQ WKHLU WUDLQLQJ KDYH OHG LQYHVWLJDWRUV WR UHMHFW WKHLU DSSOLFDWLRQ >@ ,Q IDFW PRVW QHXURFRQWURO GHVLJQV DUH EDVHG RQ LQSXWRXWSXW SURFHVV PRGHOV >@>@>@ 7KH XVH RI LQSXWRXWSXW QHXUDO QHWZRUN DUFKLWHFWXUHV LQ WKH GHVLJQ RI QHXURn FRQWUROOHUV DVVXPHV WKDW DOO LQSXW YDULDEOHV DUH LQGHSHQGHQW D VLWXDWLRQ ZKLFK LQ QRW OLNHO\ WR EH IRXQG LQ UHDOZRUOG DSSOLFDWLRQV 0DQ\ RI WKH SURFHVV YDULDEOHV WKDW DIIHFW WKH SODQWf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n JDWHV KDYH WHVWHG SXUHO\ VWDWHVSDFH DUFKLWHFWXUHV ZKHUH DOO VWDWHV DUH WUHDWHG DV KLGGHQ DQG XQNQRZQ ZLWK OLPLWHG WR QR VXFFHVV 7KLV ZRUN ZLOO HPSLULFDOO\ LQYHVWLJDWH VHYHUDO VKDGHV RI JUD\ UDQJLQJ IURP SXUHO\ PSXWRXWSXW WR SXUHO\ VWDWHVSDFH FRQWUROOHUV 7KHVH VWDWHVSDFH FRQWUROOHU GHVLJQV DUH SUH

PAGE 15

VHQWHG LQ &KDSWHU DQG WKHLU SHUIRUPDQFH LV HPSLULFDOO\ LQYHVWLJDWHG LQ FKDSWHUV DQG 7KH SULPDU\ GLIIHUHQFH LQ WKH VWDWHVSDFH UHSUHVHQWDWLRQV SURSRVHG KHUH LV WKDW WKH VWDWH YDULDEOHV DUH WUHDWHG DV QRW KLGGHQ DQG NQRZQ 7KLV ZLOO UHTXLUH DQ H[WHQVLRQ WR WKH QHXURn FRQWURO GHVLJQ VWUDWHJLHV SUHVHQWHG LQ WKH OLWHUDWXUH >@>@>@ 0HWKRGV IRU 'HDOLQJ ZLWK &RUUHODWLRQ ,QGXVWULDO SURFHVV DSSOLFDWLRQV DUH XQLTXH LQ WKDW WKHUH LV D PDVVLYH DPRXQW RI DYDLOn DEOH GDWD 7KH LQSXW YDULDEOHV IRU FRPSOH[ SURFHVV PRGHOV DUH W\SLFDOO\ KLJKO\ FRUUHODWHG D VLWXDWLRQ IRU ZKLFK WKHUH DUH IHZ VROXWLRQV LQ WKH OLWHUDWXUH 7KLV FRUUHODWLRQ FDQ FRPH IURP VHYHUDO VRXUFHV GHSHQGHQW VWDWHV DV DGGUHVVHG DERYHf SK\VLFDO OLQNDJHV DQG VRIW OLQNDJHV WKURXJK FRQWURO VWUDWHJLHV WKH ODFN RI DGHTXDWH V\VWHP SDUDPHWHUL]DWLRQ ,QGXVn WULDO SURFHVV FRQWURO DSSOLFDWLRQV ZLOO UHTXLUH WKH SURFHVV HQJLQHHU WR VHOHFW UHSUHVHQWDWLYH YDULDEOHV IURP D ODUJH VHW RI DYDLODEOH SURFHVV YDULDEOHV %HFDXVH RI LVVXHV OLNH LQSXW FRUn UHODWLRQ WKH UHSUHVHQWDWLYH YDULDEOHV VHOHFWHG ZLOO KDYH D VLJQLILFDQW LPSDFW RQ WKH SHUIRUn PDQFH RI WKH UHVXOWLQJ QHXURFRQWUROOHU $ YLDEOH QHXURFRQWURO GHVLJQ PHWKRGRORJ\ ZLOO KDYH WR EH DEOH WR FRSH ZLWK FRUUHODWHG SURFHVV YDULDEOHV /LQHDU FRQWURO WKHRU\ GHDOV ZLWK WKLV DVSHFW WKURXJK SDUDPHWHUL]DWLRQ RI WKH FRQWUROOHU 1RQOLQHDU FRQWURO WKHRU\ KRZHYHU KDV QRW VROYHG WKLV SUREOHP LQ JHQHUDO ,Q &KDSWHU ZH SURSRVH WR FRPSXWH VHQVLWLYLWLHV WKURXJK D FRPPLWWHH RI WUDLQHG QHXUDO PRGHOV WR VHOHFW WKH EHVW YDULDEOHV IRU V\VWHP LGHQn WLILFDWLRQ DQG FRQWURO 6WDWHVSDFH QHXURFRQWURO DUFKLWHFWXUHV ZLOO EH DEOH WR H[SOLFLWO\ GHDO ZLWK RQH VRXUFH RI FRUUHODWLRQ SUHVHQW LQ LQGXVWULDO SURFHVVHV QDPHO\ FRUUHODWLRQ SURGXFHG E\ GHSHQGHQW VWDWH YDULDEOHV $V PHQWLRQHG DERYH WKHUH DUH VHYHUDO RWKHU VRXUFHV RI FRUUHODWLRQ KRZn HYHU 2XU ILUVW JRDO LV WR HPSLULFDOO\ TXDQWLI\ WKH LPSDFW RI WKLV FRUUHODWLRQ RQ RXU PRGHO

PAGE 16

LQJ DQG FRQWURO REMHFWLYHV :H EHJLQ E\ LJQRULQJ RWKHU VRXUFHV RI FRUUHODWLRQ DQG LQYHVWLJDWH WKHLU LPSDFW LQ &KDSWHU 2QFH VSHFLILF SUREOHPV KDYH EHHQ LGHQWLILHG PHWKn RGV DUH GHYHORSHG IRU GHDOLQJ ZLWK FRUUHODWLRQ GXULQJ FRQWUROOHU LPSOHPHQWDWLRQ LQ &KDSn WHU DQG WKH SHUIRUPDQFH RI WKHVH PHWKRGV DUH TXDQWLILHG ZLWK UHVSHFW WR WKH SHUIRUPDQFH RI WKH UHVXOWLQJ FRQWUROOHUV $FFXUDWH &RPEXVWLRQ 0RGHOV /LWWOH LV NQRZQ DERXW KRZ 12[ LV IRUPHG IURP DLUERXQG QLWURJHQ GXULQJ FRPEXVWLRQ 7R GDWH UHOLDEOH PRGHOV IRU 12[ IRUPDWLRQ LQ HOHFWULF SRZHU ERLOHUV KDYH QRW EHHQ DYDLOn DEOH >@ 7KHUH DUH QRW DGHTXDWH PRGHOV IRU PDQ\ UHDOZRUOG LQGXVWULDO SURFHVVHV D IDFW ZKLFK KDV DOVR OLPLWHG WKH DFFHSWDQFH RI PRGHP FRQWURO VWUDWHJLHV 2QH UHDVRQ WKDW QHXURn FRQWURO KDV UHFHLYHG VXFK ZLGHVSUHDG DWWHQWLRQ LV EHFDXVH RI LWV SRWHQWLDO DELOLW\ WR GHDO ZLWK YHU\ FRPSOH[ SURFHVVHV WKDW KDYH HVFDSHG PRGHOLQJ &KDSWHU GHYHORSV DFFXUDWH FRPEXVWLRQ PRGHOV DFFRUGLQJ WR DFFHSWHG PRGHOLQJ SHUn IRUPDQFH PHWULFV &KDSWHU GHPRQVWUDWHV WKH LPSDFW WKDW FRUUHODWLRQ KDV RQ QHXURFRQWURO GHVLJQV DQG &KDSWHU LQYHVWLJDWHV LWV LPSDFW RQ WKH DFFXUDF\ RI WKH XQGHUO\LQJ SURFHVV PRGHOV +HUH LW ZLOO EH VKRZQ WKDW DFFXUDF\ LV VXEMHFWLYH DQG WKDW LQ IDFW QR JRRG PHWULFV IRU PRGHO DFFXUDF\ DUH DYDLODEOH LQ WKH OLWHUDWXUH $ QHZ PHWULF LV SURSRVHG DQG HPSLULn FDOO\ FRPSDUHG DJDLQVW DYDLODEOH PHWULFV LQ WKH OLWHUDWXUH LQ &KDSWHU $SSO\LQJ WKLV QHZ PHWULF SUHGLFWLYH FRPEXVWLRQ PRGHOV DUH GHYHORSHG DQG XVHG WR VKHG OLJKW RQ ZKLFK SURn FHVV YDULDEOHV KDYH WKH JUHDWHVW LPSDFW RQ 12[ DQG &2 IRUPDWLRQ

PAGE 17

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n WUDVW WKH IRXU FRQWURO GHVLJQV LQ &KDSWHU 7R WKH DXWKRUV NQRZOHGJH WKLV ZRUN GHYHORSHG WKH ILUVW 12[ FRQWUROOHU IRU D JDV DQG RLO FRILUHG HOHFWULF SRZHU SODQW 1HZ UHJXODWLRQV DQG WKH UHVWUXFWXULQJ RI WKH HOHFWULF SRZHU LQGXVWU\ KDYH FRPELQHG WR FUHDWH D 12[ WUDGLQJ PDUNHW 7KH DQQXDO EHQHILWV WR D JDV DQG RLO FRILUHG HOHFWULF SRZHU SODQW DVVRFLDWHG ZLWK D b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n YLGHG LQ &KDSWHU 7KH JRDO RI WKLV VHFWLRQ LV WR SURYLGH D SK\VLFDO XQGHUVWDQGLQJ IRU ZKDW ZH DUH DVNLQJ WKH FRQWUROOHUV WR SHUIRUP WKXV SURYLGLQJ MXVWLILFDWLRQ WKDW RXU REMHFWLYHV DUH IHDVLEOH &KDSWHU GHYHORSV IRXU GHWDLOHG QHXURFRQWURO GHVLJQV EHORQJLQJ WR WKH PRGHO

PAGE 18

SUHGLFWLYH PRGHO LQYHUVH DQG PRGHO UHIHUHQFH FRQWURO IDPLOLHV 7KH GHVLJQV DUH SUHVHQWHG DV JHQHUDOL]HG PHWKRGRORJLHV WKDW DUH DSSOLFDEOH WR DQ\ FRQWURO DSSOLFDWLRQ &KDSWHU SUHn VHQWV D PDQDJHPHQW DQG SUHSURFHVVLQJ PHWKRGRORJ\ IRU FROOHFWLQJ GDWD LQ VXSSRUW RI WKHVH FRQWURO GHVLJQV DQG WKH UHTXLUHG PRGHOLQJ (DFK FRQWURO GHVLJQ FRQVLGHUHG UHTXLUHV DFFXUDWH SURFHVV PRGHOV IRU LWV LPSOHPHQWDn WLRQ &KDSWHU SUHVHQWV D PRGHOLQJ PHWKRGRORJ\ IRU GHYHORSLQJ WKHVH PRGHOV 7KH FRQn WUROOHUV DUH WKHQ LPSOHPHQWHG LQ &KDSWHU 7KH SHUIRUPDQFH IRU HDFK UHVXOWLQJ FRQWUROOHU LV WKHQ TXDQWLILHG XVLQJ RIIOLQH VLPXODWLRQV DQG RQOLQH H[SHULPHQWV 6LJQLILFDQW SUREOHPV DUH GLVFRYHUHG ZLWK WKH FRQWUROOHUV IRU ZKLFK WKHUH DUH QR VROXWLRQV LQ WKH OLWHUDWXUH 7KHVH SUREOHPV DORQJ ZLWK SURSRVHG VROXWLRQV DUH LQYHVWLJDWHG IXUWKHU LQ &KDSWHU 7KLV VHFn WLRQ DGGLWLRQDOO\ GHPRQVWUDWHV WKH YDOLGLW\ RI WKHVH VROXWLRQV E\ TXDQWLI\LQJ WKH SHUIRUn PDQFH RI WKH UHYLVHG FRQWUROOHUV 7KH NH\ OHDUQLQJV DQG H[WHQVLRQV WR WKLV ZRUN DUH VXPPDUL]HG LQ &KDSWHU

PAGE 19

&+$37(5 /,7(5$785( 5(9,(: 7KLV FKDSWHU SURYLGHV EDFNJURXQG IRU WKH UHVW RI WKLV GRFXPHQW LQ WKH DUHDV RI f RSWLPL]DWLRQ f QHXUDO QHWZRUNV f QHXURFRQWURO f 12[ DQG f IRVVLOILUHG SRZHU SODQWV 2SWLPL]DWLRQ 0DWKHPDWLFDO RSWLPL]DWLRQ PHWKRGV DUH DW WKH KHDUW RI PRGHP PRGHOLQJ DQG FRQWURO DSSOLFDWLRQV 1HXUDO QHWZRUNV XVH RSWLPL]DWLRQ PHWKRGV WR IDFLOLWDWH OHDUQLQJ DQG FRQWURO DSSOLFDWLRQV DSSO\ WKHVH PHWKRGV WR PHHW WKHLU FRQWURO REMHFWLYHV 7KH QRWDWLRQ DQG PHWKn RGV SUHVHQWHG LQ WKLV VHFWLRQ ZLOO EH XVHG H[WHQVLYHO\ WKURXJKRXW WKH UHVW RI WKLV GRFXPHQW 2SWLPL]DWLRQ LV GHILQHG DV WKH SURFHVV RI ILQGLQJ WKH YDOXHV RI 1 GHFLVLRQ YDULDEOHV ] H 1 WKDW PLQLPL]H D VFDODU SHUIRUPDQFH REMHFWLYH H nLV f§! >@ )RUPDOO\ WKLV RSWLPL]DWLRQ WDVN ZLOO EH UHSUHVHQWHG DV $UJ0LQL^-]f` f ZKHUH KL LV WKH GHFLVLRQ YDULDEOH VSDFH ZKLFK LV PRVW RIWHQ WDNHQ WR EH HXFOLGHDQ 2SWLPL]DWLRQ PHWKRGV DOVR NQRZQ DV PDWKHPDWLFDO SURJUDPPLQJ PHWKRGV FDQ EH FODVVLILHG DFFRUGLQJ WR WKH DPRXQW RI D SULRUL LQIRUPDWLRQ DYDLODEOH DERXW WKH V\VWHP

PAGE 20

EHLQJ RSWLPL]HG >@ 7KH IROORZLQJ VHFWLRQV EURDGO\ FDWHJRUL]H RSWLPL]DWLRQ PHWKRGV LQWR WKH IROORZLQJ f &ODVVLFDO $QDO\WLF 2SWLPL]DWLRQ ZKHUH WKH V\VWHP EHLQJ RSWLPL]HG LV NQRZQ FRPSOHWHO\ RU QHDUO\ FRPSOHWHO\ DQG D WUDFWDEOH DQDO\WLF VROXWLRQ H[LVWV f 'HVFHQW 2SWLPL]DWLRQ ZKHUH ILUVW DQGRU VHFRQGRUGHU SDUWLDO GHULYDn WLYHV DUH DYDLODEOH HYHU\ZKHUH IRU WKH SDUDPHWHUV RI WKH V\VWHP EHLQJ RSWLPL]HG f 'LUHFW 2SWLPL]DWLRQ ZKHUH OLWWOH WR QR D SULRUL NQRZOHGJH H[LVWV DERXW WKH SK\VLFDO VWUXFWXUH RI WKH V\VWHP EHLQJ RSWLPL]HG ,I WKH RSWLPL]DWLRQ SUREOHP LQYROYHV REMHFWLYH IXQFWLRQV RU FRQVWUDLQWV ZKLFK FDQQRW EH VWDWHG DV H[SOLFLW IXQFWLRQV RI WKH GHVLJQ YDULDEOHV RU DUH WRR FRPSOLFDWHG WR PDQLSXODWH ZH FDQQRW VROYH LW E\ XVLQJ FODVVLFDO DQDO\WLF RSWLPL]DWLRQ PHWKRGV 7KLV ZRUN ZLOO EH GHDOLQJ ZLWK FRPSOH[ V\VWHPV ZKHUH OLWWOH LV NQRZQ D SULRUL DQG ZLOO WKHUHIRUH QRW FRQn VLGHU DQDO\WLF RSWLPL]DWLRQ PHWKRGV ,WHUDWLYH 2SWLPL]DWLRQ 0HWKRGV $OO GLUHFW DQG GHVFHQW RSWLPL]DWLRQ PHWKRGV DUH LWHUDWLYH LQ QDWXUH LH WKH\ VWDUW IURP DQ LQLWLDO WULDO VROXWLRQ DQG SURFHHG WRZDUG WKH PLQLPXP SRLQW LQ D VHTXHQWLDO PDQQHU $Q LWHUDWLYH RSWLPL]DWLRQ PHWKRG LV W\SLFDOO\ MXGJHG EDVHG RQ LWV UDWH RI FRQYHUJHQFH >@ ,Q JHQHUDO DQ RSWLPL]DWLRQ PHWKRG LV VDLG WR KDYH FRQYHUJHQFH RI RUGHU S LI __eQ f+, ! S! f __LQfLr_3 ZKHUH ]Qf DQG f GHQRWH WKH SRLQWV REWDLQHG DW WKH HQG RI LWHUDWLRQV Q DQG Q UHVSHFWLYHO\ ]r UHSUHVHQWV WKH RSWLPXP SRLQW DQG __[__ GHQRWHV WKH OHQJWK RU QRUP RI WKH YHFWRU [

PAGE 21

,I S DQG N WKH PHWKRG LV VDLG WR EH OLQHDUO\ FRQYHUJHQW FRUUHVSRQGV WR VORZ FRQYHUJHQFHf ,I S WKH PHWKRG LV VDLG WR EH TXDGUDWLFDOO\ FRQYHUJHQW FRUUHn VSRQGV WR IDVWHU FRQYHUJHQFHf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n RGV LQFOXGH >@ f 5DQGRP 6HDUFK f *ULG 6HDUFK f 3DWWHUQ 'LUHFWLRQV (OLPLQDWLRQ PHWKRGV 7KH H[KDXVWLYH VHDUFK PHWKRGV DUH VLPLODU WR D ODUJHU FODVV RI DOJRULWKPV NQRZQ DV HOLPLQDWLRQ PHWKRGV EHFDXVH WKH\ VHDUFK E\ HOLPLQDWLQJ SDUWV RI WKH LQWHUYDO (OLPLQDWLRQ PHWKRGV GLIIHU LQ KRZ WKH\ VHDUFK DQG GLVFDUG VXELQWHUYDOV 7KH PRUH FRPPRQ HOLPLQDn WLRQ PHWKRGV LQFOXGH >@ f 'LFKRWRPRXV 6HDUFK

PAGE 22

f ,QWHUYDO +DOYLQJ f )LERQDFFL 0HWKRG f *ROGHQ 6HFWLRQ 0HWKRG ,QWHUSRODWLRQ PHWKRGV ,QWHUSRODWLRQ PHWKRGV LWHUDWLYHO\ ILW WKH ORFDO SHUIRUPDQFH VXUIDFH ZLWK D VLPSOH SRO\n QRPLDO IRUP DQG WKHQ DSSUR[LPDWH WKH PLQLPXP SRLQW RI WKH V\VWHP DV WKH PLQLPXP SRLQW RI WKH SRO\QRPLDO >@ 7KHVH PHWKRGV DUH JHQHUDOO\ PRUH HIILFLHQW WKDQ HOLPLQDWLRQ PHWKRGV DQG FDQ EH DFFHOHUDWHG LI JUDGLHQW LQIRUPDWLRQ LV DYDLODEOH 6RPH RI WKH PRUH SRSXODU LQWHUSRODWLRQ PHWKRGV LQFOXGH >@ f 4XDGUDWLF 0HWKRG f &XELF 0HWKRG f 1HZWRQ 0HWKRG f 4XDVL1HZWRQ 0HWKRG f 6HFDQW 0HWKRG 8QUHVWULFWHG PHWKRGV :KHQ WKH GHVLJQ YDULDEOH UDQJH LV QRW NQRZQ WKH VHDUFK PXVW EH SHUIRUPHG ZLWKRXW UHVWULFWLRQV RQ WKH YDOXHV RI WKH YDULDEOHV 0RVW RI WKHVH PHWKRGV XVH D VWHS VL]H DQG PRYH IURP DQ LQLWLDO JXHVV LQ IDYRUDEOH GLUHFWLRQ SRVLWLYH RU QHJDWLYHf >@ 7KH VWHS VL]H XVHG PXVW EH VPDOO LQ UHODWLRQ WR WKH ILQDO DFFXUDF\ GHVLUHG 7KLV PHWKRG LV RIWHQ DFFHOHUDWHG E\ XVLQJ D YDULDEOH VWHS VL]H 7KHVH PHWKRGV LQFOXGH >@ f 6LPSOH[ 0HWKRG f 5HYLVHG 6LPSOH[ 0HWKRG f .DUPDUNDUfV 0HWKRG f +RRNfV DQG -HHYHVf 0HWKRG

PAGE 23

f 5RVHQEURFNfV 0HWKRG ,Q DGGLWLRQ HYROXWLRQDU\ FRPSXWLQJ WHFKQLTXHV OLNH JHQHWLF DOJRULWKPV EHORQJ WR WKLV FDWHJRU\ /LQH VHDUFK $OO RI WKH GLUHFW VHDUFK PHWKRGV SUHVHQWHG DERYH FDQ EH DSSOLHG WR ERWK RQHGLPHQn VLRQDO RU QGLPHQVLRQDO VHDUFKHV $ RQHGLPHQVLRQDO VHDUFK LV RIWHQ UHIHUUHG WR DV D OLQH VHDUFK VLQFH ZH DUH VHDUFKLQJ DORQJ D OLQH 7KH DLP RI DOO OLQH VHDUFKHV LV WR ILQG Tr H VXFK WKDW U_r $UJ0LQU>^-] U?]Gf` f ZKHUH ] LV WKH GHVLJQ YHFWRU DQG ]G LV D NQRZQ VHDUFK GLUHFWLRQ 2QH RI WKH PRVW HIILFLHQW DQG KHQFH PRVW SRSXODU OLQH VHDUFKHV XVHV WKH 4XDGUDWLF 0HWKRG WR ILQG Tr >@ 7KLV PHWKRG KDV EHHQ DSSOLHG LQ WKLV ZRUN XVLQJ WKH IROORZLQJ DOJRULWKP 1RUPDOL]H WKH VHDUFK YHFWRU ]G E\ GLYLGLQJ HDFK FRPSRQHQW E\ WKH DEVROXWH YDOXH RI WKH HOHPHQW RI ]G ZLWK WKH PD[LPXP DEVROXWH YDOXH ,, (YDOXDWH WKH IXQFWLRQ DW WKH SRLQWV $ DQG T ZKHUH T LV DQ LQLWLDO VWHS VL]H ,,, ,I -'!-$ WKHQ VHW & DQG % T ,9 (OVH VHW % DQG HYDOXDWH DW WKH SRLQW ( T ,I -( -' WKHQ VHW & ( (OVH VHW ( DQG T T DQG JRWR VWHS ,,, 9 &DOFXODWH W7 -E -$ -F -% a AF A -$ f

PAGE 24

9, ,I -% $ WKHQ VHW U_r UMr DQG TXLW 9,, ,I I_r % WKHQ VHW & % DQG % UMr (OVH VHW $ % DQG % IMr 9,,, *RWR VWHS 9 ZKHUH $ IX LV WKH PLQLPXP FKDQJH LQ WR GHWHFW HDUO\ VWRSSLQJ 'HVFHQW%DVHG 2SWLPL]DWLRQ ] :KHQ DOO YDOXHV RI ] H DUH SRVVLEOH DQG WKH IXQFWLRQ -]f KDV ILUVW DQG VHFRQG SDUn WLDO GHULYDWLYHV HYHU\ZKHUH WKH QHFHVVDU\ FRQGLWLRQV IRU D ORFDO PLQLPXP DUH A 6L E\ ZKLFK ZH PHDQ G-G]L 9L DQG f f E\ ZKLFK ZH PHDQ WKDW WKH P [ P PDWUL[ ZKRVH FRPSRQHQWV DUH G-G]IL]M PXVW EH SRVn LWLYH VHPL GHILQLWH LH KDYH HLJHQYDOXHV WKDW DUH ]HUR RU SRVLWLYH >@ $OO SRLQWV WKDW VDWLVI\ f DUH FDOOHG VWDWLRQDU\ SRLQWV 6XIILFLHQW FRQGLWLRQV IRU D ORFDO PLQLPXP DUH f DQG T!R f VU WKDW LV DOO HLJHQYDOXHV PXVW EH SRVLWLYH ,I f LV VDWLVILHG EXW G -G]a WKDW LV WKH GHWHUPLQDQW RI WKH PDWUL[ LV ]HUR PHDQLQJ WKDW RQH RU PRUH RI LWV HLJHQYDOXHV LV ]HURf

PAGE 25

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n RGV >@ 'HVFHQW VHDUFK PHWKRGV DUH LWHUDWLYH DOJRULWKPV IRU LPSURYLQJ HVWLPDWHV RI WKH GHFLn VLRQ YDULDEOH ] VR DV WR FRPH FORVHU WR VDWLVI\LQJ WKH FRQGLWLRQV IRU D VWDWLRQDU\ SRLQW 7KH VWHSV LQ XVLQJ WKH GHVFHQW PHWKRG DUH DV IROORZV 6HW Q DQG JXHVV DW WKH LQLWLDO GHVLJQ YHFWRU ]Qf XVXDOO\ UDQGRP ,, 'HWHUPLQH WKH YDOXHV RI G-G]Qf ,,, ,QWHUSUHWLQJ G-G]Qf DV WKH JUDGLHQW YHFWRU GHWHUPLQH WKH VHDUFK G GLUHFWLRQ ] Qf I G-G]Qff DV D IXQFWLRQ RI WKLV JUDGLHQW ,9 'HWHUPLQH WKH VWHS VL]H WR EH WDNHQ U_mf DV D IXQFWLRQ RI WKLV GLUHFWLRQ 9 8SGDWH WKH HVWLPDWHV RI ]Q f ]Qf U?Qf]GQf 9, 5HSHDW ,, XQWLO G+G]QffG+G]Qff7 LV YHU\ VPDOO

PAGE 26

7KH YDULDWLRQV LQ GHVFHQWEDVHG RSWLPL]DWLRQ FDQ EH H[SUHVVHG DV YDULDWLRQV LQ WKH YUI GHWHUPLQDWLRQ RI WKH GLUHFWLRQ YHFWRU I DQG WKH VWHS VL]H 6RPH RI WKH PRUH FRPPRQ YDULDWLRQV LQFOXGH >@ >@>@ f 6WHHSHVW 'HVFHQW ]G^Qf G+G]Qf f U_mf FRQVWDQW f f 6WHHSHVW 'HVFHQW ZLWK 0RPHQWXP ]GQf SfG+GLQf S]GQf f U? mf FRQVWDQW f f &RQMXJDWH *UDGLHQWV G G+G]MQ ff7?G+G]Q fG+GLQf@ G+G]Q ff7G+G]^Q f U_mf /LQH6HDUFK]mff f f 1HXUDO 1HWZRUNV $UWLILFLDO QHXUDO QHWZRUNV $11Vf DUH ELRORJLFDOO\ PRWLYDWHG GDWD SURFHVVLQJ VWUXFWXUH WKDW FRQVLVW RI D ODUJH QXPEHU RI UHODWLYHO\ VLPSOH KLJKO\ LQWHUFRQQHFWHG QHXURQV RU SURn FHVVLQJ HOHPHQWV 3(Vf >@ ,Q JHQHUDO WKHVH VWUXFWXUHV SURYLGH DQ LQGXFWLYH PDWKHPDWLn FDO PRGHO WKDW FDQ EH UHSUHVHQWHG E\ M! 0UL!f f \A ZKHUH } LV WKH PRGHOfV LQSXWRXWSXW PDS DQG M! H H DQG WY H A DUH LWV RXWSXWV LQSXWV DQG FRHIILFLHQWV UHVSHFWLYHO\ 7KH FRHIILFLHQWV LQ DQ

PAGE 27

$11 PDS DUH FRPPRQO\ UHIHUUHG WR DV ZHLJKWV $UWLILFLDO QHXUDO QHWZRUNV LQIHU RU OHDUQ WKH UHODWLRQVKLSV EHWZHHQ S DQG Z E\ REVHUYLQJ DFWXDO SURFHVV GDWD ,Q WKLV ZD\ $11V FDQ EH DSSOLHG WR JHQHUDOL]HG UHJUHVVLRQ DQG FODVVLILFDWLRQ LQIHUHQFH SUREOHPV $11 DUFKLWHFWXUHV SRVVHVV WZR IXQGDPHQWDO SURSHUWLHV f 7KH\ DUH FDSDEOH RI DSSUR[LPDWLQJ WR DUELWUDU\ DFFXUDF\ DQ\ FRQWLQXRXV IXQFWLRQ LH WKH\ DUH XQLYHUVDO PDSSHUV >@ f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fV DUFKLWHFWXUH GHILQHV WKH ZD\ LQ ZKLFK LW SURFHVVHV LQSXW LQIRUPDWLRQ WR SURGXFH RXWSXW LQIRUPDWLRQ LH WKH IRUP RI WKHLU PDWKHPDWLFDO LQSXWRXWSXW PDS 0XOWLOD\HU SHUFHSWURQ 0RVW $11V SUHVHQWHG LQ WKH OLWHUDWXUH DUH VWDWLF PDSSHUV LH WKH\ DUH RQO\ FDSDEOH RI PRGHOLQJ VWDWLF RU VWHDG\VWDWH SURFHVV UHODWLRQVKLSV %\ IDU WKH PRVW SRSXODU DQG ZLGHO\ DSSOLHG $11 DUFKLWHFWXUH LV FDOOHG WKH PXOWLOD\HU SHUFHSWURQ 0/3f >@ 7KLV QHWZRUN FRQVLVWV RI IXOO\LQWHUFRQQHFWHG OD\HUV RI 3(V ZLWK ORJLVWLF UHVSRQVH FKDUDFWHULVWLFV 7KH 0/3 QHWZRUN LV W\SLFDOO\ FRQILJXUHG ZLWK RQH RU WZR KLGGHQ OD\HUV RI 3(V $ WZROD\HU 0/3 LV LOOXVWUDWHG LQ )LJXUH

PAGE 28

)LJXUH 0XOWLOD\HU SHUFHSWURQ PRGHO DUFKLWHFWXUH )RUPDOO\ XVLQJ PDWUL[ DOJHEUD WKLV DUFKLWHFWXUH LV JLYHQ E\ f ZKHUH H An LV WKH PDWUL[ RI ZHLJKWV IRU WKH ILUVW KLGGHQ OD\HU WfKr H A LV D YHFWRU RI ELDV YDOXHV IRU WKLV OD\HU H A 1Onn LV WKH PDWUL[ RI ZHLJKWV IRU WKH VHFn RQG KLGGHQ OD\HU Wfn H An LV WKH ELDV YHFWRU IRU WKLV OD\HU -S9 H I9! 9n LV WKH PDWUL[ 1\ [ 1K s\ 1 RI ZHLJKWV IRU WKH RXWSXW OD\HU E H LV WKH FRUUHVSRQGLQJ ELDV YHFWRU 1 LV WKH QXPEHU RI 3(V LQ WKH ILUVW KLGGHQ OD\HU 1KO LV WKH QXPEHU RI 3(V LQ WKH VHFRQG KLGGHQ OD\HU D LV WKH WDQK ORJLVWLF IXQFWLRQ DQG WKH VHW 7Y EKa -3\ En` UHSUHn VHQWV WKH PRGHOf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f 7KH G\QDPLF 3(V

PAGE 29

FDQ KDYH UHVSRQVH FKDUDFWHULVWLFV EDVHG RQ D SULRUL SURFHVV NQRZOHGJH RU FRQWDLQ DGDSWLYH PHPRU\ PHFKDQLVPV RU IGWHUV 7LPHGHOD\ QHXUDO QHWZRUN 7KH PRVW FRPPRQ '11 LV FDOOHG WKH WLPHGHOD\ QHXUDO QHWZRUN 7'11f >@ 7KLV DUFKLWHFWXUH FRQVLVWV RI D 0/3 ZKHUH HDFK LQSXW 3( KDV DQ DGDSWLYH OLQHDU ),5 ILOWHU DV LOOXVWUDWHG LQ )LJXUH )LJXUH 7'11 LQSXW 3( FRQQHFWLYLW\ )RUPDOO\ WKH 7'11 FDQ EH GHVFULEHG E\ n}f 17f n,8 MH17@ 9 \ IGQn?^Wf f IQOSIGOWf 1Wf X!f WQLS LGO f f $GL 1X [ 7 ZKHUH f§} [ UHSUHVHQWV WKH WDSSHGGHOD\ OLQH RSHUDWRU DQG 1 EHU RI WDSV LQ WKH GHOD\ OLQH LV WKH QXP *DPPD QHXUDO QHWZRUN 7KH PDLQ GLVDGYDQWDJH WR PRVW '11V LV WKDW WKH\ SUHSURFHVV WKH LQSXW WR H[WUDFW IL[HG DQG NQRZQ G\QDPLFV RI WKH SURFHVV GDWD UDWKHU WKDQ OHDUQ WKHVH G\QDPLFV IURP WKLV GDWD 7KH 7'11 FDQ EH FRQVLGHUHG DV DQ H[FHSWLRQ WR WKLV UXOH EXW KHUH WKH SURFHVV PXVW KDYH

PAGE 30

ILQLWH LPSXOVH UHVSRQVH ),5f G\QDPLFV RI NQRZQ RUGHU 7KH *DPPD 1HXUDO 1HWZRUN *11f >@ UHSUHVHQWV DQ LPSRUWDQW FODVV RI G\QDPLF $11 PRGHOV WKDW LV DEOH WR OHDP LQILQLWH LPSXOVH UHVSRQVH +5f SURFHVV G\QDPLFV ZLWKRXW D SULRUL NQRZOHGJH DERXW WKH VWUXFWXUH RU RUGHU RI WKHVH G\QDPLFV )LJXUH *DPPD PHPRU\ SURFHVVLQJ HOHPHQW 7KH *11 DUFKLWHFWXUH LV FRQFHSWXDOO\ DQ 0/3 ZLWK DGDSWLYH *DPPD )LOWHUV *)f SODFHG DW WKH RXWSXW RI LWV LQSXW DQG KLGGHQ OD\HU 3(V $ VLQJOH *) LV LOOXVWUDWHG LQ )LJXUH )RUPDOO\ WKH *11 LV JLYHQ E\ MIM:2 QWf m M 9f \\IM Bnf f M H 17@ ,Wf fRWI-nIR -Sn 1Uf Vnnf 1M-ff Wnf LW f RI 1f 1f ; 1f 9 ZKHUH A UHSUHVHQWV WKH *) 1 LV WKH QXPEHU RI *) WDSV LQ WKH LQSXW U OD\HU DQG 1 LV WKH QXPEHU RI *) WDSV LQ WKH ILUVW KLGGHQ OD\HU 7KLV DUFKLWHFWXUH KDV EHHQ SUHVHQWHG ZLWKRXW D *) LQ WKH VHFRQG KLGGHQ OD\HU EXW VXFK D FRQILJXUDWLRQ ZRXOG EH D VWUDLJKWIRUZDUG H[WHQVLRQ

PAGE 31

1RQOLQHDU VWDWHVSDFH PRGHO 7KH *11 XVHV WKH *DPPD )LOWHU WR UHSUHVHQW SURFHVV G\QDPLFV 7KH *) DSSUR[LPDWHV WKHVH G\QDPLFV IURP D *DPPD PHPRU\ NHUQHO EDVLV 7KH *DPPD NHUQHOV DUH DEOH WR PRGHO DQ LPSRUWDQW FODVV RI G\QDPLFV EXW PD\ QRW EH WKH EHVW UHSUHVHQWDWLRQ IRU JHQHUDO SURFHVV G\QDPLFV $Q DOWHUQDWLYH DSSURDFK WR XVLQJ *DPPD NHUQHOV LV WR GHVLJQ WKH $11 DUFKLWHFWXUH ZLWK DQ H[SOLFLW VWDWH DQG GDWD IORZ VWUXFWXUH WKDW LV FDSDEOH RI OHDUQLQJ XQLn YHUVDO SURFHVV G\QDPLFV 7KLV DSSURDFK LV WKH JRDO RI WKH QRQOLQHDU VWDWHVSDFH PRGHO 1/66f ZKLFK LPSOHPHQWV SURFHVV G\QDPLFV GLUHFWO\ DV D QRQOLQHDU VWDWH HYROXWLRQ HTXDn WLRQ DQG DQ RXWSXW REVHUYDWLRQ HTXDWLRQ >@ DV JLYHQ E\ LR LRLfRWIJ f 3 O!: f ZKHUH f H A LV WKH PRGHOV VWDWH YHFWRU FRQVLVWLQJ RI 1r KLGGHQ 3(Vr LV DQ $11 PDS GHVFULELQJ WKH WLPH HYROXWLRQ RI WKLV VWDWH WYY DUH WKH ZHLJKWV RI WKLV VWDWH QHWZRUN LV D VHFRQG $11 PDS GHVFULELQJ KRZ RXWSXWV DUH SURGXFHG IURP WKLV VWDWH DQG WL!\ DUH WKH ZHLJKWV RI WKLV RXWSXW QHWZRUN )LJXUH LOOXVWUDWHV WKH FRQILJXUDWLRQ RI D 1/66 QHWZRUN )LJXUH 1RQOLQHDU VWDWH VSDFH QHXUDO QHWZRUN FRQILJXUDWLRQ

PAGE 32

/HDUQLQJ $OJRULWKPV 7KH ELRORJLFDO URRWV RI QHXUDO QHWZRUNV DUH UHVSRQVLEOH IRU WKH ZLGHVSUHDG XVH RI WKH WHUP OHDUQLQJ WR GHVFULEH WKH SURFHVV GXULQJ ZKLFK WKH QHWZRUN SDUDPHWHUV DUH FKDQJHG WR LPSURYH WKH SHUIRUPDQFH RI WKH QHXUDOQHWZRUNEDVHG V\VWHP $Q $11 OHDUQLQJ DOJRn ULWKP VSHFLILHV KRZ LWV ZHLJKWV DUH XSGDWHG LQ UHVSRQVH WR WUDLQLQJ GDWD 7KHVH DOJRULWKPV DUH VLPSO\ RSWLPL]DWLRQ PHWKRGV DSSOLHG WR WKH WDVN RI ILQGLQJ WKH EHVW PRGHO ZHLJKWV WR PLQLPL]H D VSHFLILHG PRGHOLQJ REMHFWLYH LH $UJ0LQA^-f ,Q JHQHUDO DQ\ RI WKH RSWLPL]DWLRQ PHWKRGV SUHVHQWHG DERYH FDQ EH XVHG WR VROYH WKLV SUREOHP 2QH RI WKH PRVW VLJQLILFDQW EUHDNWKURXJKV LQ WKH ILHOG RI $11V ZDV WKH UHDOL]DWLRQ WKDW WKH FKDLQ UXOH IRU RUGHUHG SDUWLDO GHULYDWLYHV SURYLGHV D PHFKDQLVP IRU GHULYLQJ WKH ILUVW RUGHU JUDGLHQWV IRU DOO ZHLJKWV LQ D PRGHO HYHQ WKRXJK WKH PRGHOLQJ REMHFWLYH LV RQO\ DQ H[SOLFLW IXQFWLRQ RI WKH PRGHOfV RXWSXWV >@ 5HFDOO WKDW WKH FKDLQ UXOH IRU RUGHUHG SDUWLDO GHULYDWLYHV LV JLYHQ E\ LLR 7 [ 7 W $[ f $SSO\LQJ WKH FKDLQ UXOH DOORZV VHQVLWLYLWLHV WR EH FDOFXODWHG IURP WKH RXWSXW RI WKH PRGHO EDFN WR LWV LQSXW ZKLFK LV ZK\ WKH UHVXOWLQJ DOJRULWKP KDV EHHQ FRLQHG fEDFNSURSD JDWLRQf LQ WKH OLWHUDWXUH >@ :KHQ WKH YDULDEOHV DUH WHPSRUDOO\ UHODWHG WKH FKDLQ UXOH KDV WKH IROORZLQJ IRUP f +HUH LQ DGGLWLRQ WR EDFNSURSDJDWLQJ VHQVLWLYLWLHV IURP WKH PRGHOfV RXWSXW WR LWV LQSXW WKH VHQVLWLYLWLHV DUH EDFNSURSDJDWHG WKURXJK WLPH

PAGE 33

7KH PRVW FRPPRQ RSWLPL]DWLRQ PHWKRG LV VLPSO\ VWHHSHVWGHVFHQW ZLWK PRPHQWXP DOWKRXJK PDQ\ YDULDWLRQV KDYH EHHQ GHPRQVWUDWHG WR VLJQLILFDQWO\ LPSURYH FRQYHUJHQFH >@ 7KH LVVXHV OHDGLQJ WR WKH VHOHFWLRQ RI RQH RSWLPL]DWLRQ PHWKRG RYHU DQRWKHU DUH f &RQYHUJHQFH 5DWH f ,PSOHPHQWDWLRQ &RPSOH[LW\ f &RQILJXUDWLRQ &RPSOH[LW\ f $YRLGDQFH RI /RFDO 0LQLPD f 6HQVLWLYLW\ WR &RUUHODWLRQ 7KH PRVW FRPPRQ PRGHOLQJ REMHFWLYH XVHG LQ $11V LV WKH PHDQ VTXDUHG HUURU 06(f D EHWZHHQ WKH PRGHOfV RXWSXW \ DQG D VSHFLILHG GHVLUHG UHVSRQVH G H DV JLYHQ E\ WY: 1n “r ; LZR02f W L f ZKHUH 79r LV WKH QXPEHU RI VDPSOHV LQ WKH WUDLQLQJ GDWDVHW /HDUQLQJ UXOHV ZKLFK XVH WKH 06( FULWHULRQ DUH FRPPRQO\ FODVVLILHG DV VXSHUYLVHG OHDUQLQJ UXOHV EHFDXVH RI WKH SUHVn HQFH RI D fWHDFKHUf LPSOLHG E\ WKH H[SOLFLW VSHFLILFDWLRQ RI D GHVLUHG UHVSRQVH /HDUQLQJ UXOHV ZLWKRXW H[SOLFLW UHIHUHQFH WR D GHVLUHG UHVSRQVH IRU WKH PRGHO LQ WKH REMHFWLYH IXQFn WLRQ DUH FRPPRQO\ UHIHUUHG WR DV XQVXSHUYLVHG OHDUQLQJ UXOHV *HQHUDOL]DWLRQ $V XQLYHUVDO PDSSHUV $11V DUH DOPRVW DOZD\V PRUH FRPSOH[ WKDQ WKH UHODWLRQVKLSV WKDW ZH VHHN WR XQFRYHU 7KH QHW UHVXOW LV WKDW $11V DUH QRWRULRXV IRU RYHUILWWLQJ D WUDLQn LQJ GDWDVHW LH SHUIRUPLQJ ZHOO RQ WUDLQLQJ GDWD EXW SRRUO\ RQ D EOLQG WHVW GDWDVHW >@ ,W

PAGE 34

LV YHU\ LPSRUWDQW WR RSWLPL]H WKH FRPSOH[LW\ RI WKH QHXUDO QHWZRUN LQ RUGHU WR DFKLHYH WKH EHVW JHQHUDOL]DWLRQ %LDV DQG YDULDQFH &RQVLGHUDEOH LQVLJKW LQWR WKLV SKHQRPHQRQ FDQ EH REWDLQHG E\ LQWURGXFLQJ WKH FRQFHSW RI WKH ELDVYDULDQFH WUDGHRII %LVKRS >@ REVHUYHV WKDW WKH JHQHUDOL]DWLRQ HUURU e XVLQJ WKH (XFOLGHDQ QRUP ZLOO GHSHQG RQ D SDUWLFXODU GDWDVHW RQ ZKLFK WKH QHWZRUN ZDV WUDLQHG 7KH GHSHQGHQFH RQ FDQ EH HOLPLQDWHG E\ FRQVLGHULQJ DQ DYHUDJH RYHU WKH FRPn SOHWH HQVHPEOH RI GDWDVHWV ZKLFK FDQ EH ZULWWHQ DV A ('>G?fIff@ f ZKHUH G?! GHQRWHV WKH FRQGLWLRQDO DYHUDJH RU UHJUHVVLRQ RI WKH GHVLUHG GDWD JLYHQ E\ G?Âœf IS?fGW f DQG SG?Xf LV WKH FRQGLWLRQDO GHQVLW\ RI WKH GHVLUHG YDULDEOH G FRQGLWLRQHG RQ WKH LQSXW YHFWRU X %LVKRS >@ GHPRQVWUDWHV WKDW WKLV JHQHUDOL]DWLRQ HUURU FDQ EH GHFRPSRVHG LQWR WKH VXP RI WKH ELDV VTXDUHG SOXV WKH YDULDQFH ('?IÂœf@G?Âœff ('>If('>I7rf@@ f $ PRGHO ZKLFK LV WRR VLPSOH RU WRR LQIOH[LEOH ZLOO KDYH D ODUJH ELDV ZKLOH RQH ZKLFK KDV WRR PXFK IOH[LELOLW\ LQ UHODWLRQ WR WKH SDUWLFXODU GDWDVHW ZLOO KDYH D ODUJH YDULDQFH %LDV DQG YDULDQFH DUH FRPSOHPHQWDU\ TXDQWLWLHV DQG WKH EHVW JHQHUDOL]DWLRQ LV REWDLQHG ZKHQ ZH KDYH WKH EHVW FRPSURPLVH EHWZHHQ WKH FRQIOLFWLQJ UHTXLUHPHQWV RI VPDOO ELDV DQG VPDOO YDULDQFH 7KH YDULDQFH RI WKH SUHGLFWLRQ ZLOO EH IXUWKHU DGGUHVVHG EHORZ 6HFWLRQ 6WDQGDUG (UURUV

PAGE 35

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n VWUDLQHG E\ WKH GDWD WKHUHE\ DOVR UHGXFLQJ WKH YDULDQFH ,I WKH QXPEHU RI GDWD VDPSOHV LV LQFUHDVHG UDSLGO\ LQ UHODWLRQ WR WKH PRGHO FRPSOH[LW\ ZH FDQ ILQG D VHTXHQFH RI PRGHOV VXFK WKDW ERWK ELDV DQG YDULDQFH GHFUHDVH 0RGHOV VXFK DV $11V FDQ LQ SULQFLSOH SURYLGH FRQVLVWHQW HVWLPDWRUV RI DUELWUDU\ DFFXUDF\ DV WKH QXPEHU RI GDWD SRLQWV LV LQFUHDVHG WR LQILQLW\ 1RWH WKDW HYHQ LI ERWK WKH ELDV DQG YDULDQFH FDQ EH UHGXFHG WR ]HUR WKH JHQHUDOLn ]DWLRQ HUURU ZLOO VWLOO EH QRQ]HUR GXH WR WKH LQWULQVLF QRLVH LQ WKH GDWD 2QH UDUHO\ KDV LQILQLWH GDWD DQG SUDFWLFDO LVVXHV OLNH WUDLQLQJ WLPH PDNH VLPSO\ DGGLQJ PRUH GDWD SRLQWV LPSUDFWLFDO 7KHUH DUH VHYHUDO SUDFWLFDO DQG SUDFWLFHG ZD\V WR LPSURYH PRGHO JHQHUDOL]DWLRQ ZH VWDUW ZLWK UHJXODUL]DWLRQ 5HJXODUL]DWLRQ 5HJXODUL]DWLRQ ZDV RULJLQDOO\ SURSRVHG E\ 7LNKDQRY >@ DV D PHWKRG IRU VROYLQJ LOO SRVHG SUREOHPV 7KH EDVLF LGHD LV WR VWDELOL]H WKH VROXWLRQ E\ PHDQV RI VRPH DX[LOLDU\ QRQQHJDWLYH IXQFWLRQDO WKDW HPEHGV SULRU LQIRUPDWLRQ HJ VPRRWKQHVV FRQVWUDLQWV RQ WKH LQSXWRXWSXW PDSSLQJ 5HJXODUL]DWLRQ LV DEOH WR WUDQVIRUP DQ LOOSRVHG SUREOHP LQWR D ZHOOSRVHG SUREOHP >@

PAGE 36

7LNKDQRYfV UHJXODUL]DWLRQ WKHRU\ XVHV D UHJXODUL]DWLRQ SHQDOW\ WHUP RI WKH IRUP p MSIM f ZKHUH 3 LV D OLQHDU SVHXGRf GLIIHUHQWLDO RSHUDWRU 7KLV SHQDOW\ WHUP LV DGGHG WR WKH REMHFn WLYH IXQFWLRQ WR JLYH M \UR f ZKHUH < LV WKH UHJXODUL]DWLRQ SDUDPHWHU 3ULRU LQIRUPDWLRQ DERXW WKH IRUP RI WKH VROXWLRQ LH WKH SODQWf LV HPEHGGHG LQ WKH RSHUDWRU 3 7KH RSHUDWRU 3 LV UHIHUUHG WR DV D VWDELOL]HU LQ WKH VHQVH WKDW LW VWDELOL]HV WKH VROXWLRQ \ PDNLQJ LW VPRRWK 7KH DSSURSULDWH FKRLFH IRU 3 DQG WKH VROXWLRQ WR f UHTXLUHV IXQFWLRQDO DQDO\VLV DQG LV EH\RQG WKH VFRSH RI WKLV ZRUN 7KH PRVW FRPPRQO\ XVHG IRUP RI UHJXODUL]HU KRZHYHU LV TXLWH VLPSOH WR LPSOHPHQW :HLJKW GHFD\ UHJXODUL]HU WHUPV FRQVLVW RI WKH VXP RI VTXDUHV RI WKH DGDSWLYH SDUDPHWHUV LQ WKH QHWZRUN f ZKHUH WKH VXP UXQV RYHU WKH ZHLJKWV DQG ELDVHV ,Q FRQYHQWLRQDO FXUYH ILWWLQJ WKH XVH RI WKLV IRUP RI UHJXODUL]HU LV FDOOHG ULGJH UHJUHVVLRQ ,W KDV EHHQ IRXQG HPSLULFDOO\ WKDW D UHJn XODUL]HU RI WKLV IRUP FDQ OHDG WR VLJQLILFDQW LPSURYHPHQWV LQ JHQHUDOL]DWLRQ >@ *URZLQJ DQG SUXQLQJ DOJRULWKPV 7KH WRSRORJ\ RI D QHXUDO QHWZRUN QXPEHU RI XQLWV DQG LQWHUFRQQHFWLRQV FDQ KDYH D VLJQLILFDQW LPSDFW RQ LWV SHUIRUPDQFH 5HJXODUL]DWLRQ KHOSV WR PLQLPL]H WKLV LPSDFW ZKHQ WKH FRPSOH[LW\ RI WKH QHWZRUN LV ODUJHU WKDQ UHTXLUHG IRU WKH SDUWLFXODU DSSOLFDWLRQ &OHDUO\ KRZHYHU D EHWWHU DSSURDFK LV WR PDWFK WKH FRPSOH[LW\ RI WKH PRGHO ZLWK WKH FRP

PAGE 37

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n ZRUN DV WKH LQLWLDO JXHVV IRU WUDLQLQJ WKH QH[W QHWZRUN ZLWK WKH H[WUD ZHLJKWV LQLWLDOL]HG UDQGRPO\f 7HFKQLTXHV RI WKLV IRUP DUH FDOOHG JURZLQJ DOJRULWKPV $Q DOWHUQDWLYH DSSURDFK LV WR VWDUW ZLWK D UHODWLYHO\ ODUJH QHWZRUN DQG JUDGXDOO\ UHPRYH XQLWV WKHVH DUH NQRZQ DV SUXQLQJ DOJRULWKPV 0RVW RI WKHVH SURFHGXUHV DUH DG KRF DQG WDLORU WR VSHFLILF DSSOLFDWLRQV WKDW LV QRW WR VD\ KRZHYHU WKDW WKH\ DUH LQHIIHFWLYH 0RUH UHFHQW ZRUN KDV WDNHQ DGYDQWDJH RI GHYHORSPHQWV LQ GLVFUHWH RSWLPL]DWLRQ XVLQJ JHQHWLF DOJRULWKPV >@ *HQHWLF DOJRULWKPV SURYLGH D PHWKRGLFDO ZD\ RI VHDUFKLQJ ODUJH GLVFUHWH VSDFHV PRUH HIILFLHQWO\ &URVVYDOLGDWLRQ $Q DOWHUQDWLYH WR UHJXODUL]DWLRQ DV D ZD\ RI FRQWUROOLQJ WKH HIIHFWLYH FRPSOH[LW\ RI D QHWZRUN LV WKH SURFHGXUH RI FURVVYDOLGDWLRQ >@ 7KH WUDLQLQJ RI D QRQOLQHDU PRGHO FRUUHn VSRQGV WR WKH LWHUDWLYH UHGXFWLRQ RI WKH HUURU IXQFWLRQ GHILQHG ZLWK UHVSHFW WR WKH WUDLQLQJ

PAGE 38

GDWDVHW 'XULQJ WUDLQLQJ WKH HUURU ZLOO JHQHUDOO\ PRQRWRQLFDOO\ GHFUHDVH DV D IXQFWLRQ RI WKH QXPEHU RI SUHVHQWDWLRQV RI WKH WUDLQLQJ GDWDVHW LH HSRFKV +RZHYHU WKH JHQHUDOL]Dn WLRQ HUURU ZLWK UHVSHFW WR DQ LQGHSHQGHQW GDWDVHW FDOOHG WKH YDOLGDWLRQ GDWDVHW RIWHQ VKRZV D GHFUHDVH DW ILUVW IROORZHG E\ DQ LQFUHDVH DV WKH QHWZRUN VWDUWV WR RYHUILW 7UDLQLQJ FDQ WKHUHIRUH EH VWRSSHG DW WKH SRLQW RI VPDOOHVW HUURU ZLWK UHVSHFW WR WKH YDOLGDWLRQ GDWDVHW DV WKLV SURGXFHV D QHWZRUN ZLWK WKH VPDOOHVW JHQHUDOL]DWLRQ HUURU RU DW OHDVW DQ DSSUR[LPDn WLRQ WKHUHRIf &RPPLWWHHV RI QHWZRUNV ,Q SUDFWLFH EXLOGLQJ QHXUDO QHWZRUN PRGHOV UHTXLUHV WKH WUDLQLQJ RI PDQ\ GLIIHUHQW FDQGLGDWH QHWZRUNV DQG WKHQ WKH VHOHFWLRQ RI WKH EHVW SHUIRUPHU 7\SLFDOO\ SHUIRUPDQFH LV EDVHG RQ WKH QHWZRUNV SHUIRUPDQFH RQ D WKLUG GDWDVHW QRW XVHG IRU WUDLQLQJ RU FURVVYDOLGDn WLRQ 7KHUH DUH WZR GLVDGYDQWDJHV WR WKLV DSSURDFK )LUVW DOO RI WKH HIIRUW LQYROYHG LQ WUDLQn LQJ WKH UHPDLQLQJ QHWZRUNV LV ZDVWHG DQG VHFRQGO\ WKH JHQHUDOL]DWLRQ SHUIRUPDQFH RQ WKH YDOLGDWLRQ GDWDVHW KDV D UDQGRP FRPSRQHQW GXH WR WKH QRLVH RQ WKH GDWD >@ 7KH QHWn ZRUN ZKLFK SHUIRUPHG WKH EHVW RQ WKLV GDWDVHW PLJKW QRW EH WKH RQH ZLWK WKH ORZHVW JHQHUn DOL]DWLRQ HUURU 5HFDOO WKDW WKH JHQHUDOL]DWLRQ HUURU LV DYHUDJHG RYHU DOO GDWDVHWV f 7KHVH OLPLWDWLRQV FDQ EH RYHUFRPH E\ FRPELQLQJ WKH QHWZRUNV WRJHWKHU WR IRUP D FRPPLWWHH >@>@ 7KLV DSSURDFK ZDV VKRZQ WR SURYLGH VLJQLILFDQW LPSURYHPHQWV LQ WKH JHQHUDOL]DWLRQ HUURU 'HQRWH WKH FRPPLWWHH SUHGLFWLRQ DV
PAGE 39

ZKHUH WY DUH WKH ZHLJKWV RI FRPPLWWHH PHPEHU L DQG MIU @ LV WKH VHW RI DOO ZHLJKWV IRU WKH FRPPLWWHH %LVKRS >@ VKRZV WKDW LI WKH HUURUV RI WKH LQGLYLGXDO FRPPLWWHH PHPEHU DUH GHFRUUH ODWHG WKHQ WKH FRPPLWWHH ZLOO DOZD\V KDYH D ORZHU JHQHUDOL]DWLRQ HUURU WKDQ DQ\ RI LWV LQGLYLGXDO PHPEHUV 6WDQGDUG (UURUV 7LEVKLUDQL >@ UHYLHZV D QXPEHU RI PHWKRGV IRU HVWLPDWLQJ WKH VWDQGDUG HUURU RI SUHn GLFWHG YDOXHV IURP D PXOWLOD\HU SHUFHSWURQ 7KHVH LQFOXGH GLUHFW HYDOXDWLRQ RI PD[LPXP OLNHOLKRRGV EDVHG RQ WKH +HVVLDQ PDWUL[ WKH fVDQGZLFKf HVWLPDWRU DQG WKH ERRWVWUDS PHWKRG 7LEVKLUDQL RIIHUV WKH IROORZLQJ REVHUYDWLRQV f 7KH ERRWVWUDS PHWKRGV SURYLGHG WKH PRVW DFFXUDWH HVWLPDWHV RI WKH VWDQn GDUG HUURUV RI SUHGLFWHG YDOXHV f 7KH QRQVLPXODWLRQ PHWKRGV GHOWD DQG VDQGZLFKf PLVVHG WKH VXEVWDQn WLDO YDULDELOLW\ GXH WR WKH UDQGRP LQLWLDO ZHLJKWV IURP WKH PXOWLSOH WUDLQn LQJ UXQV 7KH QRQVLPXODWLRQ PHWKRGV DUH VROYHG DQDO\WLFDOO\ DQG WKHUHIRUH UHTXLUH XQLTXH VROXWLRQV IRU HDFK QHWZRUN WRSRORJ\ :KHUHDV WKH ERRWVWUDS PHWKRGV DSSO\ WR DOO QHWZRUN WRSRORJLHV DV ZHOO DV QRQQHXUDO QHWZRUN SDUDGLJPV 7KH DGGLWLRQDO IDFW DV QRWHG DERYH WKDW WKH ERRWVWUDS PHWKRGV DFFRXQW IRU ORFDO PLQLPD SURYLGHV VWURQJ DUJXPHQW IRU WKHLU XVH %RRWVWUDS PHWKRGV %RRWVWUDS PHWKRGV ZRUN E\ FUHDWLQJ PDQ\ SVHXGRUHSOLFDWHV fERRWVWUDS GDWDVHWVff IURP WKH WUDLQLQJ GDWDVHW DQG WKHQ UHHVWLPDWLQJ WKH PRGHOV ZHLJKWV WY RQ HDFK ERRWVWUDS GDWDVHW 7KHUH DUH WZR GLIIHUHQW DSSURDFKHV WR ERRWVWUDSSLQJ >@ 2QH FDQ FRQVLGHU HDFK

PAGE 40

WUDLQLQJ FDVH DV D VDPSOLQJ XQLW DQG VDPSOH ZLWK UHSODFHPHQW IURP WKH WUDLQLQJ GDWDVHW FDVHV WR FUHDWH D ERRWVWUDS VDPSOH 7KLV LV RIWHQ FDOOHG WKH fERRWVWUDS SDLUVf PHWKRG 7KH ERRWVWUDS SDLUV VDPSOLQJ DOJRULWKP LV JLYHQ E\ *HQHUDWH 1E VDPSOHV HDFK RQH RI VL]H 79 GUDZQ ZLWK UHSODFHPHQW r A $7 IURP WKH E WUDLQLQJ REVHUYDWLRQV ^XLf GLf`c L DQG WKH E WK VDPn SOH E\ ^X?Lf G?Lf`L ,, )RU HDFK ERRWVWUDS VDPSOH E H > @ ILQG $UJ0LQAK^-GEI^OEEff` f ,,, (VWLPDWH WKH VWDQGDUG HUURU RI WKH L WK SUHGLFWLRQ DV F\cf -f§ P r!f \f f ZKHUH 1E \ ? A 2Q WKH RWKHU KDQG RQH FDQ FRQVLGHU WKH SUHGLFWRUV DV IL[HG WUHDW WKH PRGHO UHVLGXDOV $ G \ DV WKH VDPSOLQJ XQLWV DQG FUHDWH D ERRWVWUDS VDPSOH E\ DGGLQJ UHVLGXDOV WR WKH PRGHO ILW 7KLV LV FDOOHG WKH fERRWVWUDS UHVLGXDOVf DSSURDFK )LQG $UJ0LQA^-GI[ WYff` IURP WKH 9V WUDLQLQJ REVHUYDWLRQV ^mf GLf`I L DQG OHW ULf GLf ILf WYf

PAGE 41

,, *HQHUDWH $A ERRWVWUDS VDPSOHV HDFK RQH RI VL]H @?I GUDZQ ZLWK UHSODFHPHQW IURP ^ULf [ DQG WKH E WK VDPSOH E\ ^U f`‘ L OHWn WLQJ \E ^ WIf$]f`IO f ,,, )RU HDFK ERRWVWUDS VDPSOH E H > 9D@ ILQG $UJ0LQAK^-\E IX WY$ff` f ,9 (VWLPDWH WKH VWDQGDUG HUURU RI WKH L WK SUHGLFWLRQ DV 9 A)f -f§ ; : Gf \Wf f $n L 1RWH WKDW ERWK RI WKHVH PHWKRGV UHTXLUH ILWWLQJ D PRGHO UHWUDLQLQJ WKH QHWZRUNf 9D WLPHV 7\SLFDOO\ 9D LV LQ WKH UDQJH 9D ,Q VLPSOH OLQHDU OHDVW VTXDUHV UHJUHVn VLRQ LW FDQ EH VKRZQ WKDW WKH ERRWVWUDS PHWKRGV ERWK DJUHH ZLWK WKH VWDQGDUG OHDVW VTXDUHV IRUPXOD DV $A f§} FR 7KH ERRWVWUDS PHWKRGV ZLOO DUULYH DW FRQILGHQFH LQWHUYDOV \c FFRQI\cf \ L FFRQI\cf f ZKHUH FFRQM GHSHQGV RQ WKH GHVLUHG FRQILGHQFH OHYHO D 7KH IDFWRU F&QM FDQ EH WDNHQ IURP D WDEOH ZLWK WKH SHUFHQWDJH SRLQWV RI WKH 6WXGHQWfV W GLVWULEXWLRQ ZLWK WKH QXPEHU RI GHJUHHV RI IUHHGRP HTXDO WR WKH QXPEHU RI ERRWVWUDS UXQV 9D

PAGE 42

&RQWURO 7KHRU\ $Q $11 LV FDSDEOH RI PRGHOLQJ DQ\ SURFHVV PDNLQJ WKHP LGHDO FDQGLGDWHV IRU FRPn SOH[ SURFHVV RSWLPL]DWLRQ DQG FRQWURO VWUDWHJLHV 1HXURFRQWURO LV EXW D VXEILHOG RI FODVVLn FDO FRQWURO WKHRU\ >@ 7R SXW QHXURFRQWURO LQ SHUVSHFWLYH LW LV LPSRUWDQW WR FRQVLGHU LWV SODFH ZLWKLQ WKLV ILHOG &ODVVLFDO &RQWURO 7KHRU\ &ODVVLFDO FRQWURO WKHRU\ LV VWURQJO\ ELDVHG WRZDUGV OLQHDU WLPHLQYDULDQW V\VWHPV >@ *HQHUDO QRQOLQHDU V\VWHPV VLPSO\ GR QRW DOORZ XV EHFDXVH RI WKHLU DQDO\WLFDO LQWUDFWDELOn LW\ WR IRUPXODWH D WKHRU\ WKDW LV DV VWURQJ DV WKDW RI OLQHDU V\VWHP WKHRU\ 2Q WKH RWKHU KDQG QRQOLQHDU V\VWHPV FDQ EH TXDOLWDWLYHO\ VLPLODU WR OLQHDU RQHV XQGHU VRPH FLUFXPn VWDQFHV /LQHDU FRQWURO /LQHDU FRQWURO LV FRQFHUQHG ZLWK V\VWHPV RI WKH IRUP [ ‹ f ZLWK VWDWH [ LQSXW X PHDVXUDEOH RXWSXW S f DQG FRQWUROOHUV RI WKH IRUP X )[ •[r f ZKHUH nr LV WKH UHIHUHQFH VWDWH WKDW LV WKH VWDWH WR ZKLFK WKH SODQW LV WR EH EURXJKW ZLWK WKH KHOS RI WKH FRQWUROOHU >@ 7KH JRDOV RI OLQHDU FRQWURO DUH f $OWHULQJ WKH FORVHGORRS EHKDYLRU RI WKH V\VWHP WR VRPH XVHUGHILQHG UHVSRQVH FKDUDFWHULVWLFV

PAGE 43

f&RQWUROOLQJ WKH FORVHGORRS VWDELOLW\ LH FRQYHUJHQFH EDFN WR DQ HTXLn OLEULXP SRLQW DIWHU GLVWXUEDQFH 7KH GLVDGYDQWDJHV RI OLQHDU FRQWURO GHVLJQV f $VVXPH WKDW WKH ZRUOG LV OLQHDU *DXVVLDQ DQG VWDWLRQDU\ ZKHQ LQ UHDOLW\ WKH ZRUOG LV QRQH RI WKH DERYH f 5HTXLUH FRPSOHWH D SULRUL NQRZOHGJH RI WKH SURFHVV G\QDPLFV f 5HTXLUH WKDW WKH SURFHVV LV FRQWUROODEOH DQG REVHUYDEOH f &DQQRW IROORZ D UHIHUHQFH WUDMHFWRU\ SURGXFHG E\ D V\VWHP RI ORZHU RUGHU WKDW WKH SURFHVV 5REXVW FRQWURO 5REXVW FRQWURO DGGUHVVHV WKH SUREOHP RI FRQWUROOLQJ D SODQW ZKRVH EHKDYLRU LV VOLJKWO\ GLIIHUHQW IURP WKDW RI D SODQW PRGHO >@ 7KH UHDVRQV IRU WKH GLIIHUHQFH DUH SUHGRPLQDWHO\ WKH HIIHFW RI WKH QRQOLQHDU QRQ*DXVVLDQ DQG QRQVWDWLRQDU\ ZRUOG $ SRSXODU SUDJPDWLF FODVVLFDO DSSURDFK WR UREXVW FRQWURO LV FRQFHUQHG ZLWK SUHVHUYLQJ VWDELOLW\ >@ 7KH FORVHG ORRS HLJHQYDOXHV DUH FKRVHQ VR WKDW WKH\ UHPDLQ LQ WKH VWDELOLW\ UHJLRQ HYHQ LI WKH SODQW PRGHO VKRXOG FKDQJH LQ D GHILQHG UDQJH $OWKRXJK UREXVW FRQWURO VWUDWHJLHV DUH SULPDULO\ GHVLJQHG WR FRPSHQVDWH IRU GLIIHUn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

PAGE 44

$GDSWDWLRQ VFKHPHV FDQ EH EDVHG RQ ERWK D UHIHUHQFH PRGHO DQG D FRVW IXQFWLRQDO >@ 7KH DSSURDFK FDOOHG PRGHO UHIHUHQFH DGDSWLYH FRQWURO 05$&f LV E\ LWV QDPH FRPPLWWHG WR WKH IRUPHU 7KLV DSSURDFK LV EDVHG RQ IRUPXODWLQJ WKH UXOHV IRU FRPSXWLQJ WKH GLUHFWLRQ RI FKDQJH RI FRQWUROOHU SDUDPHWHUV DV D IXQFWLRQ RI WKH GLIIHUHQFH EHWZHHQ WKH EHKDYLRU RI WKH FORVHGORRS V\VWHP DQG D UHIHUHQFH PRGHO &RQWUROOHU SDUDPHWHUV FDQ EH DGDSWHG HLWKHU GLUHFWO\ RU YLD WKH HVWLPDWLRQ RI SODQW PRGHO SDUDPHWHUV $ PRUH JHQHUDO DSSURDFK LV WKDW RI VHOIWXQLQJ UHJXODWRUV 675f >@ ZKLFK FRQVLVWV LQ DGDSWLYH HVWLPDWLRQ RI D SODQW PRGHO DQG DSSO\LQJ D IRUPDOL]HG FRQWUROOHU GHVLJQ PHWKRG WR WKH SODQW PRGHO 7KLV GHVLJQ PHWKRG FDQ EH EDVHG RQ FRVW IXQFWLRQ RSWLPL]DWLRQ /LNH UREXVW FRQWURO KRZHYHU DGDSWLYH FRQWURO LPSOHPHQWDWLRQV KDYH EHHQ EDVHG RQ OLQHDU RU VLPSOH QRQOLQHDU SDUDPHWULF DVVXPSWLRQV DERXW WKH SURFHVV $V D UHVXOW DGDSn WLYH FRQWURO GHVLJQV KDYH QRW GHPRQVWUDWHG VLJQLILFDQW VXFFHVVIXO ZLWK FRPSOH[ UHDOZRUOG SURFHVVHV 1RQOLQHDU FRQWURO 1RQOLQHDU FRQWURO WKHRU\ LV FRQFHUQHG ZLWK JHQHUDO V\VWHPV RI WKH IRUP >@ L r •f f ZLWK PHDVXUDEOH RXWSXW 3 Jrf f DQG FRQWUROOHUV RI WKH IRUP F FA erf f 7KH JHQHUDO IRUPXODWLRQ RI QRQOLQHDU FRQWURO KROGV SURPLVH WR RYHUFRPH DOO RI WKH OLPn LWDWLRQV RI WKH FODVVLFDO FRQWURO VFKHPHV SUHVHQWHG DERYH 7KH ILHOGV WUDFN UHFRUG KRZn HYHU GRHV QRW GHOLYHU RQ WKLV SURPLVH 7KH SUREOHP LV WKDW DQ DQDO\WLFDO VROXWLRQ LV NQRZQ

PAGE 45

RQO\ IRU D UHVWULFWLYH VXEFODVV RI QRQOLQHDU V\VWHPV 7KH GLIILFXOWLHV ZLWK JHQXLQH QRQOLQn HDU FRQWUROOHU GHVLJQV KDYH W\SLFDOO\ OHDG WR D OLQHDUL]DWLRQ DSSURDFK DOVR NQRZQ DV JDLQ VFKHGXOLQJ 2SWLPDO FRQWURO 7KH WRSLF RI RSWLPDO FRQWURO WKHRU\ LV WR GHVLJQ FRQWUROOHUV WKDW DUH RSWLPL]HG WR D FHUn WDLQ SHUIRUPDQFH FULWHULRQ >@ &ODVVLFDO RSWLPDO FRQWURO KDV SULPDULO\ IRFXVHG RQ DSSOLn FDWLRQV ZKHUH VXFK RSWLPDOLW\ FRXOG EH SURYHQ DQDO\WLFDOO\ )RU H[DPSOH IRU D OLQHDU SODQW DQG D TXDGUDWLF SHUIRUPDQFH FULWHULRQ WKH 5LFDWWL FRQWUROOHU UHSUHVHQWV DQ H[SOLFLW DQG JORn EDO VROXWLRQ >@ '\QDPLF RSWLPL]DWLRQ SURYLGHV DQRWKHU H[DPSOH IRU VWDWH HYDOXDWLRQ DQG VHOHFWLRQ RI WKH RSWLPDO DFWLRQ ZKLFK FDQ EH SURYHQ RSWLPDO LQ FHUWDLQ DSSOLFDWLRQV $OWHUn QDWLYHO\ LI HDFK VWDWH DW HDFK VDPSOLQJ SHULRG LV UHSUHVHQWHG E\ D QRGH LQ D GLUHFWHG JUDSK DQG DFWLRQV DUH UHSUHVHQWHG E\ FRQQHFWLQJ HGJHV RI WKH VXEVHTXHQW VWDWHV WKHQ WKH WDVN FDQ DOVR EH WUDQVIRUPHG WR WKH FULWLFDO JUDSK SUREOHP RI JUDSK WKHRU\ >@ ,Q LWV PRVW JHQHUDO IRUP WKH RSWLPDO FRQWURO SUREOHP FDQ EH IRUPXODWHG DV DQ RSWLPLn ]DWLRQ SUREOHP >@ 7KH SODQW FDQ EH JHQHUDOL]HG DV LQ f DQG WKH JRDO LV WR ILQG D FRQn WUROOHU GHVFULEHG E\ f WKDW PLQLPL]HV e>rrrf@ f ZKHUH (> @ LV WKH PHDQ YDOXH RYHU WLPH 1HXURFRQWURO 0RVW QHXURFRQWURO DUFKLWHFWXUHV DUH HLWKHU H[SOLFLW RU GLVJXLVHG DQDORJLHV RI FODVVLFDO FRQWURO GHVLJQ VXFK DV RSWLPDO FRQWURO RU QXPHULFDO O\DSXQRYIXQFWLRQEDVHG GHVLJQ PHWKRGV ,W KDV EHHQ DUJXHG >@ WKDW LW LV RQO\ WKH UHSUHVHQWDWLRQ RI IXQFWLRQV E\ QHXUDO

PAGE 46

QHWZRUNV WKDW GHILQHV WKH ILHOG RI QHXURFRQWURO LQ WKH EURDG VHQVH DQG GLIIHUHQWLDWHV QHXURn FRQWURO IRUP FODVVLFDO FRQWURO PHWKRGV 7KH DXWKRU DJUHHV WKDW WKH XVH RI QRQSDUDPHWULF PRGHOV GRHV GLIIHUHQWLDWH QHXURFRQn WURO IURP FODVVLFDO FRQWUROV EXW ZRXOG DUJXH WKDW WKH SULPDU\ GHSDUWXUH IURP DQG H[WHQn VLRQ WR WKH SRWHQWLDO RI FRQWURO WKHRU\ LV QHXURFRQWUROfV ZLOOLQJQHVV WR GHSDUW IURP D UHTXLUHPHQW IRU DQDO\WLF VROXWLRQV 1HXURFRQWURO GHVLJQV VHHN WR UHDOL]H WKH SURPLVH RI JHQHUDO QRQOLQHDU FRQWURO UHSODFLQJ DQDO\WLF RSWLPL]DWLRQ PHWKRGV ZLWK QXPHULF RQHV 5HVHDUFKHUV LQ WKH ILHOG RI QHXUDO QHWZRUNV DUH DFFXVWRPHG WR ZRUNLQJ LQ DQ LQWUDFWDEOH ZRUOG DQG KDYH EHHQ ZLOOLQJ WR UHVROYH LPSRUWDQW TXHVWLRQV OLNH VWDELOLW\ UREXVWQHVV DQG FRQVLVWHQF\ HPSLULFDOO\ 7KH UHVXOW KDV EHHQ DQG ZLOO FRQWLQXH WR EH DQ LPSRUWDQW H[WHQn VLRQ WR FRQWURO WKHRU\ 7KHUH DUH PDQ\ W\SHV RI QHXURFRQWURO DUFKLWHFWXUHV LQ WKH OLWHUDWXUH HDFK ZLWK VSHFLILF DGYDQWDJHV DQG GLVDGYDQWDJHV 7KH IROORZLQJ VHFWLRQV UHYLHZ VRPH RI WKH FRQFHSWXDO QHXn URFRQWURO VWUDWHJLHV ZKLFK KDYH EHHQ SURSRVHG LQ WKH OLWHUDWXUH 0RGHOSUHGLFWLYH FRQWURO :KHQ D PRGHO LV XVHG LQGLUHFWO\ DQG RIIOLQH WKH FRQWURO VFKHPH LV XVXDOO\ UHIHUUHG WR DV PRGHOSUHGLFWLYH FRQWURO 03&f >@ ,Q PRVW LQGXVWULDO SURFHVV FRQWURO DSSOLFDWLRQV D SULRUL NQRZOHGJH DERXW WKH SURFHVV LV KDUG WR REWDLQ DQG EODFNER[ PRGHOV PXVW EH XVHG 7KH RIIOLQH WUDLQLQJ SKDVH SHUIRUPV VXSHUYLVHG OHDUQLQJ WR GHYHORS DQ $11 PRGHO IRU WKH SURFHVV WR EH FRQWUROOHG LH WKH $11 DWWHPSWV WR PLPLF WKH SURFHVV DIWHU EHLQJ H[SRVHG WR DFWXDO SURFHVV GDWD 7KLV SKDVH FDQ EH VWDWHG DV $UJ0LQA^-I IXS WYff` f

PAGE 47

ZKHUH I DQG LI DUH WKH SURFHVV RXWSXWV WR EH FRQWUROOHG DQG LQSXWV WR EH PDQLSXODWHG UHVSHFWLYHO\ $W WKH RQOLQH FRQWURO SKDVH WKH $11 PRGHO FDQQRW EH XVHG DORQH LW PXVW EH LQFRUSRn UDWHG ZLWK D PRGHOEDVHG FRQWURO VFKHPH >@ 7KLV FRQWURO VFKHPH LV RQFH DJDLQ DQ RSWLn PL]DWLRQ SUREOHP ZKLFK FDQ EH VWDWHG DV $UJ0LQcI^-Sr LI r!ff` f ZKHUH \r J GHQRWHV WKH GHVLUHG FORVHGORRS SURFHVV RXWSXW 7KLV RSWLPL]DWLRQ LV SHUn IRUPHG UHSHDWHGO\ DW HDFK WLPH LQWHUYDO GXULQJ WKH FRXUVH RI IHHGEDFN FRQWURO 0RGHOLQYHUVH FRQWURO $Q $11 FDQ EH WUDLQHG WR GHYHORS DQ LQYHUVH PRGHO RI D SURFHVV >@ +HUH WKH PRGHOfV LQSXW LV WKH SURFHVV RXWSXW DQG WKH PRGHOfV RXWSXW LV WKH SURFHVV LQSXW 7KH RIIOLQH WUDLQLQJ SKDVH FDQ EH VWDWHG DV $UJ0LQ^-WIIIr!ff` f &OHDUO\ WKH LQYHUVH PRGHO LV D VWHDG\VWDWH PRGHO RU WKH UHVXOWLQJ FRQWUROOHU ZRXOG EH QRQFDXVDO *LYHQ D GHVLUHG SURFHVV VHWSRLQW \r WKH DSSURSULDWH RQOLQH FRQWURO VLJQDO LI FDQ EH LPPHGLDWHO\ FDOFXODWHG DV WI $3rWL!f f 6XFFHVVIXO DSSOLFDWLRQV RI LQYHUVH PRGHOLQJ DUH GLVFXVVHG LQ >@ DQG >@ 2EYLRXVO\ DQ LQYHUVH PRGHO H[LVWV RQO\ ZKHQ WKH SURFHVV EHKDYHV PRQRWRQLFDOO\ DV D fIHHGIRUZDUGf IXQFWLRQ DW VWHDG\VWDWH ,I QRW WKLV DSSURDFK LV QRW DSSOLFDEOH

PAGE 48

&RQWUROOHU PRGHOLQJ $QRWKHU VLPSOH GLUHFW QHXURFRQWURO VFKHPH LV WR XVH D QHXUDO QHWZRUN WR PRGHO DQ H[LVWLQJ FRQWUROOHU 7KH LQSXW WR WKH H[LVWLQJ FRQWUROOHU LV XVHG DV WUDLQLQJ LQSXW WR WKH $11 PRGHO DQG WKH FRQWUROOHU RXWSXW VHUYHV DV WKH GHVLUHG UHVSRQVH 7KLV DSSURDFK LV VLPLODU WR WKH PRGHOLQYHUVH FRQWURO H[FHSW WKDW WKH GHVLUHG UHVSRQVH KHUH LV QRW D SURFHVV EXW D FRQn WUROOHU 7KLV DSSURDFK FDQ EH IRUPXODWHG DV $UJ0LQr^-^+F "!ff` f ZKHUH XF DUH WKH GHFLVLRQ YDULDEOHV JHQHUDWHG E\ DQ H[LVWLQJ FRQWUROOHU LQ UHVSRQVH WR WKH SODQW VWDWHV [" /LNH D SURFHVV D FRQWUROOHU LV JHQHUDOO\ G\QDPLF DQG RIWHQ FRPSULVHV LQWHJUDWRUV RU GLIIHUHQWLDWRUV ,I DQ DOJHEUDLF IHHGIRUZDUG QHWZRUN LV XVHG WR PRGHO WKH H[LVWLQJ FRQWUROn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r If WI [S WYFf` f

PAGE 49

ZKHUH I LV DQ $11 WKDW LV GLUHFWO\ FRQWUROOLQJ WKH SURFHVV LQSXWV DQG WYF DUH WKH ZHLJKWV RI WKLV QHWZRUN 1RWLFH WKDW WKH RSWLPL]DWLRQ FULWHULRQ LV RQO\ D IXQFWLRQ RI WKH DFWXDO DQG GHVLUHG SURFHVV RXWSXWV 7KLV PHDQV WKDW WKH RSWLPL]DWLRQ PHWKRGRORJ\ HPSOR\HG PXVW EH DEOH WR OHDUQ WYF ZLWKRXW DQ H[SOLFLW GHVLUHG UHVSRQVH RU HYHQ D PDWKHPDWLFDO OLQNn DJH WR WKH FULWHULRQ 7KH NH\ IHDWXUH RI WKLV GLUHFW DGDSWLYH DSSURDFK LV WKDW D SURFHVV PRGHO LV QHLWKHU NQRZQ LQ DGYDQFH QRU H[SOLFLWO\ GHYHORSHG GXULQJ FRQWURO GHVLJQ 7KLV PRVW FRPPRQ OHDUQLQJ DOJRULWKP IRU WKLV W\SH RI FRQWURO GHVLJQ LV UHIHUUHG WR DV UHLQIRUFHPHQW OHDUQLQJ 7KH ILUVW ZRUN LQ WKLV DUHD ZDV WKH fDGDSWLYH FULWLFf DOJRULWKP SURSRVHG E\ %DUWR HW DO >@ 6XFK DQ DOJRULWKP FDQ EH FRQVLGHUHG DV DQ DSSUR[LPDWH YHUVLRQ RI G\QDPLF SURJUDPPLQJ >@>@ ODWHU FRLQHG DV 1HXUR'\QDPLF 3URJUDPPLQJ >@ 'HVSLWH LWV KLVWRULFDO LPSRUWDQFH DQG LQWXLWLYH DSSHDO PRGHOIUHH DGDSWLYH QHXURFRQn WURO LV QRW DSSURSULDWH IRU PRVW UHDO ZRUOG DSSOLFDWLRQV 7KH SODQW LV PRVW OLNHO\ RXW RI FRQn WURO GXULQJ WKH OHDUQLQJ SURFHVV DQG IHZ LQGXVWULDO SURFHVVHV FDQ WROHUDWH WKH ODUJH QXPEHU RI IDLOXUHV UHTXLUHG WR DGDSW WKH FRQWUROOHU 0RGHOUHIHUHQFH GLUHFW FRQWURO )URP D SUDFWLFDO SHUVSHFWLYH RQH ZRXOG SUHIHU WR OHW IDLOXUHV WDNH SODFH LQ D VLPXODWHG HQYLURQPHQW ZLWK D SURFHVV PRGHO UDWKHU WKDQ LQ D UHDO SODQW (YHQ LI IDLOXUHV DUH QRW GLVDVn WURXV WKH\ FDQ FDXVH VXEVWDQWLDO ORVVHV 7KH SHUIRUPDQFH RI D FRQWUROOHU FRXOG EH HYDOXDWHG EDVHG RQ D PRGHO IRU WKH SURFHVV UDWKHU WKDQ WKH SURFHVV LWVHOI 7KH WUDLQLQJ VWDJH RI WKH FRQWURO GHVLJQ FDQ EH JLYHQ E\ $UJ0LQ AP^-\SI?X WYnff` f

PAGE 50

DQG WKH FRQWURO GHVLJQ EHFRPHV $UJ0LQAF^-\r I?X WYnff F r" f` f ,Q WKH FRXUVH RI PRGHOLQJ WKH SODQW WKH SODQW PXVW EH RSHUDWHG fQRUPDOO\f LQVWHDG RI EHLQJ GULYHQ RXW RI FRQWURO $IWHU WKH PRGHOLQJ VWDJH WKH PRGHO FDQ EH XVHG IRU FRQWUROOHU GHVLJQ ,I D SURFHVV PRGHO LV DOUHDG\ DYDLODEOH DQ $11 FRQWUROOHU FDQ EH GHYHORSHG LQ D VLPXODWLRQ LQ ZKLFK IDLOXUHV FDQQRW FDXVH DQ\ ORVV EXW WKDW RI FRPSXWHU WLPH $ QHXUDO QHWn ZRUN FRQWUROOHU DIWHU H[WHQVLYH WUDLQLQJ LQ WKH VLPXODWLRQ FDQ WKHQ EH LQVWDOOHG LQ WKH DFWXDO FRQWURO V\VWHP 0RGHO5HIHUHQFH GLUHFW FRQWURO VFKHPHV KDYH QRW RQO\ SURYHQ HIIHFWLYH LQ VHYHUDO VWXGLHV > @>@ EXW KDYH DOVR DOUHDG\ SURGXFHG QRWDEOH HFRQRPLF EHQHILWV >@ 7KHVH DSSURDFKHV FDQ EH XVHG IRU ERWK RIIOLQH FRQWURO DQG IRU RQOLQH DGDSWDWLRQ 12[ 7KH &OHDQ $LU $FW $PHQGPHQWV RI UHTXLUH WKDW HOHFWULF XWLOLWLHV PDNH VLJQLILFDQW UHGXFWLRQV LQ QLWURJHQ R[LGH 12[f HPLVVLRQV IURP WKHLU IRVVLOILUHG SRZHU SODQWV 7R GDWH PRVW HIIRUWV WR UHGXFH 12[ HPLVVLRQV KDYH FRPH IURP H[SHQVLYH KDUGZDUH UHWURILWV ZLWK OHVV WKDQ VDWLVIDFWRU\ SHUIRUPDQFH )XUWKHU FRPSOLFDWLQJ PDWWHUV FRQGLWLRQV WKDW GHFUHDVH 12[ IRUPDWLRQ ORZHU WHPSHUDWXUH H[FHVV IXHOf UHVXOW LQ WKH IRUPDWLRQ RWKHU SROn OXWLQJ FRPSRXQGV PDLQO\ FDUERQ PRQR[LGH &2f 6LPLODU HPLVVLRQV UHGXFWLRQV DUH EHLQJ UHTXLUHG LQ (XURSH WKURXJK ORFDO DQG (XURSHDQ (FRQRPLF &RPPXQLW\ (&&f LQLWLDWLYHV 1LWURJHQ PRQR[LGH 12f DQG QLWURJHQ GLR[LGH 1f DUH E\SURGXFWV RI WKH FRPEXVn WLRQ SURFHVV RI YLUWXDOO\ DOO IRVVLO IXHOV +LVWRULFDOO\ WKH TXDQWLW\ RI WKHVH LQRUJDQLF FRPn SRXQGV LQ WKH SURGXFWV RI FRPEXVWLRQ ZDV QRW VXIILFLHQW WR DIIHFW ERLOHU SHUIRUPDQFH WKHLU

PAGE 51

SUHVHQFH ZDV ODUJHO\ LJQRUHG ,Q UHFHQW \HDUV R[LGHV RI QLWURJHQ KDYH EHHQ VKRZQ WR EH NH\ FRQVWLWXHQWV LQ WKH FRPSOH[ SKRWRFKHPLFDO R[LGDQW UHDFWLRQ ZLWK VXQOLJKW WR IRUP VPRJ 7RGD\ WKH HPLVVLRQ RI 1 DQG 12 FROOHFWLYHO\ UHIHUUHG WR DV 12[f LV UHJXODWHG E\ WKH &OHDQ $LU $FW $PHQGPHQWV DQG KDV EHFRPH DQ LPSRUWDQW FRQVLGHUDWLRQ LQ WKH GHVLJQ RI IXHO ILULQJ HTXLSPHQW 12[ LV IRUPHG E\ WZR SULPDU\ PHFKDQLVPV WKHUPDO 12[ DQG IXHOERXQG 12[ 7KHUn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n WDLQLQJ IXHOV &RQGLWLRQV WKDW GHFUHDVH 12[ IRUPDWLRQ ORZHU WHPSHUDWXUH H[FHVV IXHOf FDQ UHVXOW LQ LQFRPSOHWH FRPEXVWLRQ 7KHVH FRQGLWLRQV UHVXOW LQ WKH IRUPDWLRQ RWKHU SROOXWn LQJ FRPSRXQGV PDLQO\ FDUERQ PRQR[LGH &2f 5HGXFWLRQ 7KH DYDLODEOH 12[ UHGXFWLRQ WHFKQRORJLHV FDQ EH FDWHJRUL]HG LQWR RQH RI WKH IROORZn LQJ f %HIRUH &RPEXVWLRQ 1LWURJHQ LV H[WUDFWHG IURP WKH IXHO 7KLV LV UHODWLYHO\ LQHIn IHFWLYH VLQFH PRVW RI WKH QLWURJHQ LQ WKH IRUPDWLRQ RI 12[ FRPHV IURP WKH DLU FRQWDLQLQJ 1f

PAGE 52

f $IWHU &RPEXVWLRQ 12[ LV FKHPLFDOO\ UHGXFHG EHIRUH OHDYLQJ WKH VWDFN 7KLV SURn FHVV LV DOVR H[SHQVLYH UHTXLULQJ KDUGZDUH UHWURILWV f 'XULQJ &RPEXVWLRQ $OWHULQJ IXHO DQG DLU IORZV DQG LQWURGXFLQJ WKHP DW GLIIHUHQW SRLQWV RI WKH IXUQDFH FDQ FUHDWH VHYHUDO ]RQHV ZLWK GLIIHUHQW WHPSHUDWXUHV DQG VWRLFKLRPHWU\ 7KHVH SDUDPHWHUV VLJQLILFDQWO\ HIIHFW WKH UDWH RI 12[ IRUPDWLRQ 7KH IROORZLQJ VHFWLRQ UHYLHZV DYDLODEOH 12[ UHGXFWLRQ VWUDWHJLHV DQG WHFKQRORJLHV IRU FRPEXVWLRQ VRXUFHV f )XHO 6ZLWFKLQJ )XHOERXQG 12[ IRUPDWLRQ LV PRVW HIIHFWLYHO\ UHGXFHG E\ VZLWFKLQJ WR D IXHO ZLWK ORZHU QLWURJHQ FRQWHQW 1R IXHO RLO RU DQRWKHU UHVLGXDO IXHO KDYLQJ D UHODWLYHO\ KLJK QLWURJHQ FRQWHQW FDQ EH UHSODFHG ZLWK 1R IXHO RLO DQRWKHU GLVWLOODWH RLO RU QDWXUDO JDV ZKLFK LV HVVHQWLDOO\ QLWURJHQIUHHf WR UHGXFH 12[ HPLVVLRQV f )OXH *DV 5HFLUFXODWLRQ )*5f )OXH JDV UHFLUFXODWLRQ LQYROYHV H[WUDFWLRQ RI VRPH RI WKH IOXH JDV IURP WKH VWDFN DQG UHFLUFXODWLRQ ZLWK WKH FRPEXVWLRQ DLU VXSSOLHG WR WKH EXUQHUV 7KH SURFHVV UHGXFHV ERWK WKH R[\JHQ FRQFHQWUDWLRQV DW WKH EXUQn HUV DQG WKH WHPSHUDWXUH E\ GLOXWLQJ WKH FRPEXVWLRQ DLU ZLWK IOXH JDV &2 FDQ EHFRPH D VLJQLILFDQW SUREOHP KHUH f /RZ 12[ %XUQHUV ,QVWDOODWLRQ RI EXUQHUV HVSHFLDOO\ GHVLJQHG WR OLPLW 12[ IRUn PDWLRQ FDQ UHGXFH 12[ HPLVVLRQV +LJKHU UHGXFWLRQ HIILFLHQFLHV FDQ EH DFKLHYHG E\ FRPELQLQJ D ORZ 12[ EXUQHU ZLWK )*5 /RZ 12[ EXUQHUV DUH GHVLJQHG WR UHGXFH WKH SHDN IODPH WHPSHUDWXUH E\ LQGXFLQJ UHFLUFXODWLRQ ]RQHV VWDJLQJ FRPEXVWLRQ ]RQHV DQG UHGXFLQJ ORFDO R[\JHQ FRQFHQWUDWLRQV

PAGE 53

f 'HUDWLQJ 6RPH LQGXVWULDO ERLOHUV PD\ EH GHUDWHG WR SURGXFH D UHGXFHG TXDQWLW\ RI VWHDP RU KRW ZDWHU 'HUDWLQJ ZLOO GHFUHDVH WKH IODPH WHPSHUDWXUH ZLWKLQ WKH XQLW UHGXFLQJ IRUPDWLRQ RI WKHUPDO 12[ 'HUDWLQJ FDQ EH DFFRPSOLVKHG E\ UHGXFLQJ WKH ILULQJ UDWH RU E\ LQVWDOOLQJ D SHUPDQHQW UHVWULFWLRQ VXFK DV DQ RULILFH SODWH LQ WKH IXHO OLQH &OHDUO\ WKLV VROXWLRQ ZRXOG KDYH VLJQLILFDQW HFRQRPLF LPSDFW RQ WKH XQLW f 6WHDP RU :DWHU ,QMHFWLRQ %\ LQMHFWLQJ D VPDOO DPRXQW RI ZDWHU RU VWHDP LQWR WKH LPPHGLDWH YLFLQLW\ RI WKH IODPH WKH IODPH WHPSHUDWXUH ZLOO EH ORZHUHG DQG WKH ORFDO R[\JHQ FRQFHQWUDWLRQ UHGXFHG 7KH UHVXOW ZRXOG EH WR GHFUHDVH WKH IRUPDn WLRQ RI WKHUPDO DQG IXHOERXQG 12[ 7KLV SURFHVV JHQHUDOO\ ORZHUV WKH FRPEXVn WLRQ HIILFLHQF\ RI WKH XQLW E\ RQH RU WZR SHUFHQW f 6WDJHG &RPEXVWLRQ (LWKHU DLU RU IXHO LQMHFWLRQ FDQ EH VWDJHG FUHDWLQJ HLWKHU D IXHOULFK ]RQH IROORZHG E\ DQ DLUULFK ]RQH RU DQ DLUULFK ]RQH IROORZHG E\ D IXHOULFK ]RQH $ ORZ 12[ EXUQHU XWLOL]LQJ VWDJHG FRPEXVWLRQ FDQ EH LQVWDOOHG RU WKH IXUQDFH LWVHOI FDQ EH UHWURILWWHG IRU VWDJHG FRPEXVWLRQ f )XHO 5HEXPLQJ 6WDJHG FRPEXVWLRQ FDQ EH DFKLHYHG WKURXJK WKH IXHO UHEXPLQJ SURFHVV $ *DV 5HEXPLQJ =RQH *5=f LV FUHDWHG DERYH WKH SULPDU\ FRPEXVWLRQ ]RQH ,Q WKH *5= DGGLWLRQDO QDWXUDO JDV LV LQMHFWHG FUHDWLQJ D IXHOULFK UHJLRQ ZKHUH K\GURFDUERQ UDGLFDOV UHDFW ZLWK 12[ WR IRUP PROHFXODU QLWURJHQ f 5HGXFHG 2[\JHQ &RQFHQWUDWLRQ 'HFUHDVLQJ H[FHVV DLU UHGXFHV WKH R[\JHQ DYDLOn DEOH LQ WKH FRPEXVWLRQ ]RQH DQG OHQJWKHQV WKH IODPH UHVXOWLQJ LQ D ORZHU KHDW UHOHDVH UDWH SHU XQLW IODPH YROXPH 12[ HPLVVLRQV DUH UHGXFHG LQ DQ DSSUR[L

PAGE 54

PDWHO\ OLQHDU IDVKLRQ ZLWK GHFUHDVLQJ H[FHVV DLU +RZHYHU DV H[FHVV DLU LV UHGXFHG EH\RQG D WKUHVKROG YDOXH FRPEXVWLRQ HIILFLHQF\ ZLOO GHFUHDVH GXH WR LQFRPSOHWH PL[LQJ DQG &2 HPLVVLRQV ZLOO LQFUHDVH f 6HOHFWLYH &DWDO\WLF 5HGXFWLRQ 6&5f 6HOHFWLYH FDWDO\WLF UHGXFWLRQ 6&5f LV D SRVWIRUPDWLRQ 12[ FRQWURO WHFKQRORJ\ WKDW XVHV D FDWDO\VW WR IDFLOLWDWH D FKHPLn FDO UHDFWLRQ EHWZHHQ 12[ DQG DPPRQLD WR SURGXFH QLWURJHQ DQG ZDWHU $Q DPPRQLDDLU RU DPPRQLDVWHDP PL[WXUH LV LQMHFWHG LQWR WKH H[KDXVW JDV ZKLFK WKHQ SDVVHV WKURXJK D FDWDO\VW ZKHUH 12[ LV UHGXFHG 7R RSWLPL]H WKH UHDFWLRQ WKH WHPSHUDWXUH RI WKH H[KDXVW JDV PXVW EH LQ D FHUWDLQ UDQJH ZKHQ LW SDVVHV WKURXJK WKH FDWDO\VW EHG $PRQJ LWV GLVDGYDQWDJHV 6&5 UHTXLUHV DGGLWLRQDO VSDFH IRU WKH FDWDO\VW DQG UHDFWRU YHVVHO DV ZHOO DV DPPRQLD VWRUDJH GLVWULEXn WLRQ DQG LQMHFWLRQ V\VWHP 3UHFLVH FRQWURO RI DPPRQLD LQMHFWLRQ LV FULWLFDO $Q LQDGHTXDWH DPRXQW RI DPPRQLD FDQ UHVXOW LQ XQDFFHSWDEOH KLJK 12[ HPLVVLRQ UDWHV ZKLOH H[FHVV DPPRQLD FDQ OHDG WR DPPRQLD fVOLSf RU WKH YHQWLQJ RI XQGHn VLUDEOH DPPRQLD WR WKH DWPRVSKHUH f 6HOHFWLYH 1RQ&DWDO\WLF 5HGXFWLRQ 61&5f 6HOHFWLYH QRQFDWDO\WLF 12[ UHGXFn WLRQ LQYROYHV LQMHFWLRQ RI D QLWURJHQRXV DJHQW VXFK DV DPPRQLD RU XUHD LQWR WKH IOXH JDV 7KH RSWLPXP LQMHFWLRQ WHPSHUDWXUH ZKHQ XVLQJ DPPRQLD LV GHJUHHV ) DW ZKLFK SHUFHQW 12[ UHPRYDO FDQ EH DSSURDFKHG 7KH RSWLPXP WHPSHUDWXUH UDQJH LV ZLGHU ZKHQ XVLQJ XUHD %HORZ WKH RSWLPXP WHPSHUDWXUH UDQJH DPPRQLD LV IRUPHG DQG DERYH 12[ HPLVVLRQV DFWXDOO\ LQFUHDVH 7KH VXFFHVV RI 12[ UHPRYDO GHSHQGV QRW RQO\ RQ WKH LQMHFWLRQ WHPSHUDWXUH EXW DOVR

PAGE 55

RQ WKH DELOLW\ RI WKH DJHQW WR PL[ VXIILFLHQWO\ ZLWK IOXH JDV )RVVLO)LUHG 3RZHU *HQHUDWLRQ ,Q JHQHUDO &DQDO 8QLW LV D ODUJH IRVVLO IXHO FRPEXVWLRQ HQJLQH )URP DQ DEVWUDFW SHUn VSHFWLYH WKH FRPEXVWLRQ SURFHVV WDNHV LQ DLU DQG IXHO DQG SURGXFHV HQHUJ\ DQG H[KDXVW DV GHVFULEHG E\ f $LU )RVVLO IXHO FRPEXVWLRQ UHTXLUHV DLU RU PRUH VSHFLILFDOO\ WKH R[\JHQ FRQWDLQHG LQ DLU 6XEV\VWHPV ZLWKLQ WKH SODQW PHDVXUH SUHSDUH DQG LQWURGXFH WKLV DLU f )XHO &RPEXVWLRQ DOVR UHTXLUHV IXHO ,Q WKH FDVH RI &DQDO 8QLW WKH IXHO FDQ EH HLWKHU UHVLGXDO RLO OHIWRYHU IURP WKH UHILQLQJ SURFHVVf RU QDWXn UDO JDV &DQDO 8QLW FDQ ILUH RLO RQO\ JDV RQO\ RU D PL[WXUH RI ERWK %RWK IXHOV PXVW EH PHDVXUHG SUHSDUHG DQG LQWURGXFHG WR WKH IXUQDFH f (QHUJ\ 7KH HQHUJ\ UHOHDVHG E\ WKH R[LGDWLRQ RI IRVVLO IXHOV GXULQJ FRPn EXVWLRQ LV XVHG WR PDNH VWHDP 7KH SURSHUWLHV RI ZDWHU VWLOO PDNH LW WKH EHVW FKRLFH ZKHQ FRQYHUWLQJ WKHUPDO HQHUJ\ WR ZRUN &DQDO XVHV WKH UDGLDWLYH DQG FRQYHFWLYH KHDW IURP WKH FRPEXVWLRQ SURFHVV WR WUDQVIRUP XOWUDFOHDQ ZDWHU LQWR VXSHUKHDWHG VWHDP 7KH H[SDQVLRQ RI WKLV VWHDP LV XVHG WR WXUQ D WXUELQH WKDW WXUQV D FRLO LQ D PDJQHWLF ILHOG SURGXFLQJ HOHFWULF SRWHQWLDO 7KH VWHDP KDYLQJ GRQH WKLV ZRUN IORZV WKURXJK RFHDQ ZDWHU ILOOHG FRQGHQVHUV WKDW FRQYHUW LW EDFN WR VXSHUFOHDQ ZDWHU f ([KDXVW 7KH JDVHRXV SURGXFWV RI FRPEXVWLRQ KDYLQJ FRQWULEXWHG PXFK RI WKHLU KHDW FRQWHQW WR WKH SURGXFWLRQ RI VWHDP DUH FOHDQHG HOHFWURVWDWLn FDOO\ DQG HMHFWHG LQWR WKH DWPRVSKHUH 3URFHVV 9DULDEOHV 7KH VSHFLILF SURFHVV YDULDEOHV DV WKH\ DSSO\ WR WKH &DQDO JHQHUDWLQJ XQLW DUH GHVFULEHG LQ PRUH GHWDLO LQ WKH IROORZLQJ VHFWLRQV 7KHVH YDULDEOHV DUH DOVR OLVWHG LQ $SSHQGL[ DQG ZLOO EH UHIHUUHG WR WKURXJKRXW WKLV ZRUN $LU 7KH DLU UHTXLUHG IRU IRVVLO IXHO FRPEXVWLRQ LV SUHSDUHG DQG LQWURGXFHG LQ WZR ZD\V 7ZR ODUJH V\PPHWULFDO )DQV FDOOHG )RUFHG 'UDIW IDQV SXVK DPELHQW DLU WKURXJK D VHULHV RI

PAGE 56

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n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

PAGE 57

7KH )' IDQV DUH WKH SULQFLSDO DLU WKURWWOHV RI WKH %RLOHU DQG VR KDYH D IXQGDPHQWDO HIIHFW RQ QHDUO\ HYHU\ RWKHU V\VWHP )RUFHG GUDIW IDQ LQOHW YDQHV 6LQFH WKH LQOHW YDQHVf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f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

PAGE 58

&RPEXVWLRQ DLU WHPSV 7KH FRPEXVWLRQ DLU WHPSHUDWXUH WDJV UHSUHVHQW WKH WHPSHUDWXUH RI WKH LQFRPLQJ DLU DIWHU WKH )' )DQV 7KH WHPSHUDWXUH RI WKLV DLU LV D GLUHFW UHVXOW RI HQHUJ\ DGGHG WR DPELHQW DLU E\ WKH *O\FRO $LU 3UHKHDWHU *$+f DQG WKH &RPEXVWLRQ $LU 3UHKHDWHU &$),f 6LQFH GHQVLW\ LV D IXQFWLRQ RI WHPSHUDWXUH WKH WHPSHUDWXUH RI WKLV DLU FDQ LPSDFW WKH FRPEXVWLRQ SURFHVV WKDW LV VHQVLWLYH WR WKH 2[\JHQ FRQWHQW RI DLU DV ZHOO DV WKH RSHUDWLRQ RI RWKHU YROn XPHWULF V\VWHPV OLNH WKH ,QGXFHG 'UDIW ,'f )DQV ,W DOVR KDV D SULPDU\ LPSDFW RQ H[KDXVW JDV WHPSHUDWXUH DQG UHVXOWDQW VWDFN JDV YHORFLW\ 3ULPDU\ DLU VKURXGV 7KH 3ULPDU\ $LU 3$f 6KURXGV UHSUHVHQW WKH FLUFXODU DUWLFXODWLQJ YHQWV WKDW VXUURXQG WKH LQGLYLGXDO EXUQHU RULILFHV 7KHVH DUH FORVHVW WR WKH IXHO JXQ FRQFHQWULFDOO\ LQVLGH RI WKH 6HFRQGDU\ $LU 6$f 6KURXGV 7KH\ DUH UHVSRQVLEOH IRU VXSSO\LQJ SULPDU\ FRPEXVWLRQ DLU WR WKH IODPH IURQW 7KHVH WDJV UHSUHVHQW DFWXDWRU SRVLWLRQV 7KH 3$ VKURXGV DUH FRQWUROOHG E\ WKH %XUQHU 0DQDJHPHQW 6\VWHP %06f DQG WKH\ PRYH DV D JURXS IURP PLQLPXP SRVLWLRQ b RSHQ WR SURYLGH FRROLQJ DLUf WRZDUG RSHQ DV ORDG LQFUHDVHV 7KH VLJQDO WKDW FRQWUROV WKHP LV FDOOHG WKH 3ULPDU\ $LU 0DVWHU 'HPDQG 3$0'f 6HSDUDWH 3$0' VLJQDOV H[LVW IRU IXHO JDV SULPDU\ DLU GHPDQG DQG IRU IXHO RLO SULPDU\ DLU GHPDQG (DFK VSHFLILF EXUQHU HIIHFWLYHO\ OLVWHQV WR WKH FXUUHQW IXHO VWDWH +DYn LQJ UHFHLYHG WKLV VLJQDO HDFK EXUQHUfV RZQ 3$ VKURXGV UHVSRQGV WR WKH 3$0' LQ DFFRUn GDQFH ZLWK RQH RI WZR IXQFWLRQV WKDW DUH XQLTXH WR LW D %XUQHU 3ULPDU\ $LU 6KURXG )XQFWLRQ IRU RLO RSHUDWLRQ DQG D %XUQHU 3ULPDU\ $LU 6KURXG )XQFWLRQ IRU JDV RSHUDWLRQ 7KH FRUUHFW XQLTXH ORFDO VKURXG IXQFWLRQ LV FKDQJHG DFFRUGLQJ WR WKH FRUUHFW PDVWHU VLJQDO GHSHQGLQJ RQ WKH IXHO VWDWH RI WKH EXUQHU 7KHVH EXUQHU DQG IXHO VSHFLILF UHVSRQVH IXQF

PAGE 59

WLRQV ZHUH VHW XS WR JLYH URXJKO\ DSSURSULDWH DLU IORZ WR FRPEXVWLRQ DW DOO ORDG SRLQWV DQG IXHO VWDWHV EDVHG RQ WKH DLU IORZ LQKHUHQW WR WKH IXUQDFH $VLGH IURP QRUPDO RSHUDWLRQ WKH 3$ VKURXGV FDQ EH ELDVHG IURP WKH IXHO VSHFLILF PDVn WHU VLJQDO RU RQ DQ LQGLYLGXDO EDVLV IURP WKHLU UHVSHFWLYH XQLTXH IXQFWLRQV 6HFRQGDU\ DLU VKURXGV 7KH VHFRQGDU\ DLU VKURXG WDJV UHSUHVHQW WKH EURDGFDVW DFWXDWRU SRVLWLRQV RI WKH VHFRQG RXWHU FRQFHQWULF VHW RI FLUFXODU DUWLFXODWLQJ YHQWV WKDW VXUURXQG WKH LQGLYLGXDO EXUQHU RULn ILFHV 7KH ILUVW IXQFWLRQ RI WKH 6HFRQGDU\ $LU 6$f 6KURXGV LV WR LQWURGXFH FRPEXVWLRQ DLU WR WKH IODPH IURQW IROORZLQJ ORDG 7KHLU VHFRQG IXQFWLRQ LV WR EDODQFH ZLQGER[ SUHVVXUH DQG WKHUHIRUH WRWDO DLUIORZ DJDLQVW WKH DFWXDWLRQ RI WKH 2YHUILUH $LU 3RUWV DQG WKH 3$ VKURXGV 7KH 6$ 6KURXGV KDYH D PDVWHU VLJQDO DJDLQVW ZKLFK D PDVWHU ELDV FDQ EH VHW ,Q DGGLn WLRQ WKH\ KDYH LQGLYLGXDO DFWXDWLQJ IXQFWLRQV DQG LQGLYLGXDO ELDVHV WKDW FDQ EH VHW DJDLQVW WKHVH LQGLYLGXDO IXQFWLRQV 2YHU ILUH DLU SRUWV 7KH 2YHUILUH $LU 2)$f SRUWV DUH UHFWDQJXODU ORXYHUHG SRUWV WKDW SDVV FRPEXVWLRQ DLU IURP WKH :LQGER[ WR WKH )XUQDFH DERYH WKH WRS EXUQHU OHYHO ,Q GRLQJ WKLV WKH\ UHR[\JHQn DWH WKH R[\JHQ GHSOHWHG IODPH IURQW 7KH WDJV WKHPVHOYHV UHSUHVHQW WKH SRVLWLRQV EURDGFDVW IURP WKH DFWXDWRUV WKDW FRQWURO WKH DUWLFXODWLQJ ORXYHUV 7KH 2)$ SRUWV ZHUH LQVWDOOHG DV D SDUW RI WKH ORZ 12[ UHWURILW RI 7KH )RUQH\ ORZ 12[ EXUQHU V\VWHP LV GHVLJQHG WR EXP PRUH FRROO\ DQG LQFRPSOHWHO\ WKDQ QRUPDO 12[ IRUPDWLRQ KDV EHHQ SRVLWLYHO\ OLQNHG ZLWK WLPH H[SRVXUH WR KLJKHU WHPSHUDWXUHV $IWHU SDUWLDO FRPEXVWLRQ KDV WDNHQ SODFH ORZ LQ WKH IODPH IURQW H[WUD R[\JHQ ULFK FRP

PAGE 60

EXVWLRQ DLU LV LQWURGXFHG WKURXJK WKH 2$ SRUWV WR FRPSOHWH WKH SURFHVV ,Q WKLV ZD\ WKH ORZ 12[ EXUQHU V\VWHP VWDJHV RIIVWRLFKLRPHWULF FRPEXVWLRQ WR PDQDJH FRPEXVWLRQ SURGXFWV 7KH 2$ SRUW DFWXDWRUV UHFHLYH WKHLU PDVWHU VLJQDO IURP ORDG 7KLV VLJQDO FDQ EH ELDVHG (DFK DFWXDWRUfV UHVSRQVH LV EDVHG RQ D XQLTXH IXQFWLRQ WKDW ZDV SDUDPHWULFDOO\ GHWHUn PLQHG LQ FRQFHUW ZLWK WKH 3ULPDU\ DQG 6HFRQGDU\ $LU 6KURXGV GXULQJ LQVWDOODWLRQ WR JLYH EHVW DLUIORZ WR FRPEXVWLRQ DW DOO ORDG SRLQWV $LU SUHKHDWHU WHPSV 7KHVH UHSUHVHQW WKH WHPSHUDWXUH RI WKH H[KDXVW JDVVHV HQWHULQJ DQG OHDYLQJ WKH OMXQJ VWURP FRPEXVWLRQ DLU KHDW H[FKDQJHU 7KH OMXQJVWURP LV D ODUJH IW GLDf URWDWLQJ ZKHHO DUUDQJHG SHUSHQGLFXODU WR WKH JDV IORZ ,W LV KDOI HQFORVHG E\ WKH H[KDXVW GXFWV DQG KDOI HQFORVHG E\ IUHVK DLU GXFWV $V WKLV ZKHHO VORZO\ URWDWHV KHDW LV DEVRUEHG E\ D JLYHQ DUHD RI WKH ZKHHO H[SRVHG WR H[KDXVW JDV 7KH DEVRUEHG KHDW LV WKHQ LPSDUWHG WR WKH LQFRPLQJ DLU ZKLOH WKDW VDPH VHFWLRQ WUDYHUVHV WKH IUHVK DLU GXFW (ODERUDWH VHDOV DQG SUHVVXUL]HG VHDOLQJ DLU NHHS WKH WZR JDVVHV IURP PLQJOLQJ DFURVV WKLV GHYLFH 7KH DLU SUHKHDWHU WDJV DUH VRPHZKDW UHGXQGDQW 7KH f,Q 7HPSVf UHSUHVHQW WKH WHPSHUn DWXUH RI WKH JDV RQ LWV ZD\ LQ ZKLOH WKH f2XW 7HPSVf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

PAGE 61

:LQGER[ DQG IXUQDFH 7KH ZLQGER[ LV DQ HQFORVHG YROXPH WKDW VXUURXQGV WKH ZDLVW RI WKH IXUQDFH DQG WKH EXUQHU RSHQLQJV 3UHKHDWHG R[\JHQ ULFK DLU LV SUHVVXUL]HG LQ WKLV YROXPH E\ WKH )' IDQV )URP KHUH WKLV DLU FDQ SDVV RQO\ LQWR WKH IXUQDFH DQG RQO\ WKURXJK YDQHV WKDW VXUURXQG WKH EXUQHU RSHQLQJV FDOOHG SULPDU\ DQG VHFRQGDU\ VLU VKURXGV RU WKURXJK WKH RYHUILUH DLU SRUWV DERYH WKH EXUQHUV &DQDO 8QLW LV D EDODQFHG GUDIW IXUQDFH ZKLFK PHDQV WKDW DLU IORZ WKURXJK WKH IXUQDFH LV FRQWUROOHG DURXQG D GHVLUHG IXUQDFH SUHVVXUH E\ ERWK SXVKLQJ DQG SXOOLQJ IDQ V\VWHPV 7KH SXVKLQJ IDQV DUH WKH )' IDQV ZKLOH WKH SXOOLQJ IDQV DUH WKH LQGXFHG GUDIW IDQV 7KH )' IDQV KDYH WKH SULPDU\ UHVSRQVLELOLW\ RI JHWWLQJ WKH FRPEXVWLRQ ]RQH DOO WKH R[\JHQ LW UHTXLUHV 7KH LQWURGXFWLRQ RI WKLV SUHVVXUL]HG DLU LV DFFRPSOLVKHG QRW RQO\ E\ SRVLWLYHO\ SUHVVXUL]LQJ WKH ZLQGER[ EXW DOVR E\ QHJDWLYHO\ SUHVVXUL]LQJ WKH IXUn QDFH :LWK WKH ZLQGER[ GULYHQ WR D SRVLWLYH SUHVVXUH DQG WKH IXUQDFH NHSW DW D IL[HG UHODn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

PAGE 62

WKURWWOHG EDFN 7UDQVLHQW FKDQJHV LQ WKH ZLQGER[ WR IXUQDFH SUHVVXUH GLIIHUHQWLDO FDQ DOVR SURGXFH DXWRPDWHG FKDQJHV LQ WKH )' DQG ,' IDQ IORZV :LQG%R[ SUHVVXUH 7KLV WDJV UHSUHVHQWV WKH SRVLWLRQDO DYHUDJH ZLQGER[ DLU SUHVVXUH ,W LV FRQWUROOHG DURXQG LQZF )XUQDFH SUHVVXUH 7KLV WDJ UHSUHVHQWV WKH DFWXDO IXUQDFH DLU SUHVVXUH )XHO 7KH )XHO UHTXLUHG IRU &RPEXVWLRQ PD\ EH HLWKHU )XHO 2LO RU 1DWXUDO *DV ,Q ERWK FDVHV WKH IXHO LV WDNHQ IURP VWRUDJH ILOWHUHG KHDWHG WR JUHDWHU RU OHVVHU GHJUHH SUHVVXUn L]HG DQG LQMHFWHG ,Q WKH FDVH RI )XHO 2LO WKH WHPSHUDWXUH UHTXLUHG WR DFKLHYH D SXPS DEOH FRQVLVWHQF\ LV XVXDOO\ DURXQG GHJUHHV 1DWXUDO *DV FRPHV IURP KLJK SUHVVXUH WUDQVPLVVLRQ OLQHV DQG RQFH VWHSSHG GRZQ WR XVDEOH SUHVVXUH LV ZDUPHG XS WR DURXQG GHJUHHV ) %RWK IXHOV DUH WKHQ SUHVVXUL]HG LQ WKHLU UHVSHFWLYH KHDGHUV ,W LV IURP WKHVH KHDGHUV WKDW EXUQHUV ZKHQ WKH\ DUH OLW WDS WKHLU IXHO %XUQHUVRQIXHO 7KHVH WDJV UHSUHVHQW WKH UHDGLQJV RI DQ DUUD\ RI DLU FRROHG RSWLFDO IODPH VFDQQHUV ORFDWHG LQ WKH IXUQDFH LWVHOI WKDW REVHUYH WKH UHVSHFWLYH EXUQHU IODPHV 6LQFH HDFK EXUQHU FDQ ILUH HLWKHU QDWXUDO JDV RU IXHO RLO $ VFDQQHU FDOLEUDWHG IRU HDFK IXHO VSHFLILF IODPH LV SHUPDQHQWO\ DVVLJQHG WR HDFK EXUQHU $OWKRXJK WKHVH VFDQQHUV DUH DQDORJ GHYLFHV WKHLU SULPDU\ IXQFWLRQ LV WR FRQILUP WKDW WKH IODPH HPDQDWLQJ IURP HDFK OLW EXUQHU LV RI D WKUHVKn ROG TXDOLW\ ,I WKH IODPH WKH\ DUH PRQLWRULQJ LV QRW RI D WKUHVKROG TXDOLW\ WKH VFDQQHU KDV

PAGE 63

WKH ZLOO WR GHFODUH D 0DVWHU )XHO 7ULS DQG FXW RII DOO IXHO WR WKH IXUQDFH 7KLV LV WR SUHYHQW WKH LQWURGXFWLRQ RI XQEXPHG IXHO WR WKH IXUQDFH 7KHVH DUH DQDORJ GHYLFHV EXW EHFDXVH WKH\ DUH FDOLEUDWHG ZLWK WKH VLQJOH SXUSRVH RI HLWKHU SRVLWLYHO\ RU QHJDWLYHO\ FRQILUPLQJ WKLV WKUHVKROG WKH\ HVVHQWLDOO\ UHDG HLWKHU RU 7KLV VSHFLILF VHW UHSUHVHQWV WKH IODPH TXDOLW\ RI LWV EXUQHU LI WKDW EXUQHU LV RQ QDWXUDO JDV %XUQHU FHOOV $ t % 01 JDV IODPH 7KHVH DUH WKH VLJQDOV IRU JDV IODPH VWDWXV IRU HDFK EXUQHU %XUQHU FHOOV $ t % 01 RLO IODPH 7KHVH DUH WKH VLJQDOV IRU RLO IODPH VWDWXV IRU HDFK EXUQHU )XHO W\SH $V WKH %RLOHU 0DVWHU UHTXHVW LQFUHDVHG RXWSXW %78V DUH UHTXHVWHG IURP WKH )XHO 6XSn SO\ 6\VWHPV $V D GHIDXOW WKLV UHTXHVW LV GLYLGHG HYHQO\ LQ SURSRUWLRQ WR EXUQHUV LQ VHUYLFH HDFK RI ZKLFK KDYH %78 FRQWHQW SHU XQLW RI IXHO VHWWLQJV 7KH WRWDO %78V HQWHULQJ WKH IXUn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

PAGE 64

3UHVVXUH &KDQJHV LQ WKH QXPEHU RI EXUQHUV OLW FDQ YDU\ WKH IXHO RLO SUHVVXUH LQ WKH KHDGHU EHWZHHQ DQG 36,* )XHO 7HPS )LUHG PXVW EH DW OHDVW WKH WHPS UHTXLUHG IRU SXPSDELOLW\ ZKLFK LV VSHFLILF WR WKH YLVFRVLW\ RI WKH IXHO RLO EHLQJ XVHG 1DWXUDO JDV ,Q D IDVKLRQ VLPLODU WR WKH IXHO RLO LQWURGXFWLRQ V\VWHP WKH XQLW PDVWHU GHPDQG UHTXHVWV %78V IURP WKH JDV V\VWHP )XHO JDV IURP WKH SLSHOLQH LV VWHSSHG GRZQ WR RSHUDWn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

PAGE 65

(QHUJ\ 'XULQJ RSHUDWLRQ DW &DQDO 8QLW IHHGZDWHU SUHVVXUL]HG E\ D ODUJH SDUDVLWLF WXUELQH GULYHQ SXPS LV FLUFXODWHG WKURXJK VHULHV RI SUHKHDWHUV DQG WKHQ WKURXJK WKH YHU\ ZDOOV RI WKH IXUQDFH 'XULQJ WKLV SDVVDJH LW LV FRQYHUWHG WR VWHDP 7KLV VWHDP LV WKHQ FROOHFWHG LQ D SUHVVXUH YHVVHO FDOOHG D 6WHDP 'UXP ORFDWHG DW WKH WRS RI WKH ERLOHU ZKHUH LW LV fGULHGf )URP WKH 6WHDP 'UXP WKLV GU\ VDWXUDWHG VWHDP LV SDVVHG WKURXJK UDGLDWRU OLNH 3ULPDU\ 6XSHUKHDWHU DQG 6HFRQGDU\ 6XSHUKHDWHUV WKDW KDQJ DW WKH WRS RI WKH IXUQDFH ZKHUH FRQYHFn WLYH DQG UDGLDWLYH KHDW WUDQVIHU RFFXUV )URP WKH RXWOHW RI WKH 6HFRQGDU\ 6XSHUKHDWHU WKH VWHDP JRHV GLUHFWO\ WR WKH +LJK 3UHVVXUH LQOHW RI WKH 7XUELQH 8QLW LV D VLQJOH UHKHDW ERLOHU ZKLFK PHDQV WKDW WKH H[KDXVW IURP WKH KLJK SUHVVXUH WXUELQH LQVWHDG RI EHLQJ FRQn GHQVHG LV SDVVHG EDFN WR WKH ERLOHU DQG UH VXSHUKHDWHG 7KLV UH VXSHUKHDWHG VWHDP WKHQ WXUQV WKH ,QWHUPHGLDWH DQG /RZ 7XUELQH 6WDJHV $WWHPSHUDWLQJ VSUD\V LQMHFW FRRO IHHGZDn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

PAGE 66

7KLV WDJ UHSUHVHQWV WKH DFWXDO LQVWDQWDQHRXV XQLW RXWSXW LQ XQLWV RI SRZHU +HDW UDWH 7KLV LV D VLPSOH FDOFXODWHG WDJ UHSUHVHQWLQJ WKH VXP RI %78V IORZLQJ LQWR FRPEXVWLRQ IURP RLO DQG JDV FRPELQHG GLYLGHG E\ WKH DPRXQW RI SRZHU FUHDWHG ,W FDQ VKRZ WKH UHODn WLYH HIILFLHQF\ RI FRPEXVWLRQVWHDPSRZHU V\VWHP LQ DQ HQHUJ\ LQ YV HQHUJ\ RXW UHODWLRQn VKLS $V ORDG LQFUHDVHV KHDW UDWH GHFUHDVHV GXH WR WKH WKHUPDO SURSHUWLHV RI WKH VWHDP ORRS 0DLQ VWHDP 7KH PDLQ VWHDP WHPSHUDWXUH LQ FRQFHUW ZLWK WKH WKURWWOH SUHVVXUH LV UHODWHG YLD VWHDP WDEOHV WR YROXPH HQWKDOS\ DQG HQWURS\ DQG GHVFULEHV WKH RXWSXW VWDWH RI WKH VWHDP JHQHUDWn LQJ V\VWHP 8QLW LV D VOLGLQJ WKURWWOH XQLW FDSDEOH RI PRGXODWHG VWHDP WHPS RXWSXW DFURVV GLIIHUHQW WKURWWOH YDOYH FRQILJXUDWLRQV 6WHDP RXWSXW LV HVVHQWLDOO\ FRQWUROOHG E\ IORZ $V WKH XQLW UDPSV XS LQ ORDG PRUH VWHDP LV JHQHUDWHG IURP LQFUHDVHG FRPEXVWLRQ 6WHDP WHPSHUDWXUH LV KHOG URXJKO\f VWHDG\ YLD PRGXODWLRQ RI IORZ WKURXJK WKH WXUELQH WKURWWOH YDOYHV ZKLFK DUH VHTXHQWLDOO\ RSHQHG 2QFH WKH XQLW UHDFKHV D FHUWDLQ OHYHO RI RXWSXW #0:f DOO WKURWWOH YDOYHV DUH VHW LQ WKH IXOO\ RSHQ SRVLWLRQ DQG VWHDP IORZ LV PRGXn ODWHG E\ FRQWLQXLQJ WR LQFUHDVLQJ VWHDP RXWSXW WKURXJK FRPEXVWLRQ WKURWWOLQJ $W DOO OHYHOV RI RXWSXW 6WHDP WHPSHUDWXUH LV FRQWUROOHG DURXQG GHJUHHV ) IRU RSWLPXP WXUELQH RSHUDWLRQ 7HPSHUDWXUH 7KHVH WDJV UHSUHVHQW WKH WHPSHUDWXUH RI VXSHUKHDWHG VWHDP DV LW H[LWV WKH VHFRQGDU\ VXSHUKHDWHU KHDGHU DQG KHDGV WR WKH KLJK SUHVVXUH WXUELQH WKURWWOH YDOYHV

PAGE 67

$WWHPSHUDWLRQ VSUD\ 7KHVH UHSUHVHQW WKH DPRXQW RI FRRO IHHGZDWHU WKDW LV VSUD\HG LQWR PDLQ VWHDP EHWZHHQ WKH SULPDU\ DQG VHFRQGDU\ VXSHUKHDWHUV WR FRQWURO WKH WHPSHUDWXUH RI WKH VWHDP DW WKH VHFn RQGDU\ VXSHUKHDWHU RXWOHW WR WKH WXUELQH 8 UHSUHVHQWV ILQH FRQWURO 7KLV YDOYH UHVSRQGV DXWRPDWLFDOO\ DQG LQ DQDORJ IDVKn LRQ WR DOO FKDQJHV LQ VWHDP WHPSHUDWXUH DW WKH VHFRQGDU\ VXSHUKHDWHU RXWOHW 8 UHSUHn VHQWV EXON FRQWURO ,W UHVSRQGV RQO\ WR FKDQJHV LQ 66+ RXWOHW WHPS WKDW DUH H[FHHG SUHVHW GHDGEDQG 7KHVH FRDUVH DQG ILQH FRROLQJ FRQWUROV DUH FRPELQHG WR GDPSHQ DQG FRQWURO VWHDP RXWOHW WHPS DJDLQVW RVFLOODWLRQV RU LPEDODQFHV LQKHUHQW LQ WKH VWHDP V\VWHP $OO GHVXSHUKHDWLQJ VSUD\V UHFHLYH WKHLU YROXPH RI IHHGZDWHU IURP WRWDO IHHGZDWHU IORZ 5HKHDW VWHDP ([KDXVW IURP WKH KLJK SUHVVXUH WXUELQH VWDJH LV F\FOHG EDFN WR WKH IXUQDFH YLD WKH UHKHDW VWHDP ORRS ZKHUH LW LV VSUD\HG WKHQ UHLQWURGXFHG WR KHDW H[FKDQJH LQ WKH UHKHDW VXSHUKHDWHU 7KURXJK WKH UHKHDW VXSHUKHDWHU WKLV VWHDP LV EURXJKW EDFN XS WR O222GHJ) DQG SLVD XSRQ ZKLFK LW LV VHQW WR WKH LQWHUPHGLDWH VWDJH RI WKH WXUELQH ([KDXVW IURP WKH LQWHUPHGLDWH VWDJH WXUELQH IORZV WR WKH ORZ SUHVVXUH WXUELQH VWDJH 7HPSHUDWXUH 7KLV WHPSHUDWXUH UHSUHVHQW WKH WHPSHUDWXUH RI UH VXSHUKHDWHG VWHDP DV LW KHDGV WR WKH LQWHUPHGLDWH WXUELQH VWDJH LQOHW $WWHPSHUDWLRQ VSUD\V 7KHVH VSUD\V IXQFWLRQ OLNH WKH VXSHUKHDWHU VSUD\V 7KH\ LQMHFW UHODWLYHO\ FRRO IHHGZDn WHU LQWR WKH UHKHDW VWHDP DIWHU LW KDV EHHQ H[WUDFWHG IURP WKH WXUELQH DQG EHIRUH LW LV

PAGE 68

UHKHDWHG 7KH\ IXQFWLRQ WR FRQWURO WKH WHPSHUDWXUH RI WKH VWHDP DW WKH RXWOHW RI WKH UHKHDW VXSHUKHDWHU 8QOLNH WKH VXSHUKHDW GHVXSHUKHDWHUV WKHVH VSUD\V GR QRW KDYH VHSDUDWH ILQH DQG FRDUVH FRQWURO IXQFWLRQV )XUQDFH PHWDO WHPSV 7KHVH WDJV UHSUHVHQW DQ DUUD\ RI WKHUPRFRXSOHV LQVWDOOHG RQ WKH YHUWLFDO OHJV RI WKH SHQGDQW VXSHUKHDWHUV (VSHFLDOO\ LQ JDV EXUQLQJ WKH ILUH VLGH PDWHULDO WHPSHUDWXUH RI WKHVH KHDW H[FKDQJHUV FDQ EHFRPH SUREOHPDWLF 8QLW KDV DQ HVSHFLDOO\ ODUJH DUHD RI VXSHUn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

PAGE 69

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n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n LW\ RI FRPEXVWLRQ .H\ SDUDPHWHUV LQFOXGH KRZ PXFK R[\JHQ KDV EHHQ OHIW E\ WKH FRPn EXVWLRQ SURFHVV DQG KRZ PXFK &2 KDV EHHQ FUHDWHG 7KH ULFKQHVV RU OHDQQHVV RI FRPEXVWLRQ LV GLUHFWO\ HYLGHQW

PAGE 70

)OXH JDV &2 7KHVH WDJV UHSUHVHQW WKH &2 FRQWDLQHG LQ H[KDXVW JDVVHV DV PHDVXUHG LQ WKH VLGH $ 8f DQG VLGH % 8f IXUQDFH RXWOHWV WR WKH H[KDXVW GXFWV MXVW DIWHU WKH HFRQRn PL]HU 7KHVH DUH SRLQW PHDVXUHV RI &2 LQ D YHU\ ODUJH GXFW DQG PD\ QRW FDSWXUH H[DFW &2 FRQWHQW 7KH\ DOVR GLVSOD\ H[WUHPH VLGH WR VLGH ELDV ZLWK VLGH % VKRZLQJ KLJKHU &2 FRQn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n WLRQ WKHVH WDJV DUH XVHG E\ &DQDO DV D SDUW RI WKH &(0 12[ FDOFXODWLRQ 7HPSV 7KH WHPSHUDWXUH RI WKH DLU EHLQJ IRUFHG WKURXJK WKH ERLOHU DW &DQDO 8QLW LPSDFWV DQG UHSUHVHQWV PDQ\ SURFHVV SDUDPHWHUV IURP FRPEXVWLRQ TXDOLW\ WR KHDW WUDQVIHU GLVWULEXn WLRQ WR LQGXFHG GUDIW IDQ RXWSXW ,W DOVR LV FRQWURO UHIHUHQFH IRU WKH WHPSHUDWXUH DQG YHORFn LW\ RI H[KDXVW OHDYLQJ WKH VWDFN

PAGE 71

6WDFN 7KH &(0 &RQWLQXRXV (PLVVLRQV 0RQLWRULQJ 8QLWf FRQVLVWV RI DQ DUUD\ RI H[WUDFWLRQ JDV DQDO\]HUV LQ D FRPSXWHU URRP DW WKH EDVH RI &DQDOfV IW VWDFN 7KH SLWRWV RI WKHVH DQDO\]HUV VQLII PL[HG H[KDXVW IURP WKH WRS RI WKH IRRW ZLGH 8QLW IOXH 7KH VSHFLILF DPRXQWV RI FHUWDLQ FRPSRXQGV PHDVXUHG LQ WKLV JDV DUH HQWHUHG LQWR D GDWDEDVH 7KLV GDWDn EDVH VHUYHV DV D ELQGLQJ OHJDO KLVWRU\ RI &DQDOV HQYLURQPHQWDO FRPSOLDQFH (DFK YLRODWLRQ RI HPLVVLRQV OLPLWV SODFHG RQ FHUWDLQ FRPSRXQGV OLNH 12[ DQG &2 LV UHFRUGHG ,I WKH XQLW LV LQ GDQJHU RI EUHDNLQJ LWV DOORZHG GDLO\ DYHUDJH RXWSXW RI WKHVH UHJXODWHG SROOXWDQWV PHDVXUHG IURP PLGQLJKW WR PLGQLJKW DOO VWHSV PXVW EH WDNHQ WR UHJDLQ FRPSOLDQFH LQFOXGn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n RPHWULF FRPEXVWLRQ 8QIRUWXQDWHO\ VXFK ULFK DQG FRRO LQFRPSOHWHf FRPEXVWLRQ LQKHUHQWO\ SURGXFHV LQFUHDVHG &2 12[ 7KLV WDJ LV FDOFXODWHG XVLQJ D UHJXODWRU\ DSSURYHG PHWKRG DQG LV XVHG WR UHSUHVHQWV WKH SRXQGV RI 12[ SURGXFHG E\ &DQDO 8QLW SHU PLOOLRQ %78V

PAGE 72

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n DWXUH WR RWKHU SURSHUWLHV RI FRPEXVWLRQ VWDFN WHPSHUDWXUH FDQ EH DQ LPSRUWDQW DQG WULFN\ FRQWURO SRLQW 6LQFH WKH KHDW ORVVHV WR WKH H[KDXVW WKURXJK WKH H[KDXVW GXFWV SUHFLSLWDWRUV DQG LQGXFHG GUDIW IDQV DUH IL[HG IOXHJDV WHPSHUDWXUH DOVR UHSUHVHQWV VWDFN WHPSHUDWXUH EXW ZLWKRXW JDV SDWK WUDYHO GHOD\

PAGE 73

&+$37(5 %2,/(5 237,0,=$7,21 7KH PRVW HIILFLHQW PHWKRG IRU UHGXFLQJ 12[ HPLVVLRQV LV FOHDUO\ GXULQJ WKH FRPEXVn WLRQ SURFHVV >@ $V SUHVHQWHG LQ 6HFWLRQ 12[ VLPSO\ FKDQJLQJ WKH FRPEXVWLRQ WHPSHUDWXUH DQG IXHODLU GLVWULEXWLRQ FDQ GUDPDWLFDOO\ DIIHFW 12[ HPLVVLRQV 7KH FRPEXVn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f ,I WKH ERLOHU RSHUDWHV LQ DQ R[\JHQULFK HQYLURQPHQW LH ZLWK XQQHFHVVDU\ H[FHVV DLU ERLOHU HIILFLHQF\ ZLOO GHFUHDVH GXH WR WKH ORVV RI VHQVLEOH KHDW XS WKH VWDFN 12[ HPLVVLRQV ZLOO LQFUHDVH FRQFXUUHQWO\ f ,I WKH ERLOHU RSHUDWHV LQ D IXHOULFK HQYLURQPHQW LH ZLWK LQVXIILFLHQW DLU ERLOHU HIILFLHQF\ ZLOO GHFUHDVH GXH WR WKH ORVV RI XQEXPHG IXHO ,Q DGGLWLRQ LQVXIILFLHQW

PAGE 74

DLU OHDGV WR &2 IRUPDWLRQ ZKLFK FDXVHV VODJJLQJ DQG ZDWHU ZDOO FRUURVLRQ XOWLn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n UHVHQWHG DV D OLJKW JUD\ EDQG LQ )LJXUH 7KLV YDULDELOLW\ GHWHUPLQHV WKH DPRXQW RI DJJUHn JDWH DLU UHTXLUHG WR HQVXUH WKDW DOO UHJLRQV LQVLGH WKH ERLOHU DYRLG IXHOULFK FRPEXVWLRQ )LJXUH &RPEXVWLRQ HPLVVLRQV FKDUDFWHULVWLF YHUVXV DLU IORZ

PAGE 75

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n WXUH 6XFK ERLOHU WXQLQJ LPSURYHV XQLW SHUIRUPDQFH EXW GRHV QRW EHJLQ WR JHQHUDWH WKH VDYn LQJV DFKLHYDEOH WKURXJK LPSURYHG FRQWURO

PAGE 76

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f KRZ ZHOO PL[HG WKH IXHO DQG DLU DUH LQ WKH IXUQDFH f KRZ LQGLYLGXDO VHWSRLQWV FDQ EH XVHG WR LPSURYH WKLV PL[LQJ DQG f FRQGLWLRQV ZKLFK OHDG WR VODJJLQJ DQG ZDWHUZDOO FRUURVLRQ 7KH &2 PHDVXUHPHQW VHUYHV DV D NH\ VDIHW\ FRQVWUDLQW ZKHQ RSWLPL]LQJ WKH ERLOHU %\ FRQWUROOLQJ WR &2 OHYHOV WKH ERLOHU FDQ EH RSWLPL]HG ZLWKRXW FRPSURPLVLQJ VDIHW\ PDUn JLQV ,PSURYHG DLU DQG IXHO GLVWULEXWLRQ ZLOO PHUHO\ WLJKWHQ WKH &2 FXUYH UHVXOWLQJ LQ LPSURYHG HIILFLHQF\ DQG ORZHU 12[

PAGE 77

&+$37(5 &21752/ '(6,*16 7KLV UHVHDUFK LQYHVWLJDWHV WKH DSSOLFDELOLW\ RI QHXURFRQWURO WHFKQLTXHV WR FRPSOH[ SURn FHVV FRQWURO SUREOHPV DQG GHYHORSV D PHWKRGRORJ\ IRU LPSOHPHQWLQJ WKHP 7RZDUGV WKLV HQG WKLV ZRUN ZLOO GHYHORS VHYHUDO GHWDLOHG QHXUDO QHWZRUNEDVHG FRQWURO GHVLJQV DQG DSSO\ WKHP WR WKH UHGXFWLRQ RI 12[ DQG WKH PDLQWHQDQFH RI DFFHSWDEOH &2 OHYHOV LQ HOHFn WULF SRZHU SODQWV 6XEVHTXHQW VHFWLRQV LPSOHPHQWDWLRQ WKHVH FRQWURO GHVLJQV DQG XVH RXU 12[ FDVH VWXG\ WR FRPSDUH DQG FRQWUDVW WKHP 7KH FRQWURO PHWKRGRORJ\ ZLOO EH SUHVHQWHG DV IROORZV f $ PHWKRGRORJ\ IRU FDWHJRUL]LQJ NH\ SURFHVV YDULDEOHV LQWR JURXSV WKDW DUH UHTXLUHG IRU DOO FRQWURO GHVLJQV f $ PHWKRGRORJ\ IRU IRUPDOO\ VWDWLQJ WKH FRQWURO RSWLPL]DWLRQ REMHFWLYHV DQG RSHUDWLQJ FRQVWUDLQWV XVLQJ WKH DIRUHPHQWLRQHG YDULDEOH GHILQLWLRQV f 3HUIRUPDQFH FULWHULD E\ ZKLFK WKH YDULRXV FRQWURO GHVLJQV ZLOO EH MXGJHG EDVHG RQ WKHVH IRUPDO REMHFWLYHV DQG FRQVWUDLQWV f )RXU IRUPDO FRQWURO GHVLJQV ZLWK H[SOLFLWO\ DFFRXQW IRU VWDWH YDULDEOH GHSHQGHQFLHV 9DULDEOH 'HILQLWLRQV :KHQ GHVLJQLQJ D FRQWUROOHU IRU ODUJHVFDOH LQGXVWULDO SURFHVVHV WKHUH DUH D ODUJH QXPn EHU RI YDULDEOHV WR EH FRQVLGHUHG 7KH SK\VLFDO SURFHVVHV DUH W\SLFDOO\ FRQVLGHUHG WR KDYH LQSXWV GLVWXUEDQFHV VWDWHV DQG RXWSXWV 7KH IROORZLQJ YDULDEOH GHILQLWLRQV DUH SURSRVHG DV D PHWKRGRORJ\ IRU FDWHJRUL]LQJ DOO SURFHVV YDULDEOHV LQWR VXEVHWV WKHVH VXEVHWV ZLOO SURYH XVHIXO ZKHQ GHVLJQLQJ FRQWUROOHUV LQ JHQHUDO

PAGE 78

f 0DQLSXODWHG 9DULDEOHV 09Vf SURFHVV LQSXWV ZKLFK KDYH EHHQ VHOHFWHG IRU RXU FRQWUROOHU WR PDQLSXODWH 7KH 09V VKRXOG EH LQGHSHQGHQW RI RQH DQRWKHU LH PDQLSXODWLQJ RQH ZLOO QRW FDXVH D FKDQJH LQ DQ\ RI WKH RWKn HUV f 'LVWXUEDQFH 9DULDEOHV '9Vf SURFHVV LQSXWV RU GLVWXUEDQFHV WKDW DIIHFW WKH VWDWH RU RXWSXW RI WKH SURFHVV EXW ZH HLWKHU FDQQRW RU KDYH FKRVHQ QRW WR PDQLSXODWH 7KH '9 VKRXOG EH LQGHSHQGHQW RI ERWK HDFK RWKHU DQG WKH 09V f &RQWURO 9DULDEOHV &9Vf WKH SURFHVV VWDWH RU RXWSXW YDULDEOHV WKDW WKH FRQWUROOHU ZLOO EH GHVLJQHG WR FRQWURO 7KH &9V VKRXOG EH D IXQFWLRQ RI WKH 09V DQG '9V RU WKHUH LV OLWWOH KRSH RI WKH FRQWUROOHU EHLQJ DEOH WR FRQWURO WKHP f 6WDWH 9DULDEOHV 69Vf SURFHVV VWDWH YDULDEOHV ZKLFK DUH D IXQFWLRQ RI WKH 09V DQGRU '9V WKDW DIIHFW WKH &9V $OWHUQDWLYHO\ WKH 69V PD\ EH SURFHVV RXWSXW YDULDEOHV WKDW KDYH QRW EHHQ VHOHFWHG IRU FRQWURO EXW QHHG WR EH FRQVLGHUHG DV FRQVWUDLQWV 1RWLFH WKDW WKH 09 '9 69 DQG &9 GHILQLWLRQV FDWHJRUL]H WKH SURFHVV ORJLFDOO\ DQG QRW SK\VLFDOO\ 7KHVH GHILQLWLRQV GLYLGH YDULDEOHV EDVHG RQ KRZ WKH FRQWUROOHU ZLOO EH FRQn ILJXUHG UDWKHU WKDQ KRZ WKH SK\VLFDO SURFHVV LV FRQILJXUHG 7KH 09V ZLOO DOZD\V EH SURn FHVV LQSXWV LH FDQ EH PDQLSXODWHG E\ RSHUDWRUV EXW WKH '9V FDQ FRQWDLQ ERWK SURFHVV LQSXWV DQG GLVWXUEDQFHV GHSHQGLQJ RQ ZKLFK LQSXWV DUH EHLQJ PDQLSXODWHG /LNHZLVH 69V DQG &9V FDQ HDFK FRQVLVW RI DQ\ FRPELQDWLRQ RI SURFHVV VWDWHV DQGRU RXWSXWV EDVHG RQ ZKLFK ZLOO XOWLPDWHO\ EH FRQWUROOHG 1RWDWLRQ 7KH FDWHJRUL]DWLRQ RI YDULDEOHV LQWR &9V 69V '9V DQG 09V ZLOO EH XVHG H[WHQVLYHO\ WKURXJKRXW WKLV ZRUN DQG LV FRQFHSWXDOO\ FRQVLVn WHQW ZLWK WKH OLWHUDWXUH RQ RSWLPL]DWLRQ DQG FRQWURO >@ 2SWLPL]DWLRQ 2EMHFWLYHV 7KH FRQWURO REMHFWLYH LV WR ORZHU 12[ )RUPDOO\ WKLV REMHFWLYH QHHGV WR EH VWDWHG DV DQ REMHFWLYH IXQFWLRQ IRU RSWLPL]DWLRQ 6LQFH VHYHUDO RI WKH FRQWUROOHUV GHYHORSHG KHUH DUH WUD

PAGE 79

MHFWRU\ PXOWLVWDJHf FRQWUROOHUV WKLV REMHFWLYH IXQFWLRQ ZLOO EH D IXQFWLRQ RI WLPH &RQn VLGHU WKH VLQJOH FRQWURO YDULDEOH 12[Wf H DV WKH PHDVXUHG YDOXH RI 12[ DW WLPH W $Q RSWLPDO FRQWURO REMHFWLYH ZLWK IL[HG WHUPLQDO WLPH 7 IRU PLQLPL]LQJ 12[Wf RYHU WKH LQWHUYDO W H W W 7@ FDQ EH JLYHQ E\ f W W ,Q JHQHUDO WKHUH ZLOO EH PRUH WKDQ RQH &9 ,I DOO &9V DUH WR KDYH HTXDO LPSDFW RQ WKLV REMHFWLYH IXQFWLRQ WKHQ WZR HIIHFWV ZLOO KDYH WR EH UHPRYHG IURP WKH RSWLPL]DWLRQ REMHFn WLYH f WKH HIIHFW RI SRZHU GLIIHUHQFHV EHWZHHQ WKHVH &9V DQG f WKH FXUUHQW YDOXH RI HDFK &9 7KH IROORZLQJ REMHFWLYH IXQFWLRQ H[WHQGV f WR PXOWLSOH FRQWURO YDULDEOHV f W WQL ? ZKHUH n LV WKH QXPEHU RI &9V DQG S LV D SULRULW\ ZHLJKWLQJ IDFWRU DQG =[f [ S@ $VVXPLQJ WKDW RXU FRQWUROOHU LV GHVLJQHG WR PLQLPL]H -? IRU S WKH &9 FYL ZLOO EH PLQLPL]HG RYHU WKH WUDMHFWRU\ ZKLOH VHWWLQJ S ZLOO PD[LPL]H WKH RXWSXW (TXDWLRQ f FRQVLGHUV WKH FDVH ZKHUH &9V DUH WR EH PD[LPL]HG RU PLQLPL]HG ,Q JHQHUDO WKH JRDO LV WR GHVLJQ D FRQWUROOHU FDSDEOH RI PDLQWDLQLQJ D FRQWURO VHWSRLQW $ JHQn HUDOL]HG RSWLPL]DWLRQ REMHFWLYH LV WKHUHIRUH SUHVHQWHG DV f

PAGE 80

ZKHUH 'c LV D GHVLUDELOLW\ IXQFWLRQ WKDW FDQ EH WDLORUHG IRU HDFK &9 WR 'fD[ =FY2f=&Yff f 'nIQ =FYWff=FYfRU f 'fS M=FYf=Yff f 2SHUDWLQJ &RQVWUDLQWV &RQVWUDLQWV ZLOO EH XVHG WR HQVXUH WKDW WKH RSWLPL]HU SURGXFHV D IHDVLEOH VROXWLRQ %\ IHDVLEOH ZH PHDQ f WKH 09 PRYHV FDQ EH PDGH DQG f WKDW ZKHQ WKHVH 09V DUH DSSOLHG WKH SODQW ZLOO HQG XS LQ D GHVLUDEOH VWDWH )HDVLEOH VROXWLRQV ZLOO EH JXDUDQWHHG E\ GHVLJQn LQJ FRQWUROOHUV ZKLFK DUH DEOH WR PDLQWDLQ 09 DQG 69 FRQVWUDLQWV 0DQLSXODWHG 9DULDEOH &RQVWUDLQWV 7R HQVXUH WKDW WKH 09 PRYHV FDQ EH PDGH WKH FRQWUROOHUV ZLOO PDLQWDLQ VLPSOH UDQJH FRQVWUDLQWV $ UDQJH FRQVWUDLQW FRQVLVWV RI WKH XSSHU DQG ORZHU OLPLWV WKDW DQ 09 ZLOO EH DOORZHG WR PRYH )RUPDOO\ WKH UDQJH FRQVWUDLQW IRU 09 PYc ZLOO EH JLYHQ E\ fZY U fPLQ fPD[ &L /& f &L f f ZKHUH &f LV WKH 09fV DEVROXWH PLQLPXP DQG &!fDn[ LV LWV PD[LPXP &RQWUROOHUV ZLOO EH W7? UHTXLUHG WR SURYLGH DQ RSWLPDO 09 WUDMHFWRU\ ^PY Wf`W 4 VXFK WKDW 9 W H &IY f

PAGE 81

6WDWH 9DULDEOH &RQVWUDLQWV 6LPLODUO\ WR HQVXUH WKDW FRQWUROOHUV GULYH WKH SODQW WR D GHVLUDEOH VWDWH 69 FRQVWUDLQWV ZLOO DOVR KDYH WR EH DGGUHVVHG )RUPDOO\ FRQWUROOHUV ZLOO EH UHTXLUHG WR SURYLGH RSWLPDO WR 7 09 WUDMHFWRULHV WKDW UHVXOW LQ 69 WUDMHFWRULHV ^VY Wf` O VXFK WKDW 69 9 f 3HQDOW\ )XQFWLRQV (DFK FRQWURO GHVLJQ FRQVLGHUHG ZLOO HPSOR\ DQ RSWLPL]DWLRQ DOJRULWKP GXULQJ VRPH SKDVH RI LWV GHYHORSPHQW 6RPH RSWLPL]DWLRQ DOJRULWKPV DUH DEOH WR GHDO ZLWK FRQVWUDLQWV GLUHFWO\ LH JLYHQ NQRZOHGJH RI WKH FRQVWUDLQWV WKH\ FDQ HQVXUH D IHDVLEOH VROXWLRQV 2WKn HUV KRZHYHU ZLOO KDYH WR WUHDW FRQVWUDLQWV LQGLUHFWO\ E\ DGGUHVVLQJ WKHP ZLWK WKH REMHFn WLYH IXQFWLRQ 7KH PRVW FRPPRQ PHWKRG IRU DGGUHVVLQJ RSHUDWLQJ FRQVWUDLQWV LQ DQ REMHFWLYH IXQFWLRQ DUH WKURXJK WKH XVH RI SHQDOW\ IXQFWLRQV >@ )RU H[DPSOH 69 FRQn VWUDLQWV FDQ EH VWDWHG DV SHQDOW\ IXQFWLRQV RI WKH IRUP 0Y FYf  fPD[ fQ VYc & f VYc! & IIQLQML VLL VY & f VY & 9 HOVH f *HQHUDOL]LQJ WKH SHQDOW\ IXQFWLRQ WR PXOWLSOH 69 FRQVWUDLQWV GLIIHUHQFHV LQ WKH HQHUJ\ RI WKH UHVSHFWLYH VLJQDOV ZLOO RQFH DJDLQ KDYH WR EH QRUPDOL]HG RXW 7KHVH HIIHFWV FDQ EH FRPSHQVDWHG IRU XVLQJ D JHQHUDOL]HG SHQDOW\ IXQFWLRQ RI WKH IRUP P VYL FQ ?]^VYLf =&IQD[ff VY &PL =V9Mf =^&IQLQf< VYL&PL f 9 HOVH

PAGE 82

*LYHQ D VHW RI 1n 69 FRQVWUDLQWV DQ RSWLPL]HU PD\ VDWLVI\ WKHVH FRQVWUDLQWV E\ DSSHQGLQJ WKHLU UHVSHFWLYH SHQDOW\ IXQFWLRQV WR LWV FULWHULRQ I -neS";VYW&IOfn! f ZKHUH Srn DOORZV FRQVWUDLQWV WR EH LQGLYLGXDOO\ SULRULWL]HG 6LPLODUO\ ERWK 09 FRQn VWUDLQWV FDQ EH DSSHQGHG WR WKH RSWLPL]HUfV FULWHULRQ E\ GHILQLQJ WKH SHQDOW\ IXQFWLRQV ^;PYc &c 1RWH WKDW LPSOHPHQWLQJ FRQVWUDLQWV ZLWK SHQDOW\ IXQFWLRQV ZLOO QRW JXDUDQWHH WKDW WKH FRQVWUDLQWV DUH PHW SUHFLVHO\ ,I WKH FRQVWUDLQWV DUH SURSHUO\ SULRULWL]HG UHODWLYH WR WKH RSWLn PL]DWLRQ REMHFWLYHV KRZHYHU WKHVH FRQVWUDLQWV DUH HDVLO\ PDLQWDLQHG ZLWKLQ WKH GHVLUHG OHYHO RI DFFXUDF\ 3HUIRUPDQFH &ULWHULD )RU WKH FDVH VWXG\ FRQWUROOHUV ZLOO EH MXGJHG EDVHG RQ WKHLU DELOLW\ WR ORZHU 12[ ZKLOH PDLQWDLQLQJ GHVLUHG &2 HPLVVLRQV 7R WKLV HQG VXEVHTXHQW VHFWLRQV ZLOO PHDVXUH WKH SHUn IRUPDQFH RI FRQWUROOHUV DV D SODQW RSHUDWRU PRYHV 09V DFFRUGLQJ WR WKHLU FRQWURO ODZV &RPSDULQJ FRQWUROOHU SHUIRUPDQFH KRZHYHU ZLOO SURYH D GLIILFXOW WDVN VLQFH WKH RSHUDWRU FDQ RQO\ WDNH WKH DGYLFH IURP RQH FRQWUROOHU DW D WLPH DQG WKH SODQW LV FRQVWDQWO\ FKDQJLQJ VWDWH $OWKRXJK WKH FRQWUROOHUV PD\ EH DEOH WR GHDO ZLWK QRQVWHDG\VWDWH FRQGLWLRQV LW ZLOO EH QHDUO\ LPSRVVLEOH WR VHSDUDWH WKH SURFHVV UHVSRQVHV WR WKH VWDWH FKDQJHV YHUVXV WKH FRQn WURO DFWLRQ

PAGE 83

)XUWKHU FRPSOLFDWLQJ PDWWHUV ZKLOH RQHWLPH WHVWV ZLOO SURYLGH XVHIXO UHVXOWV ZLWK ZKLFK WR MXGJH WKH FRQWUROOHUV WKH\ DUH QRW WKH RQO\ FULWHULD 7KH FRQWUROOHUV VWXGLHG ZLOO EH MXGJHG E\ WKH IROORZLQJ FULWHULD f $ELOLW\ WR FRQWURO 12[ DQG &2 f $ELOLW\ RI WKH RSHUDWRUV WR SHUIRUP WKH UHFRPPHQGHG 09 PRYHV f )OH[LELOLW\ ZLWK UHVSHFW WR FKDQJLQJ SHUIRUPDQFH REMHFWLYHV DQG RSHUDWn LQJ FRQVWUDLQWV f $ELOLW\ WR GHDO ZLWK FKDQJLQJ RSHUDWLQJ VWDWHV HJ ORDG FKDQJHV &RQWUROOHU 'HVLJQV )RXU FRQWUROOHU GHVLJQV ZLOO EH GHYHORSHG 7KH FRQWUROOHU GHVLJQV FRQVLGHUHG IDOO LQWR WKH EURDG FDWHJRULHV RI f 0RGHO3UHGLFWLYH &RQWURO f 0RGHO,QYHUVH &RQWURO f 0RGHO%DVHG 'LUHFW &RQWURO 7KHUH DUH KRZHYHU QR VWDQGDUG UHFLSHV IRU EXLOGLQJ WKHVH FRQWUROOHUV 7KH ILHOG LV VWLOO LPPDWXUH DQG QHXURFRQWURO GHVLJQV SUHVHQWHG LQ WKH OLWHUDWXUH WHQG WR EH DG KRF 7KLV ZRUN VHHNV WR QRW RQO\ GHYHORS DQG WHVW IRXU QHXURFRQWURO GHVLJQV EXW DOVR WR GHYHORS D JHQHUDOL]HG PHWKRGRORJ\ IRU LPSOHPHQWLQJ FRQWURO GHVLJQV EHORQJLQJ WR WKH DERYH DEVWUDFW FDWHJRULHV (DFK FRQWUROOHU PXVW EH DEOH WR GHDO ZLWK WKH 09 DQG 69 FRQVWUDLQWV DQG ZLOO EH MXGJHG E\ WKH SHUIRUPDQFH FULWHULD GHVFULEHG DERYH )RU HDFK RI WKH FRQWURO GHVLJQV FRQVLGHUHG WKHUH DUH WZR GLVWLQFW SKDVHV LQ WKH LPSOHn PHQWDWLRQ f 2IIOLQH WUDLQLQJ f 2QOLQH FRQWURO

PAGE 84

6WHDG\6WDWH 2SWLPL]HU 7KH VLPSOHVW DQG PRVW SUHYDOHQW QHXURFRQWUROOHU LQ WKH OLWHUDWXUH LV WKH VWHDG\VWDWH RSWLPL]HU >@>@> @ 7KLV FRQWUROOHU EHORQJV WR WKH PRGHOSUHGLFWLYH FRQWURO IDPLO\ 0RGHOSUHGLFWLYH FRQWURO 03&f LV QRW QHZ WR FRPPHUFLDO DSSOLFDWLRQV LQ WKH SURFHVV FRQn WURO LQGXVWU\ 7KH DGYDQFH SURSRVHG KHUH LV WKH DSSOLFDWLRQ RI QHXUDO QHWZRUN UHIHUHQFH PRGHOV ZLWKLQ WKLV FRQWUROV PHWKRGRORJ\ 7KH FRQFHSW RI 03& LV VWUDLJKW IRUZDUG FRPELQH D PRGHO IRU WKH SURFHVV ZLWK DQ RSWLn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f 6WHDG\6WDWH 69 0RGHO VY VV690RGHOPY GYf f 6WHDG\6WDWH &9 0RGHO FY VV&90RGHOPY GY VYf

PAGE 85

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sr^If XVLQJ WKH 69 DQG &9 UHIHUHQFH PRGHOV GHYHORSHG GXULQJ PRGHO WUDLQ PY LQJ 7KH RSWLPL]HU VWDUWV ZLWK WKH FXUUHQW YDOXH RI WKH 09V LQ Yr PY XVHV WKH 69 ? M_^ PRGHO WR HVWLPDWH WKH FXUUHQW 69V VY ZKLFK DUH WKHQ XVHG DORQJ ZLWK WKH FXUUHQW YDOXH

PAGE 86

RI WKH '9V GY WR HVWLPDWH WKH FXUUHQW &9V FYr 7KH RSWLPL]HU WKHQ LWHUDWLYHO\ XSGDWHV f f f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n VLWLYLWLHV GI GHYr IURP ZKLFK WKH &9 0RGHO LV DEOH WR FDOFXODWH 69 VHQVLWLYLWLHV GI GVYr DQG SDUWLDO 09 VHQVLWLYLWLHV GIn GPYr IURP ZKLFK WKH 69 PRGHO FDOFXODWHV WKH UHPDLQLQJ SDUWLDO 09 VHQVLWLYLWLHV GIn GPYr DQG ILQDOO\ WKH RSWLPL]HU LV DEOH WR XSGDWH LWV RSWLPDO 09 HVWLPDWH XVLQJ WKH 09 JUDGLHQW RI GI B GIY GIY GPYr GPYr GPYr f

PAGE 87

7KLV LV UHDOO\ MXVW WKH EDFNSURSDJDWLRQ RI EDFNSURSDJDWLRQV DND PRUH IXQ ZLWK WKH FKDLQ UXOH 6HYHUDO RSWLPL]DWLRQ PHWKRGV ZHUH WHVWHG IRU WKH RSWLPL]HU DORQJ ZLWK YDULRXV WHFKn QLTXHV IRU GHDOLQJ ZLWK WKH FRQVWUDLQWV 7KH PRVW HIIHFWLYH FRPELQDWLRQ LGHQWLILHG ZDV WR XVH WKH XQFRQVWUDLQHG FRQMXJDWH JUDGLHQWV PHWKRG LQ FRPELQDWLRQ ZLWK DQ REMHFWLYH IXQFn WLRQ ZKLFK LQFOXGHG WKH 69 DQG 09 FRQVWUDLQW SHQDOW\ IXQFWLRQV DW DW DU I -U7SU:fME,S$r2 e"7AUFQ mf r mf L r DU L ? ZKHUH WKH 09 69 '9 DQG &9 YDULDEOH VHWV DORQJ ZLWK WKHLU FRUUHVSRQGLQJ FRQVWUDLQWV DUH GHILQHG LQ 6HFWLRQ )LQDO 9DULDEOH 6HWV DQG DOO SULRULWLHV KDYH EHHQ VHW WR 7KH GHWDLOV RI WKH FRQMXJDWH JUDGLHQWV PHWKRG ZLOO EH SUHVHQWHG LQ 6HFWLRQ /HDUQLQJ $OJRn ULWKP 3HQDOW\ IXQFWLRQ FDQ QHJDWLYHO\ LPSDFW WKH SHUIRUPDQFH RI D GHVFHQWEDVHG RSWLPL]HU E\ DGGLQJ FRPSOH[LW\ WR WKH SHUIRUPDQFH VXUIDFH KDYLQJ OLWWOH WR GR ZLWK WKH XQGHUO\LQJ SUREOHP 7KLV LV SDUWLFXODUO\ WUXH ZKHQ WKH FRQVWUDLQHG YDULDEOHV OLH RXWVLGH RI WKHLU FRQn VWUDLQHG YDOXHV )RU WKH 69 FRQVWUDLQWV WKHUH LV QR FKRLFH EXW WR XVH SHQDOW\ IXQFWLRQV IRU FRQVWUDLQWV )RU 09V KRZHYHU WKHUH DUH DOWHUQDWLYHV EHFDXVH WKH 09V DOZD\V VWDUW DW WKHLU FXUUHQW YDOXHV ZKLFK DUH DOZD\V ZLWKLQ WKH FRQVWUDLQWV +HQFH WKHUH LV OLWWOH WR QR RYHUKHDG WR XVLQJ 09 FRQVWUDLQWV IRU RXU RQOLQH RSWLPL]HU ,W WHUPV RI WKH SHUIRUPDQFH VXUIDFH WKH FRQVWUDLQWV FDQ EH WKRXJKW RI DV SODFLQJ D JXDUGUDLO RQ ERWK VLGHV RI RXU FXUn UHQW SRVLWLRQ LQ ZHLJKWVSDFH DORQJ RXU SDWK ZKLOH KDYLQJ OLWWOH LPSDFW RQ WKH ORFDO WRSRJn UDSK\ RI WKH URDG

PAGE 88

6WHDG\6WDWH 0RGHO,QYHUVH &RQWUROOHU 7KH QH[W FRQWUROOHU GHVLJQ EHORQJV WR WKH PRGHOLQYHUVH FRQWURO 0,&f IDPLO\ &RQn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f DQG LQYHUVH09 ,09f PRGHOV 2QFH DJDLQ WKH GHWDLOV IRU KRZ WR WUDLQ WKHVH PRGHOV ZLOO EH FRYHUHG LQ &KDSWHU 7KH PRGHO GHILQLWLRQV UHTXLUHG E\ WKH VWHDG\VWDWH PRGHOLQYHUVH FRQWUROOHU DUH f 6WHDG\6WDWH ,69 0RGHO VY VV,690RGHOFY GYf f 6WHDG\6WDWH ,09 0RGHO PY VV,090RGHOFY VY GYf

PAGE 89

)LJXUH 2IIOLQH WUDLQLQJ DQG UHWXQLQJ FRQILJXUDWLRQ IRU PRGHOLQYHUVH FRQWUROOHU $QDORJRXV WR RXU 03& FRQWUROOHU WZR PRGHOV KDYH EHHQ GHYHORSHG ZKLFK ZKHQ FRPn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n SOH[LW\ LV WKDW WKH H[DFW YDOXH IRU WKH ORZHVW DFKLHYDEOH 12[ IURP WKH FRQWUROOHU IRU D JLYHQ VHW RI FRQGLWLRQV LV QRW NQRZQ $QRWKHU FRPSOLFDWLRQ ZLWK 0,& LV KRZ WR GHDO ZLWK FRQVWUDLQWV ,I RQH DSSOLHV D WDUJHW &9 WR WKH LQSXW RI WKH LQYHUVHPRGHO LW ZLOO SUHGLFW D VHW RI LQSXWV ZKLFK LW EHOLHYHV ZRXOG KDYH DFKLHYHG WKLV WDUJHW 7KH SUREOHP LV WKDW WKH PRGHO GRHV QRW XQGHUVWDQG WKH 09 RU 69 FRQVWUDLQWV DQG LI RQH RI WKH LQSXWV LW SUHGLFWV IDOOV RXW

PAGE 90

VLGH WKHVH FRQVWUDLQWV WKH FRQWUROOHU FDQ QRW SURYLGH WKH UHTXLUHG VHWSRLQWV 7KLV LV DQDORn JRXV WR WKH SUREOHP IDFHG ZLWK 69 RU &9 FRQVWUDLQWV IRU 03& 7KH LPSOHPHQWDWLRQ RXWOLQHG LQ )LJXUH XVHV DQ RSWLPL]HU LQ RUGHU WR RYHUFRPH ERWK RI WKHVH LVVXHV &OHDUO\ &9 FRQVWUDLQWV DUH VWUDLJKWIRUZDUG )LJXUH 2QOLQH FRQWURO FRQILJXUDWLRQ IRU PRGHOLQYHUVH FRQWUROOHU 7KH 0,& FRQWUROOHU XVHV DQ RSWLPL]HU WR FDOFXODWH $UJ0LQA^I@ XVLQJ WKH ,69 DQG &9 ,09 UHIHUHQFH PRGHOV GHYHORSHG GXULQJ PRGHO WUDLQLQJ 7KH RSWLPL]HU VWDUWV ZLWK WKH FXUn UHQW YDOXH RI WKH &9V FYr FY XVHV WKH ,69 PRGHO WR HVWLPDWH WKH FXUUHQW 69V VYr ZKLFK DUH WKHQ XVHG DORQJ ZLWK WKH FXUUHQW YDOXH RI WKH '9V GY E\ WKH ,09 PRGHO WR f§A HVWLPDWH WKH FXUUHQW 09V PY 7KH RSWLPL]HU WKHQ LWHUDWLYHO\ XSGDWHV LWV HVWLPDWH IRU WKH RSWLPDO 09V PYr WR PLQLPL]H LWV REMHFWLYH IXQFWLRQ I 2QFH DJDLQ ERWK GLUHFW DQG GHVFHQWEDVHG RSWLPL]DWLRQ FDQ EH XVHG IRU 0,& DQG RQFH DJDLQ D FRQMXJDWH JUDGLHQWVEDVHG RSWLPL]HU ZDV VHOHFWHG 'HVFHQWEDVHG RSWLPL]DWLRQ LV SRVVLEOH EHFDXVH WKH ,69 DQG ,09 PRGHOV DUH FDSDEOH YLD EDFNSURSDJDWLRQ RI FDOFXODWn LQJ WKH JUDGLHQW RI I ZLWK UHVSHFW WR WKHLU LQSXWV LH WKHLU LQSXWV VHQVLWLYLWLHV JLYHQ WKH

PAGE 91

VHQVLWLYLWLHV DW WKHLU RXWSXWV ,Q WKLV PDQQHU WKH RSWLPL]HU FDOFXODWHV WKH 09 VHQVLWLYLWLHV GI ,GPYr IURP ZKLFK WKH ,09 0RGHO LV DEOH WR FDOFXODWH 69 VHQVLWLYLWLHV GIGVYr DQG SDUWLDO &9 VHQVLWLYLWLHV GMfYGFYr IURP ZKLFK WKH ,69 PRGHO FDOFXODWHV WKH UHPDLQLQJ SDUWLDO &9 VHQVLWLYLWLHV G-rnGHYr DQG ILQDOO\ WKH RSWLPL]HU LV DEOH WR XSGDWH LW RSWLPDO &9 HVWLPDWH XVLQJ WKH &9 JUDGLHQW RI GI GIVY GIOY DaAr GHY GHY GHY f 7KH RSWLPL]HUfV REMHFWLYH IXQFWLRQ ZKLFK LQFOXGHV WKH 69 DQG 09 FRQVWUDLQW SHQDOW\ IXQFWLRQV LV WKH VDPH REMHFWLYH IXQFWLRQ XVHG E\ RXU VWHDG\VWDWH RSWLPL]HU 7KH RQO\ GLIn IHUHQFH LV KRZ WKH VHQVLWLYLWLHV IORZ WKURXJK WKH V\VWHP DV RXWOLQHG DERYH '\QDPLF 0RGHO3UHGLFWLYH &RQWUROOHU 7KH VWHDG\VWDWH RSWLPL]HU FRQVLGHUHG DERYH LV D PHPEHU RI WKH 03& IDPLO\ 7KH YDVW PDMRULW\ RI 03& DSSOLFDWLRQV XVH PRGHOV ZKLFK DUH ILUVWSULQFLSOHV EDVHG >@ 6LQFH LW LV QRW SRVVLEOH WR EXLOG DQ DFFXUDWH ILUVWSULQFLSOHV PRGHO RI 12[ D QHZ VWHDG\VWDWH RSWLn PL]HU IRU 03& XVLQJ QHXUDO QHWZRUN PRGHOV ZDV GHYHORSHG 7KH YDVW PDMRULW\ RI 03& DSSOLFDWLRQV DUH G\QDPLF KRZHYHU 7KH VWHDG\VWDWH RSWLPL]HU RQO\ FRQVLGHUV WKH HIIHFW WKDW 09 FKDQJHV ZLOO KDYH RQ WKH XQLW LQ VWHDG\VWDWH FRQGLWLRQV 7KLV VHFWLRQ GHYHORSV D G\QDPLF QHXUDO QHWZRUNEDVHG 03& FRQWUROOHU 7KH PDLQ GLIn IHUHQFHV EHWZHHQ WKLV FRQWUROOHU DQG RXU VWHDG\VWDWH RSWLPL]HU LV WKDW LW f 8QGHUVWDQGV WKH G\QDPLFV RI WKH SURFHVV f 3URYLGHV D WUDMHFWRU\ RI 09 VHWSRLQWV GHVLJQHG WR RSWLPL]H WKH SDWK RI WKH XQLW LQWR WKH IXWXUH UDWKHU WKDQ D RSWLPDO VWHDG\VWDWH SRVLWLRQ ,Q RWKHU ZRUGV WKH FRQWUROOHU QRW RQO\ FRQVLGHUV ZKHUH \RXU JRLQJ EXW KRZ \RXfOO JHW WKHUH

PAGE 92

7KH FRQFHSW EHKLQG WKLV FRQWUROOHUf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n XUH +HUH WKH 69 DQG &9 PRGHOV DUH SHUIRUPLQJ VLQJOHVWDJH SUHGLFWLRQ ZLWK UHVSHFW WR WKH 09V DQG '9V QRWLFH WKDW WKH &9 PRGHO XVHV WKH FXUUHQW YDOXH RI WKH 69V VYW f 7KH UHDVRQV IRU WKLV FRQILJXUDWLRQ ZLOO EHFRPH FOHDU ZKHQ ZH FRQVLGHU WKH RQOLQH FRQWURO LPSOHPHQWDWLRQ LQ WKH QH[W VHFWLRQ )LJXUH 2IIOLQH WUDLQLQJ DQG UHWXQLQJ FRQILJXUDWLRQ IRU VWHDG\VWDWH RSWLPL]HU

PAGE 93

5HIHU WR &KDSWHU IRU GHWDLOV RQ WUDLQLQJ WKH G\QDPLF 69 DQG &9 UHIHUHQFH PRGHOV XVHG E\ WKH G\QDPLF 03& FRQWUROOHU )RU QRZ ZH VLPSO\ VWDWH WKH PRGHO GHILQLWLRQV UHTXLUHG E\ WKH G\QDPLF 03& FRQWUROOHU f '\QDPLF 69 0RGHO VYW f G690RGHOPYWf GYWff f '\QDPLF &9 0RGHO FYWOf G&90RGHOPYWf GYWf VYW ff 2QOLQH FRQWURO 7KH RQOLQH FRQWURO FRQILJXUDWLRQ IROORZV D VLPLODU FRQILJXUDWLRQ WR WKH VWHDG\VWDWH RSWLPL]HU SUHVHQWHG LQ )LJXUH +HUH D G\QDPLF RSWLPL]HU LV UHTXLUHG KRZHYHU 7KH RSWLn PL]HU FDOFXODWHV $UJ0LQ ^Wf` XVLQJ WKH G\QDPLF 69 DQG &9 PRGHOV GHYHORSHG PYrWf GXULQJ PRGHO WUDLQLQJ 7KH VWHDG\VWDWH RSWLPL]HU XVHG DQ DSSOLFDWLRQ RI WKH FKDLQ UXOH IRU RUGHUHG SDUWLDO GHULYDWLYHV ZKLFK KDV EHHQ FRLQHG fEDFNSURSDJDWLRQf >@ )URP WKH SHUn VSHFWLYH RI WKH FKDLQ UXOH RXU QHZ RSWLPL]HU LV LGHQWLFDO DQG RQO\ WKH FULWHULRQ FKDQJHV )URP WKH SHUVSHFWLYH RI WKH OLWHUDWXUH WKLV DOJRULWKP KDV EHHQ FRLQHG fEDFNSURSDJDWLRQ WKURXJK WLPHf >@>@ 7KH RSWLPL]HU VWDUWV ZLWK WKH FXUUHQW YDOXH RI WKH 09V PYrWf PYWf XVHV WKH 69 A f_ PRGHO WR HVWLPDWH WKH UHVXOWLQJ 69V VY W f ZKLFK DUH WKHQ XVHG DORQJ ZLWK WKH FXUn UHQW YDOXH RI WKH '9V GYWf WR HVWLPDWH WKH UHVXOWLQJ &9V FYr W f 1RWLFH WKDW HDFK HVWLPDWH FDQ UHO\ RQ ERWK SUHVHQW DQG SDVW YDOXHV RI WKH LQSXWV 7KH RSWLPL]HU ZLOO WKHQ UHSHDW WKLV SURFHVV RYHU WKH WLPH LQWHUYDO W H W4 W 7@ WR SURGXFH WKH 09 69 DQG &9 W7f§ W7 W7 WUDMHFWRULHV ^PY Wf`W W ^VY ` L DQG ^FY Lf` W O UHVSHFWLYHO\

PAGE 94

7KH REMHFWLYH IXQFWLRQ ZKLFK LQFOXGHG WKH 69 DQG 09 FRQVWUDLQW SHQDOW\ IXQFWLRQV FDQ QRZ EH FDOFXODWHG $Wff ,Q 7 U 1rY 1IY 1 3!rf7Y S"Y ZZ R[ SUY0Yr R fP -9nf 9  f ZKHUH WKH 09 69 '9 DQG &9 YDULDEOH VHWV DORQJ ZLWK WKHLU FRUUHVSRQGLQJ FRQVWUDLQWV DUH GHILQHG LQ 6HFWLRQ )LQDO 9DULDEOH 6HWV DQG DOO SULRULWLHV KDYH EHHQ VHW WR 7KH RSWLPL]HU WKHQ LWHUDWLYHO\ XSGDWHV LWV HVWLPDWH IRU WKH RSWLPDO 09 WUDMHFWRULHV !r W7? ^PY Wf`W WJ WR PLQLPL]H LWV REMHFWLYH IXQFWLRQ M f (DFK VWHS LQ WKH LWHUDWLRQ SHUn IRUPV WKH IROORZLQJ VWDUWLQJ ZLWK W W4 7 DQG LWHUDWLQJ GRZQ WR W W ILUVW WKH RSWLPL]HU FDOFXODWHV WKH &9 VHQVLWLYLWLHV GIWfGHYr Wf IURP ZKLFK WKH &9 0RGHO LV DEOH WR FDOFXODWH 69 VHQVLWLYLWLHV GIWfGVYr Wf DQG SDUWLDO 09 VHQVLWLYLWLHV GIYWfGPYrW f IURP ZKLFK WKH 69 PRGHO FDOFXODWHV WKH UHPDLQLQJ SDUWLDO 09 VHQn VLWLYLWLHV GI?WfGPYrW f DQG ILQDOO\ WKH 09 VHQVLWLYLW\ DW WLPH W f§ FDQ EH FDOFXn ODWHG DV GIWf GI?Wf GI?Wf GPYrW f GPYrf§ f GPYrW?f f 2QFH WKH EDFNZDUG SDVV LV FRPSOHWH WKH RSWLPL]HU LV QRZ DEOH WR XSGDWH LWfV RSWLPDO WD7a 09 WUDMHFWRU\ HVWLPDWH ^PY Wf`W W XVLQJ WKH 09 JUDGLHQW WUDMHFWRU\ VDR UU GPY W WD f

PAGE 95

1RWLFH WKDW WKH VHQVLWLYLWLHV DW WLPH W GHSHQG RQ WKH VHQVLWLYLWLHV LQ WKH IXWXUH 7KLV LV EHFDXVH WKH PRGHOV YDULDEOHV DW WLPH W GHSHQG RQ WKH YDULDEOHV LQ WKH SDVW LH WKH PRGHOV DUH G\QDPLF +HQFH WKH WHUP fEDFNSURSDJDWLRQ WKURXJK WLPHf 7KLV HQWLUH RSWLPL]DWLRQ F\FOH LV UXQ DW HDFK WLPH VWHS W4 7KH RSWLPL]HU GHULYHV WKH QH[W 7 09 PRYHV DQG WKH ILUVW 09 VHWSRLQW LV DSSOLHG WR WKH XQLW PY f $W WKLV SRLQW WKH HQWLUH SURFHVV LV UHSHDWHG 0RGHO5HIHUHQFH $GDSWLYH &RQWUROOHU 7KH ILQDO FRQWUROOHU GHVLJQ FRQVLGHUHG EHORQJV WKH PRGHOUHIHUHQFH DGDSWLYH FRQWURO IDPLO\ 05$&f /LNH WKH G\QDPLF 03& FRQWUROOHU WKH 05$& FRQWUROOHU XQGHUVWDQGV SURFHVV G\QDPLFV DQG SURYLGHV D WUDMHFWRU\ RI 09 VHWSRLQWV ZKLFK RSWLPL]H ERWK ZKHUH \RX DUH JRLQJ DQG KRZ \RX JHW WKHUH 7KH IXQGDPHQWDO GLIIHUHQFH EHWZHHQ WKHVH WZR FRQn WUROOHUV LV KRZ WKLV RSWLPDO WUDMHFWRU\ LV GHULYHG 7KH 03& GHVLJQ XWLOL]HG DQ RQOLQH RSWLn PL]HU WR FDOFXODWH WKLV WUDMHFWRU\ ZKLOH WKH 05$& GHVLJQ GHYHORSV D QHXUDOQHWZRUN EDVHG FRQWUROOHU ZKLFK LV DEOH WR FDOFXODWH WKH RSWLPDO WUDMHFWRU\ GLUHFWO\ +HQFH WKLV LV RXU ILUVW GLUHFW FRQWUROOHU LH FDOFXODWHV 09 VHWSRLQWV GLUHFWO\ 1RWLFH WKDW WKH 0,& GHVLJQ ZRXOG KDYH SURYLGHG D GLUHFW RQOLQH FRQWUROOHU LI LW ZDVQfW IRU f WKH ODFN RI D NQRZQ WDUJHW 12[ OHYHO DQG f WKH UHTXLUHPHQWV IRU 09 DQG 69 FRQVWUDLQWV 7KH 05$& GHVLJQ LV DEOH WR RYHUFRPH ERWK RI WKHVH KXUGOHV E\ EXLOGLQJ NQRZOHGJH RI WKH EHVW DFKLHYDEOH 12[ OHYHO DQG E\ EXLOGLQJ DOO RI WKH FRQVWUDLQWV GLUHFWO\ LQWR WKH FRQn WUROOHU 7KH PDLQ DGYDQWDJH WR WKH 05$& GHVLJQ LV RQOLQH UHVSRQVH WLPH 7KHUH LV QR RSWL

PAGE 96

PL]DWLRQ WR UXQ RQH VLPSO\ SUHVHQWV WKH FRQWUROOHU ZLWK WKH FXUUHQW DQG SDVW VWDWH RI WKH SURFHVV DQG LW JHQHUDWHV D 09 VHWSRLQW DV TXLFNO\ DV D QHXUDO QHWZRUN FDQ WKLQN 7KHVH EHQHILWV GR QRW FRPH IRU IUHH KRZHYHU 7KH PDLQ GUDZEDFNV WR WKH 05$& GHVLJQ DUH f ([WHQVLYH RIIOLQH WUDLQLQJ DQG UHWXQLQJ UHTXLUHPHQWV f ,QIOH[LEOH RQOLQH FRQILJXUDWLRQ ZLWK UHVSHFW WR FKDQJLQJ RSWLPL]DWLRQ REMHFWLYHV DQG RSHUDWLQJ FRQVWUDLQWV 7KLV VHFWLRQV GHWDLOV WKH GHVLJQ RI D QHXUDO QHWZRUNEDVHG G\QDPLF 05$& FRQWUROOHU WR PHHW WKH SUREOHP VSHFLILFDWLRQV SUHVHQWHG LQ 6HFWLRQV DQG 2IIOLQH WUDLQLQJ 7UDLQLQJ D 05$& FRQWUROOHU UHTXLUHV WZR VWDJHV 7KH ILUVW VWDJH LV LGHQWLFDO WR WUDLQLQJ DQG UHWXUQLQJ WKH G\QDPLF 03& FRQWUROOHU +HUH G\QDPLF 69 DQG &9 PRGHOV DUH GHYHOn RSHG XVLQJ WKH VDPH VWHSV RXWOLQHG LQ )LJXUH 7KH VHFRQG VWDJH XVHV WKHVH PRGHOV WR WUDLQ WKH FRQWUROOHU ZLWK RIIOLQH GDWD DV LOOXVWUDWHG LQ )LJXUH 7KH RIIOLQH WUDLQLQJ LV VLPn LODU WR WKH RQOLQH RSWLPL]DWLRQ ZKLFK LV SHUIRUPHG IRU WKH 03& GHVLJQ H[FHSW WKLV RSWLPLn ]DWLRQ LV SHUIRUPHG DFURVV WKH WUDLQLQJ GDWDVHW UDWKHU WKDQ RQOLQH )LJXUH 2IIOLQH WUDLQLQJ DQG UHWXQLQJ FRQILJXUDWLRQ IRU PRGHO UHIHUHQFH FRQWUROOHU

PAGE 97

2QFH DJDLQ D G\QDPLF RSWLPL]HU LV UHTXLUHG DQG WKH REMHFWLYH IXQFWLRQ LV JLYHQ E\ Qa f 7R WUDLQ WKH FRQWUROOHU WKH RSWLPL]HU FDOFXODWHV $UJ0LQAFL^I Wf` ZKHUH L3 DUH WKH ZHLJKWV RI WKH FRQWURO ODZ QHXUDO QHWZRUN 7UDLQLQJ WKH FRQWURO ODZ &/f PRGHO VWDUWV ZLWK WKH DFWXDO YDOXHV IRU WKH '9V 69V DQG ILM &9V DQG UDQGRP LQLWLDO ZHLJKWV IRU LWV &/ PRGHO -S 6WDUWLQJ DW WLPH W M ZKHUH 7c LV WKH ILUVW VDPSOH LQ WKH WUDLQLQJ GDWDVHW WUDLQLQJ XVHV WKH &/ PRGHO WR HVWLPDWH A ? MM WKH UHVXOWLQJ 09V PY Wf ZKLFK DUH WKHQ XVHG WR HVWLPDWH WKH UHVXOWLQJ 69V VY W f f§A DQG &9V FY W f 7KLV SURFHVV LV UHSHDWHG RYHU WKH WLPH LQWHUYDO W H W W4 7@ WR W7? Y W7 SURGXFH WKH 09 69 DQG &9 WUDMHFWRULHV ^PY Wf`L LR ^VY `L LL DQO W7 ^FY WfM W4 L UHVSHFWLYHO\ 7KH WUDLQLQJ DOJRULWKP WKHQ LWHUDWLYHO\ XSGDWHV LWV HVWLPDWH IRU WKH RSWLPDO &/ PRGHO ZHLJKWV -3&/ WR PLQLPL]H LWV REMHFWLYH IXQFWLRQ I ?Wf (DFK VWHS LQ WKH LWHUDWLRQ SHUn IRUPV WKH IROORZLQJ VWDUWLQJ ZLWK W W4 7 DQG LWHUDWLQJ GRZQ WR W W )LUVW WKH WUDLQLQJ DOJRULWKP FDOFXODWHV WKH &9 VHQVLWLYLWLHV GI WfGFYr Wf IURP ZKLFK WKH &9 0RGHO LV DEOH WR FDOFXODWH 69 VHQVLWLYLWLHV GI WfGVYrWf DQG SDUWLDO 09 VHQVLWLYLWLHV GIYWfGPYrW f IURP ZKLFK WKH 69 PRGHO FDOFXODWHV WKH UHPDLQLQJ SDUWLDO 09 VHQ

PAGE 98

VLWLYLWLHV GIYWfGPYrW f DQG ILQDOO\ WKH 09 VHQVLWLYLW\ DW WLPH W FDQ EH FDOFXn ODWHG DV GIWf GIYWf GI?Wf 7 GPYrW f GPYrW f GPYrWf§?f 7KH 09 VHQVLWLYLWLHV DUH ILQDOO\ SDVVHG WR WKH &/ PRGHO ZKLFK EDFNSURSDJDWHV WKHP WR GHULYH LWV ZHLJKW JUDGLHQWV GI G: ZKLFK WKH WUDLQLQJ DOJRULWKP LV DEOH WR XVH WR XSGDWH LWV FRQWURO ODZfV ZHLJKW HVWLPDWH 7KH WUDLQLQJ DOJRULWKP WKHQ LQFUHPHQWV WD DQG UHSHDWV WKH HQWLUH SURFHVV XQWLO WKH WUDLQLQJ DOJRULWKP KDV FRQYHUJHG :KHQ W 7Mf§ 7 W4 LV UHVHW WR W4 M 7KH UHDn VRQ IRU WUDLQLQJ WKH &/ PRGHO LQ LQFUHPHQWV RI 7 LV EHFDXVH WKH 69 DQG &9 KDYH D OLPLWHG SUHGLFWLRQ KRUL]RQ WKH WLPH EHIRUH WKHLU HVWLPDWHV DUH QR ORQJHU YDOLG %\ UHVHWWLQJ WKH VWDWH RI WKHVH PRGHOV WR WKH DFWXDO VWDWH RI WKH XQLW DIWHU 7 VDPSOHV ZH DUH DEOH WUDLQ WKH &/ PRGHO ZLWKLQ WKH SUHGLFWLRQ KRUL]RQ RI WKH 69 DQG &9 PRGHOV 2QOLQH FRQWURO 7KH RQOLQH FRQWURO FRQILJXUDWLRQ IRU WKH 05$& GHVLJQ LV VWUDLJKW IRUZDUG DV LOOXVn WUDWHG LQ )LJXUH 6LPSO\ VXSSO\ WKH FRQWUROOHU ZLWK WKH FXUUHQW 69V '9V DQG &9V DQG LW RXWSXWV WKH QH[W 09 VHWSRLQW 7KLV VHWSRLQWV FRQWDLQV NQRZOHGJH DERXW WKH RSWLPDO DFKLHYDEOH 12[ 69 FRQVWUDLQWV 09 FRQVWUDLQWV DQG WKH WUDMHFWRU\ WKURXJK ZKLFK LW ZLOO GULYH WKH SURFHVV LQWR WKH IXWXUH

PAGE 99

^VYWf GYWf` &RQWURO PYrWf 3ODQW FYWf IA /DZ )LJXUH 2QOLQH FRQWURO FRQILJXUDWLRQ IRU PRGHO UHIHUHQFH FRQWUROOHU &OHDUO\ WKH FRQWUROOHU LV RQO\ DV JRRG DV LWV XQGHUO\LQJ UHIHUHQFH PRGHOV ,Q DGGLWLRQ FRQVLGHUDEOH FDUH PXVW EH WDNHQ WR HQVXUH WKDW WKH WUDLQLQJ GDWD FRQWDLQV UHJLRQV RI WKH LQSXW VSDFH ZKHUH WKH 69 DQG 09 FRQVWUDLQWV KDYH EHHQ H[HUFLVHG 7KH GHVLJQ LV HDVLO\ DXJPHQWHG ZLWK OLPLWHUV WR JXDUDQWHH WKDW 09 FRQVWUDLQWV DUH PDLQWDLQHG +RZHYHU WKHUH LV OLWWOH WKDW FDQ EH GRQH WR JXDUDQWHH WKDW WKH 69 FRQVWUDLQWV DUH PDLQWDLQHG

PAGE 100

&+$37(5 '$7$ 35(3$5$7,21 *LYHQ WKH GHWDLOHG FRQWURO GHVLJQV MXVW SUHVHQWHG WKH QH[W VWHS LV WR LPSOHPHQW WKH DFWXDO FRQWUROOHUV E\ GHYHORSLQJ WKH UHTXLUHG UHIHUHQFH PRGHOV %RWK UHIHUHQFH PRGHO DQG FRQWUROOHU LPSOHPHQWDWLRQV UHTXLUH D VLJQLILFDQW DPRXQW RI SURFHVV GDWD 'DWD FROOHFWLRQ LV WKH PRVW LPSRUWDQW DVSHFW RI DQ\ PRGHOLQJ RU RSWLPL]DWLRQ SURMHFW 7KHUH LV D FRPPRQ VD\LQJ fMXQN LQ MXQN RXWf WKLV VWXG\ ZDV UHOHQWOHVV LQ UHHQIRUFLQJ WKLV OHVVRQ $SSO\LQJ WKH PRVW VRSKLVWLFDWHG PRGHOLQJ DQGRU RSWLPDO FRQWURO DOJRULWKPV LQ WKH ZRUOG ZLOO QRW PDNH XS IRU SUREOHPV ZLWK GDWD SUHSDUDWLRQ :LWK WKH DGYDQFHG GLVWULEXWHG FRQWURO V\VWHPV ''6f DQG VXSHUYLVRU\ FRQWURO DQG GDWD DFTXLVLWLRQ 6& $' $f V\VWHPV UHDGLO\ DYDLODEOH LQ WRGD\fV SURFHVV SODQWV WKH UHODWLYH TXDQWLW\ DQG TXDOLW\ RI DYDLODEOH GDWD LV RYHUZKHOPLQJ 0XFK RI WKH VWDWLVWLFV DQG PRGHOn LQJ OLWHUDWXUH KDV EHHQ GHGLFDWHG WR WKH SUREOHPV IDFHG ZKHQ GUDZLQJ LQIHUHQFHV IURP VPDOO VDPSOH VSDFHV 0RGHP SURFHVVLQJ SODQWV DUH DQ\WKLQJ EXW GDWD OLPLWHG 7KH UHOHYDQW SUREOHPV DUH MXVW WKH RSSRVLWH KRZ WR GUDZ PHDQLQJIXO LQIHUHQFHV IURP D PDVVLYH VDPSOH VSDFH 7KH IROORZLQJ VHFWLRQ SUHVHQWV VROXWLRQV IRU WKH PRVW VLJQLILFDQW FKDOOHQJHV IDFHG LQ SUHSDULQJ GDWD IRU PRGHOLQJ DQG RSWLPL]DWLRQ 0XFK RI ZKDW LV SUHVHQWHG LQ WKLV VHFWLRQ ZDV OHDUQHG WKH KDUG ZD\ GXULQJ PRGHOLQJ DQG RSWLPL]DWLRQ

PAGE 101

'DWD 0DQDJHPHQW 7KH SRZHU SODQW WUHDWHG LQ WKLV ZRUN FROOHFWV DQG VWRUHV WHQV RI WKRXVDQGV RI YDULDEOHV IURP YDULRXV VHQVRUV DQG DFWXDWRUV WKURXJKRXW WKH SODQW 7KHVH YDULDEOHV ZHUH FROOHFWHG E\ WKH '&6 DQG IRUZDUGHG WR D GDWD KLVWRULDQ FDOOHG 3, PDUNHWHG E\ 26, ,QFRUSRUDWHG 1RWDWLRQ 7KH DXWKRU ZLOO UHVHUYH WKH WHUP fYDULDEOHf WR UHSUHVHQW SURFHVV VWDWHV ZLWK D VSHFLILF SK\VLFDO LQWHUSUHWDWLRQ ZKLFK DUH FRQVLGHUHG UHOHYDQW IRU PRGHOLQJ 7KH '&6 ZLOO FROOHFW LQIRUPDWLRQ IURP PDQ\ VHQVRUV RU DFWXDWRUV HDFK RI ZKLFK UHSUHVHQW D VLQJOH SURFHVV YDULDEOH 7KH VSHFLILF SRLQWV FROOHFWHG E\ WKH '&6 ZLOO EH UHIHUUHG WR DV fWDJVf 7KH '&6 ZRUNV ZLWK H[FHSWLRQEDVHG VDPSOLQJ D QRQXQLIRUP VDPSOLQJ VFKHPH (DFK WDJ LV JLYHQ DQ DEVROXWH GHYLDWLRQ DERYH ZKLFK DQ H[FHSWLRQ LV UDLVHG 5DLVHG H[FHSn WLRQV DUH IRUZDUGHG WR WKH SURFHVV FRQWURO DOJRULWKPV ZLWKLQ WKH '&6 DV ZHOO DV WR 3, 7KH 3, GDWD KLVWRULDQ WKHQ DSSOLHV D WLPH VWDPS DQG VWRUHV WKH H[FHSWLRQ 7KH WLPH UHVROXWLRQ RI D VDPSOH FDQ EH WUXVWHG ZLWKLQ DERXW s VHFRQGV DQG WKH TXDQWL]DWLRQ HUURU LV DSSUR[Ln PDWHO\ HTXDO WR WKH GHYLDWLRQ VHWWLQJV IRU HDFK WDJ LQ WKH '&6 7KLV H[FHSWLRQEDVHG GDWD DFTXLVLWLRQ VFKHPH DOORZV WKH GDWD KLVWRULDQ WR VWRUH DQ LPSUHVVLYH DPRXQW RI GDWD 8QLIRUPO\ VDPSOHG WUHQGV DUH SURYLGHG WKURXJK VLPSOH IXQFn WLRQ FDOOV WR 3LfV $3, $Q\ VXEVHW RI WKH WDJV FDQ EH HDVLO\ UHFDOOHG IRU DUELWUDU\ WLPH UDQJHV RYHU WKH ODVW \HDU RI FRQWLQXRXV SODQW RSHUDWLRQ ,QLWLDO FRQFHUQV RYHU WKH GDWD TXDOLW\ JLYHQ WKLV QRQXQLIRUP VDPSOLQJ VFKHPH ZHUH TXLFNO\ GLVPLVVHG +RZHYHU VLJn QLILFDQW WXQLQJ RI WKH WDJ GHYLDWLRQV ZDV UHTXLUHG KRZHYHU

PAGE 102

9DULDEOH 6HOHFWLRQ )RU WKH FDVH VWXG\ RI ERLOHU RSWLPL]DWLRQ LW LV FOHDU WKDW ZH DUH ORRNLQJ IRU SURFHVV YDULDEOHV WKDW DUH UHODWHG WR WKH PHDVXUHPHQW RI 12[ RU &2 DQG LPSDFW WKH FRPEXVWLRQ SURFHVV ZLWK UHVSHFW WR WKHVH PHDVXUHPHQWV :LWK WKH PDVVLYH QXPEHU RI YDULDEOHV WR FKRRVH IURP LQ WKH GDWDEDVH WKH ILUVW OLQH RI GHIHQVH LV WR XVH RXU ILUVWSULQFLSOHV NQRZOn HGJH RI WKH SURFHVV ,QWHUYLHZV ZHUH FRQGXFWHG ZLWK HQJLQHHULQJ DQG RSHUDWLRQV SHUVRQQHO IURP WKH SODQW 7KH\ ZHUH DVNHG WR f ,GHQWLI\ DOO WDJV WKDW UHSUHVHQW WKH YDULDEOHV 12[ DQG &2 f ,GHQWLI\ DOO YDULDEOHV WKDW KDYH DQ\ DIIHFW RQ FRPEXVWLRQ SDUDPHWHUV OLNH HPLVVLRQV IXHO DQG DLU IORZV WHPSHUDWXUHV RU SUHVVXUHV LQVLGH WKH ERLOHU f 5DQN WKHVH YDULDEOHV ZLWK UHVSHFW WR WKHLU HIIHFW RQ 12[ DQG &2 DV HVVHQWLDO VHFRQGDU\ RU PLQLPDO f ,GHQWLI\ DOO RI WKH WDJV WKDW UHSUHVHQW RU GLUHFWO\ LPSDFW WKH OLVWHG YDULn DEOHV f &ODVVLI\ HDFK RI WKHVH WDJV DV HLWKHU f6HWSRLQW FDQ EH PDQLSXODWHG E\ RSHUDWRUV YLD WKH '&6 f7XQDEOH 3DUDPHWHU 73f FDQ EH PDQXDOO\ PDQLSXODWHG E\ fHQJLn QHHUV ZLWK ZUHQFKHVf f'LVWXUEDQFH FDQ QRW EH PDQLSXODWHG EXW KDV DQ DIIHFW RQ FRPn EXVWLRQ f6WDWH UHSUHVHQWV D SDUWLFXODU VWDWH RI FRPEXVWLRQ ZKLFK ZLOO KDYH DQ DIIHFW RQ HPLVVLRQV FDQ EH D IXQFWLRQ RI VHWSRLQWV DQGRU GLV

PAGE 103

WXUEDQFHV EXW FDQQRW EH PDQLSXODWHG GLUHFWO\ f ,GHQWLI\ RSHUDWLQJ FRQVWUDLQWV DQG FRQFHUQV WKDW UHTXLUH PRQLWRULQJ ZKHQ PDQLSXODWLQJ DQ\ RI WKH VHWSRLQWV 7KLV SURFHVV DQG VXEVHTXHQW LWHUDWLRQV SURGXFHG HVVHQWLDO VHFRQGDU\ DQG PLQLPDO WDJV IRU &DQDO (OHFWULF *HQHUDWLQJ 6WDWLRQ 7KH HVVHQWLDO WDJV DUH OLVWHG LQ $SSHQn GL[ 7KH RSHUDWRUV DQG HQJLQHHUV LGHQWLILHG &2 DV WKHLU RYHUZKHOPLQJ RSHUDWLQJ FRQn VWUDLQW 7KH SURFHVV RI YDULDEOH VHOHFWLRQ VKDOO EH FRQWLQXHG LQ 6HFWLRQ 0RGHO 'HILQLWLRQV 9DOLGDWLRQ ,Q RUGHU WR JHW D KDQGOH RQ WKH TXDOLW\ RI GDWD EHLQJ FROOHFWHG E\ WKH GDWD KLVWRULDQ IRU HDFK RI WKH HVVHQWLDO WDJV DERYH RQH PRQWK RI GDWD ZHUH H[WUDFWHG EHWZHHQ DQG 7KH VDPSOLQJ UDWH ZDV VHW WR 7V VHFRQGV 2QH RI WKH PRVW XVHIXO GDWD YDOLn GDWLRQ WRROV ZDV WR VLPSO\ WUHQG HDFK WDJ RYHU YDULRXV LQWHUYDOV 7KH H\H LV DEOH WR VSRW PRVW GDWD LQWHJULW\ LVVXHV WKDW DUH FRPPRQO\ PLVVHG E\ VWDWLVWLFDO LQGLFDWRUV 7KH VWDQGDUG GHVFULSWLYH VWDWLVWLFV RI PHDQ YDULDQFH PD[ DQG VWDQGDUG HUURU ZHUH FDOn FXODWHG IRU HDFK WDJ RYHU WKH GDWDVHWV 7KHVH VWDWLVWLFV ZHUH FRPSDUHG DJDLQVW ILUVWSULQFLn SOHV NQRZOHGJH WR ORRN IRU WKH GDWD LQWHJULW\ LVVXHV FRQVLGHUHG EHORZ 4XDQWL]DWLRQ DQG &OLSSLQJ 7KH PRVW FRPPRQ GDWD LQWHJULW\ SUREOHP HQFRXQWHUHG ZLWK LQGXVWULDO GDWD KLVWRULDQV LV KDYLQJ LQYDOLG UDQJH VHWWLQJV 5DQJH VHWWLQJV IRU HYHQW EDVHG GDWD DFTXLVLWLRQ V\VWHPV DUH DQDORJRXV WR VDPSOLQJ UDWH IRU XQLIRUP VDPSOLQJ 7KHUH DUH WKUHH LPSRUWDQW VHWWLQJV WKH GDWD PLQLPXP GDWD PD[LPXP DQG DEVROXWH GHYLDWLRQ :KHQ HLWKHU WKH GDWD PLQLPXP

PAGE 104

VHWWLQJV DUH WRR KLJK RU WKH GDWD PD[LPXP VHWWLQJV DUH WRR ORZ WKH DUFKLYHG GDWD LV FOLSSHG ,Q DGGLWLRQ LI WKH GHYLDWLRQ VHWWLQJ LV WRR ODUJH TXDQWL]DWLRQ HUURUV FDQ VLJQLILn FDQWO\ FRUUXSW WKH DUFKLYHG GDWD 7KH EHVW ZD\ WR GHWHFW WKHVH HUURUV LV VLPSOH YLVXDO LQVSHFWLRQ DQG WKH RQO\ UHPHG\ LV WR FRUUHFW WKH VHWWLQJV 0LVVLQJ 'DWD $ORQJ ZLWK HDFK VDPSOH WKH KLVWRULDQ SURYLGHG VWDWXV LQIRUPDWLRQ LQGLFDWLQJ DQ\ HUURUV WKDW ZHUH GHWHFWHG E\ WKH '&6 ZKHQ FROOHFWLQJ WKH VDPSOH &RQVLGHULQJ DQ\ HUURU DV PLVVLQJ WKH GDWDVHW FRQWDLQHG b PLVVLQJ GDWD 6LQFH RQH RI WKH JRDOV RI WKLV SURMHFW LV WR HYDOXDWH WKH SHUIRUPDQFH RI ERWK VWDWLF DQG G\QDPLF PRGHOV DOO PLVVLQJ YDOXHV ZLOO QHHG WR EH DFFRXQWHG IRU GHVSLWH WKH IDFW WKDW WKHUH QXPEHUV DUH TXLWH ORZ b ‘ *RRG ’ 0LVVLQJ b b b b Y6 I& Q6 . nLnY $ A A A A A 1& )LJXUH 'DLO\ b PLVVLQJ DFURVV )HEUXDU\ GDWDVHW )LJXUHV LOOXVWUDWHV WKH GDLO\ VXP RI PLVVLQJ YDOXHV DFURVV WKH GDWDVHW 2QH FDQ HDVn LO\ LGHQWLI\ D UHJLRQ RI IRXU FRQVHFXWLYH GD\V ZKHUH WKH SODQW ZDV PLVVLQJ GDWD 7KLV UHJLRQ FRUUHVSRQGV WR D XQLW VKXWGRZQ 1RWH WKDW UHPRYLQJ ODUJH EORFNV RI WLPH IURP WKH

PAGE 105

GDWD ZLOO QRW EH DQ LVVXH IRU G\QDPLF PRGHOLQJ DV ORQJ DV WKH YDOLG GDWD IDOOV LQWR D UHODn WLYHO\ VPDOO QXPEHU RI EORFNV ZLWK D ODUJH QXPEHU RI VDPSOHV WKDW DUH IUHH RI PLVVLQJ YDOn XHV 5HPRYLQJ WKLV UHJLRQ IURP RXU DQDO\VLV WKH PLVVLQJ SHUFHQWDJHV EHFRPH OHVV WKDQ b 6LQFH WKH UHPDLQLQJ HUURUV DUH KLJKO\ VSRUDGLF DQG GRZQ VDPSOLQJ RI WKH GDWD LV UHTXLUHG WKHUH H[LVW D VROXWLRQ IRU GHDOLQJ ZLWK WKH UHPDLQLQJ HUURUV 7KH FROOHFWHG GDWD ZDV VDPSOHG DW 7V VHFRQGV 'RZQVDPSOLQJ WR 7V VHFRQGV ZLOO EH DSSOLHG E\ DYHUDJLQJ FRQVHFXWLYH VDPSOHV DQG GHFLPDWLQJ ,I RU IHZHU RI WKH VDPSOHV DUH LQ HUURU WKHQ WKH UHPDLQLQJ ZLOO EH DYHUDJHV DQG WKH VWDWXV RI WKH GHFLPDWHG VDPSOHfV VWDWXV ZLOO EH VHW WR fJRRGf ,I PRUH WKDQ VDPSOHV DUH LQ HUURU WKHQ WKH GHFLPDWHG VDPSOHfV VWDn WXV ZLOO EH VHW WR fPLVVLQJf 7KLV SURFHGXUH UHPRYHG DOO UHPDLQLQJ HUURUV IURP WKLV GDWDVHW $OWKRXJK WKH DERYH SURFHGXUH ZDV DEOH WR FOHDQ DOO RI WKH GDWD IRU WKLV GDWDVHW LW LV VWLOO SRVVLEOH WKDW IXWXUH GDWDVHWV PD\ VWLOO FRQWDLQ PLVVLQJ YDOXHV )RU DOO RWKHU FDVHV PLVVLQJ GDWD ZLOO EH UHSODFHG ZLWK LQWHUSRODWHG GDWD EHWZHHQ WKH QHDUHVW VXUURXQGLQJ YDOLG VDPn SOHV 2XWOLHUV 7UHDWLQJ HDFK WDJ DV D UDQGRP YDULDEOH WKH WDJV ZHUH VWDQGDUGL]HG WR D =6FRUH E\ VXEWUDFWLQJ WKH PHDQ DQG GLYLGLQJ E\ WKHLU VWDQGDUG GHYLDWLRQ f

PAGE 106

7KH =VFRUH UHPRYHV DOO HIIHFWV RI RIIVHW DQG PHDVXUHPHQW VFDOH 7KH\ FDQ DSSURDFK SRVLWLYH DQG QHJDWLYH LQILQLW\ 7DEOH VKRZV WKH SUREDELOLW\ RI WKH DEVROXWH YDOXH RI D = VFRUH H[FHHGLQJ VRPH OLPLW IRU QRUPDOO\ GLVWULEXWHG YDULDEOHV 7DEOH 3UREDELOLW\ RI =6FRUH H[FHHGLQJ YDOXH _=6FRUH_ 3([FHHGLQJf &DOFXODWLQJ WKH SUREDELOLWLHV RI D =VFRUH YDOXH H[FHHGLQJ D IRU RXU GDWDVHWV DV 1 3= Df \ ? WIIL r ?=c;Mf? D HOVH f ZKHUH 1 LV WKH WRWDO QXPEHU RI VDPSOHV LQ WKH GDWDVHW 7DJV ZLWK 3= f ZH FRQVLGHUHG FDQGLGDWHV IRU ILOWHULQJ RU VPRRWKLQJ RSHUDWLRQV DQG WDJV ZLWK 3=c f ZHUH FRQVLGHUHG FDQGLGDWHV IRU RXWOLHU UHPRYDO )LOWHULQJ RSHUDWLRQV ZHUH PRVW FRPPRQO\ DSSOLHG WR WDJV ZLWK VHQVRU FOHDQLQJ DQG FDOn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fV YDOXH IRU D VKRUW

PAGE 107

SHULRG DIWHU WKLV SHULRG WKH WDJfV YDOXH ZRXOG LQVWDQWDQHRXVO\ UHWXUQ WR LWV WUXH YDOXH 2XWn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n XWHV 7KLV VLWXDWLRQ KDV EHHQ FRPSHQVDWHG IRU E\ XVLQJ D VLPSOH ILUVWRUGHU ORZ SDVV ILOWHU WR GDPSHQ WKH DFWXDWRU VHWSRLQW VLJQDO WR PDWFK WKH DFWXDWRUV UHVSRQVH FKDUDFWHULVWLFV 6HFRQG VRPH VHQVRUV SDUWLFXODUO\ WKH FRQWLQXRXV HPLVVLRQV PRQLWRUV &(0Vf FDQ KDYH VLJQLILFDQW H[WUDFWLRQ WLPHV 7KH &DQDO (OHFWULF *HQHUDWLQJ 6WDWLRQ &(0V PHDVXULQJ 12[ DQG &2 DGG D WR PLQXWH GHDGWLPH EHWZHHQ WKH JDVHV IRUPDWLRQ LQ WKH ERLOHU XQWLO WKH\ DUH UHFRUGHG 7KLV VLWXDWLRQ KDV EHHQ WDNHQ FDUH RI E\ VKLIWLQJ WKHVH WDJV ZLWKLQ WKH GDWDVHW VXFK WKDW VHWSRLQWV DQG SURFHVV RXWSXWV DUH DOLJQHG 1RUPDOL]DWLRQ ,Q VXSSRUW RI PRGHOLQJ DOO WDJV ZHUH QRUPDOL]HG 9DULDEOHV DSSOLHG WR D QHXUDO QHWn ZRUN VKRXOG IDOO ZLWKLQ WKH QHXURQfV DFWLYDWLRQ OLPLWV $OO RI WKH QHXUDO QHWZRUNV FRQVLGn HUHG LQ WKLV VWXG\ XWLOL]H D WDQN DFWLYDWLRQ IXQFWLRQ WKHUHIRUH DOO YDULDEOHV ZHUH QRUPDOL]HG

PAGE 108

WR IDOO ZLWKLQ WKH UDQJH >@ $VVXPLQJ QRUPDO GLVWULEXWLRQV WKH WDJ =VFRUHfV ZHUH XVHG IRU QRUPDOL]DWLRQ VXFK WKDW RQO\ b RI WKH GDWD ZRXOG IDOO RXWVLGH WKH QHXURQfV DFWLn YDWLRQ OLPLWV DFFRUGLQJ WR [ f ZKHUH U f f LV WKH UDWLR EHWZHHQ DFFHSWDEOH =VFRUH UDQJH DQG WKH QHXURQfV DFWLYDWLRQ UDQJH 'HQRUPDOL]DWLRQ ZDV WKHQ SUHIRUPHG DFFRUGLQJ WR

PAGE 109

&+$37(5 02'(/,1* 7KH PRGHOLQJ REMHFWLYHV IRU WKLV ZRUN DUH LQKHUHQWO\ WLHG WR WKH FRQWURO GHVLJQV DQG WKH FDVH VWXG\ SUHVHQWHG LQ WKH SURFHHGLQJ VHFWLRQV 7R WKLV HQG WKH REMHFWLYH RI WKLV VHFWLRQ LV WR GHWHUPLQH WKH EHVW PRGHO DUFKLWHFWXUH IRU HDFK RI WKH PRGHO GHILQLWLRQV UHTXLUHG E\ WKH YDULRXV FRQWURO GHVLJQV FRQVLGHUHG &DQGLGDWH PRGHO DUFKLWHFWXUHV ZLOO EH MXGJHG EDVHG RQ WKHLU DELOLW\ WR SUHGLFW SURFHVV G\QDPLFV 7KH IROORZLQJ DUFKLWHFWXUHV ZLOO EH FRQVLGHUHG f $XWR5HJUHVVLYH 0RYLQJ $YHUDJH 0RGHO $50$f f 0XOWLOD\HU 3HUFHSWURQ 0/3f f 7LPH'HOD\ 1HXUDO 1HWZRUN 7'11f f *DPPD 1HXUDO 1HWZRUN *11f f 1RQOLQHDU 6WDWH6SDFH 0RGHO 1/66f 1RWLFH WKDW WKH $50$ PRGHO KDV EHHQ LQFOXGHG WR SURYLGH D EHQFKPDUN DQG WR YDOLn GDWH WKH DSSOLFDWLRQ RI QRQOLQHDU FRQWURO VWUDWHJLHV IRU WKH FDVH VWXG\ DSSOLFDWLRQ 0HWKRGRORJ\ 7KH PHWKRGRORJ\ IRU GHYHORSLQJ WKH fEHVWf PRGHOV ZLOO EH DV IROORZV f 0RGHO 'HILQLWLRQV 7KH XOWLPDWH JRDO LV WR ILQG WKH EHVW SRVVLEOH PRGHOV IRU 12[ DQG &2 7KH FRQWURO GHVLJQV SUHVHQWHG UHTXLUH WKDW WKHVH SURn FHVVHV DUH UHSUHVHQWHG XVLQJ VSHFLILF PRGHO GHILQLWLRQV 'HWDLOHG VSHFLn ILFDWLRQV IRU WKH PRGHOV DUH GHYHORSHG LH LGHQWLI\LQJ WKH VSHFLILF SURFHVV LQSXWV DQG RXWSXWV WR EH XVHG E\ WKH PRGHOV f 'DWDVHWV $ PRQWK GDWDVHW ZLOO EH JHQHUDWHG XVLQJ WKH PHWKRGRORJ\ SUHVHQWHG LQ &KDSWHU DQG GLYLGHG LQWR GLVMRLQW WUDLQLQJ FURVVYDOLGDn WLRQ DQG WHVWLQJ UHJLRQV

PAGE 110

f /HDUQLQJ $OJRULWKP 7KH OHDUQLQJ DOJRULWKP XVHG IRU PRGHO WUDLQLQJ FRPELQHV WKH 3RODN5LELHUH DOJRULWKP >@ ZLWK D OLQH VHDUFK DV SUHn VHQWHG E\ %UHQW >@ 7KLV DOJRULWKP LV SUHVHQWHG DORQJ ZLWK WKH GHWDLOV GHVFULELQJ LWV DSSOLFDWLRQ f 3HUIRUPDQFH &ULWHULD 7KH FULWHULD IRU VHOHFWLQJ WKH fEHVWf PRGHO DUH SUHVHQWHG f 9DULDEOH 3UXQLQJ $ 0/3 LV FRQVWUXFWHG DQG VWDUWLQJ IURP WKH LQSXW VHWV GHWHUPLQHG IURP ILUVWSULQFLSOHV NQRZOHGJH LQ 6HFWLRQ 9DULn DEOH 6HOHFWLRQ WKH LQSXW VHWV DUH SUXQHG WR WKH VPDOOHVW SRVVLEOH VHW RI UHOHYDQW YDULDEOHV f $UFKLWHFWXUH 6HOHFWLRQ 2SWLPDO SDUDPHWHUV HJ QXPEHU RI KLGGHQ OD\n HUV SURFHVVLQJ HOHPHQWV DQG PHPRU\ WDSVf ZLOO EH LQGLYLGXDOO\ GHWHUn PLQHG IRU HDFK FRPELQDWLRQ RI DUFKLWHFWXUH DQG SURFHVV RXWSXW XVLQJ D GLUHFW VHDUFK PHWKRGRORJ\ f $QDO\VLV 7KH UHVXOWV ZLOO EH DQDO\]HG WR ILQG WKH EHVW VWHDG\VWDWH DQG G\QDPLF PRGHOV IRU HDFK PRGHO GHILQLWLRQ 0RGHO 'HILQLWLRQV ,Q VXSSRUW RI WKH FRQWURO GHVLJQV SUHVHQWHG LQ &KDSWHU PRGHOV ZLOO EH GHYHORSHG DFFRUGLQJ WR WKH IROORZLQJ PRGHO GHILQLWLRQV f 6WHDG\6WDWH 69 0RGHO VY VV690RGHOPY GYf f 6WHDG\6WDWH &9 0RGHO FY VV&90RGHOPY GY VYf f 6WHDG\6WDWH ,69 0RGHO VY VV,690RGHOFY GYf f 6WHDG\6WDWH ,09 0RGHO PY VV,090RGHOFY VY GYf f '\QDPLF 69 0RGHO VYW f G690RGHOPYWf GYWff f '\QDPLF &9 0RGHO FYW f G&90RGHOPYWf GYWf VYW ff ZKHUH WKH YHFWRUV PY GY VY DQG FY UHSUHVHQW WKH VWHDG\ VWDWH YDOXHV RI WKH 09 '9 69 DQG &9 YDULDEOH VHWV UHVSHFWLYHO\ PYWf GYWf VYWf DQG FYWf UHSUHVHQW

PAGE 111

WKHLU UHVSHFWLYH YDOXHV DW WLPH W DQG E\ FY VV&90RGHO^PY GY VYf ZH PHDQ WKDW VV&90RGHO LV D PRGHO ZKLFK WDNHV RQ YHFWRUV PY GY DQG VY DV LQSXWV DQG SURGXFHV WKH YHFWRU FY DV DQ RXWSXW 1RWDWLRQ 7KH WHUP PRGHO GHILQLWLRQ ZLOO EH XVHG WR GHVFULEH WKH LQSXWRXWn SXW VSDFH RI D PRGHO DORQJ ZLWK ZKHWKHU LW LV VWHDG\VWDWH RU G\QDPLF ZKHUH PRGHO UHIHUV WR D SDUWLFXODU UHDOL]DWLRQ RI D SUHGLFWRU ZKLFK LPSOHn PHQWV WKH PRGHO GHILQLWLRQ 7KHUH ZLOO EH PDQ\ PRGHOV GHYHORSHG ZKLFK LPSOHPHQW HDFK RI WKH DERYH PRGHO GHILQLWLRQV 9DULDEOH 'HILQLWLRQV 7KH PRGHO GHILQLWLRQV ZHUH EDVHG RQ GHILQLWLRQV IRU WKH 09 '9 69 DQG &9 YDULDEOH VHWV 7KH WDJ OLVW LQ $SSHQGL[ LV FDWHJRUL]HG DFFRUGLQJ WR WKH YDULDEOH GHILQLWLRQV ZKLFK VKDOO WKHQ EH XVHG WR LPSOHPHQW WKH PRGHO GHILQLWLRQV &RQWURO YDULDEOHV 7KH FDVH VWXG\ FRQVLGHUV D VLQJOH FRQWURO YDULDEOH 12[ :KLOH &2 LV D SURFHVV RXWSXW DQG PDLQWDLQLQJ DSSURSULDWH OHYHOV RI &2 LV DOVR DQ REMHFWLYH &2 VKDOO EH FRQVLGHUHG D FRQVWUDLQHG 69 1RWLFH WKDW 12[ KDV EHHQ PDUNHG DV D &9 LQ WKH (VVHQWLDO 7DJ /LVW XQGHU WKH ILHOG ODEHOHG f7\SHf 0DQLSXODWHG YDULDEOHV 7KH 09V DUH GHILQHG DV WKH YDULDEOHV WKDW ZH ZDQW WKH FRQWUROOHU WR PDQLSXODWH 7KHUH DUH PDQ\ LQSXWV WKDW FRXOG EH PDQLSXODWHG EXW WKHUH LV D FRVW DVVRFLDWHG ZLWK WKH QXPEHU RI PDQLSXODWHG LQSXWV 7KHVH FRVWV LQFOXGH f JHQHUDOL]DWLRQ FRVWV DVVRFLDWHG ZLWK WKH fFXUVH RI GLPHQVLRQDOLW\f

PAGE 112

f FRPSXWDWLRQDO FRVWV DVVRFLDWHG ZLWK WKH RSWLPL]DWLRQ GXULQJ ERWK PRGn HOLQJ DQG FRQWURO DQG f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n WUROOHU WR FRQVWUDLQ LWV DOORZDEOH OHYHOV 7KLV ZLOO QRW KRZHYHU EH RXU RQO\ 69 7KH LQLWLDO VHW RI 69V FRQVLGHUHG IRU PRGHOLQJ LV SUHVHQWHG LQ WKH (VVHQWLDO 7DJ /LVW 9DULDEOH UHSUHVHQWDWLRQ
PAGE 113

9DULDEOH 6HOHFWLRQ UHVXOWHG LQ VXEVHWV RI WDJV WKDW UHSUHVHQW RU LPSDFW WKH VDPH SK\VLFDO YDULDEOH EXW WKH\ HDFK UHSUHVHQW WKH YDULDEOH LQ D VOLJKWO\ GLIIHUHQW ZD\ )RU H[DPSOH FRQVLGHU WKH SK\VLFDO YDULDEOH RI JURVV DLUIORZ WKH WRWDO DPRXQW RI DLU HQWHULQJ WKH ERLOHU ,W LV FOHDU IURP ILUVWSULQFLSOHV WKDW WKLV YDULDEOH KDV D VLJQLILFDQW LPSDFW RQ 12[ VHH &KDSWHU f 7KH IRUFHG GUDIW )'f IDQV GHOLYHU JURVV DLUIORZ WR WKH FRPEXVWLRQ SURFHVV VHH 6HFWLRQ )RVVLO)LUHG 3RZHU *HQHUDWLRQf 7KH RXWSXW RI WKH )' IDQV LV GHULYHG IURP WKH ERLOHU PDVWHU VLJQDO )RUFHG GUDIW RXWSXW LV VSHFLILHG DORQJ ZLWK IXHO IORZ E\ WKH IXHODLU FXUYH RI WKH ERLOHU 7KH '&6 KDV ILYH GLIIHUHQW UHSUHVHQWDWLYH WDJV IRU WKH YDULDEOH RI JURVV DLUIORZ f )DQ YDQH SRVLWLRQ 7KH )' IDQV DUH FRQVWDQW VSHHG IDQV PHDQLQJ WKDW WKH IDQ VKDIW WXUQV DW D FRQVWDQW VSHHG ZKLOH PRUH RU OHVV DLU ZLWK PRUH RU OHVV LQLWLDO VSLQ FDQ EH GXPSHG LQWR WKH EODGHV E\ RSHQLQJ RU FORVLQJ WKH YDQHV f )DQ DPSV ,I WKH LQOHW YDQHV DUH RSHQHG ZLGHU WKH IORZ LV JUHDWHU DQG PRUH ZRUN LV EHLQJ GRQH DQG WKH DPSHUDJH PXVW LQFUHDVH f )' IDQ GHPDQG 7KH IXHODLU FXUYH JLYHV D WRWDO DLU IORZ UHTXLUHPHQW ZKLFK LV FKDUDFWHUL]HG ZLWKLQ WKH '&6 DV D GHPDQG VLJQDO IRU WKH )' IDQ FRQWUROOHU f WULP 3ULRU WR SUHVHQWLQJ WKH )' IDQ GHPDQG VLJQDO WR WKH )' IDQ FRQWUROOHU WKH RSHUDWRU LV SURYLGHG ZLWK D WULP VLJQDO WKDW FDQ VKLIW WKLV GHPDQG b RI LWV UDQJH 7KLV WULP DOORZV WKH RSHUDWRUV WR DGG RU UHPRYH JURVV DLU DW WKHLU GLVFUHWLRQ f )' IDQ VHWSRLQW 7KLV '&6 VLJQDO ZKLFK LV VLPSO\ WKH DGGLWLRQ RI WKH )' IDQ GHPDQG ZLWK WKH WULP LV SUHVHQWHG WKH )' IDQ FRQWURO ORJLF ZKHUH D 3,' FRQWURO ORRS PDLQWDLQV GHVLUHG DLUIORZ 1RWLFH WKDW DOO RI WKHVH UHSUHVHQWDWLRQV DUH LQ RXU HVVHQWLDO WDJ OLVW PDUNHG DV 09V &OHDUO\ WKHVH YDULDEOHV DUH QRW DOO LQGHSHQGHQW ,Q DGGLWLRQ WR f7\SHf \RX ZLOO QRWLFH D ILHOG LQ WKH (VVHQWLDO 7DJ /LVW FDOOHG f*URXSf 7KLV ILHOG ZLOO EH XVHG WR LGHQWLI\ WDJV ZKLFK UHSUHVHQW WKH VDPH SK\VLFDO SURFHVV YDULDEOHV 7KH IROORZLQJ VHFWLRQV SUHVHQW D

PAGE 114

PHWKRGRORJ\ IRU UHGXFLQJ WKH 09V '9V DQG 69V GRZQ WR D PLQLPDO VHW ZKLFK KDYH WKH JUHDWHVW LPSDFW RQ FRPEXVWLRQ $OO RI WKH YDULDEOH IURP D JURXS FDQ EH UHPRYHG IURP FRQn VLGHUDWLRQ EXW DW PRVW RQH YDULDEOH IURP HDFK JURXS ZLOO EH DOORZHG LQ RXU ILQDO YDULDEOH VHWV 'DWDVHWV 6HFWLRQ SUHVHQWHG WKH VDPSOLQJ PHWKRGRORJ\ IRU UHWULHYLQJ SURFHVV GDWD IURP WKH GDWD KLVWRULDQ 7KLV PHWKRGRORJ\ ZDV DSSOLHG WR SURGXFH D FRQWLJXRXV WLPH VHULHV RI PRQWKV RI SURFHVV GDWD IURP WR ZKLOH WKH XQLW ZDV RSHUDWLQJ FRQWLQXRXVO\ 7KLV GDWD DOVR LQFOXGHG SDUDPHWULF WHVWLQJ RI NH\ 09V WKRXJKW WR DIIHFW WKH SURFHVV RXWn SXWV RI LQWHUHVW 7KH PRQWKV RI DYDLODEOH GDWD KDG WR EH GLYLGHG LQWR GDWDVHWV IRU WUDLQLQJ FURVVYDOLn GDWLRQ DQG WHVWLQJ RI HDFK RI WKH PRGHOV 'XH WR WKH WHPSRUDO QDWXUH RI WKH GDWD DQG WKH G\QDPLF QDWXUH RI WKH PRGHOV EHLQJ FRQVLGHUHG WKH GDWDVHWV ZRXOG HDFK KDYH WR EH FRQWLJn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

PAGE 115

' WLPHV SURGXFLQJ WUDLQLQJ GDWDVHWV DV REVHUYDWLRQV RI WKH XQGHUO\LQJ SURn FHVV 7KHVH REVHUYDWLRQV ZHUH VDPSOHG E\ RIIVHWWLQJ WKH VWDUW RI HDFK VDPSOLQJ E\ VHFRQGV IURP WKH SUHYLRXV ZKHUH s VHFRQGV LV WKH DSSUR[LPDWH UHVROXWLRQ RI WKH GDWD KLVWRULDQ 3HUIRUPDQFH &ULWHULD 7KH REMHFWLYH RI WKLV VHFWLRQ LV WR GHYHORS WKH fEHVWf SRVVLEOH PRGHOV EDVHG RQ WKH PRGHO GHILQLWLRQV SUHVHQWHG LQ 6HFWLRQ 7R WKLV HQG PDQ\ PRGHOV ZLOO EH GHYHORSHG IRU HDFK RI WKH VL[ PRGHO GHILQLWLRQV WKHVH PRGHOV ZLOO WKHQ FRPSHWH IRU EHLQJ WKH fEHVWf IRU D SDUWLFXODU PRGHO GHILQLWLRQ 7KH fEHVWf PRGHOV ZLOO WKHQ EH XVHG LQ &KDSWHU 7KLV VHFWLRQ ZLOO IRUPDOO\ SUHVHQW WKH FULWHULD E\ ZKLFK WKH PRGHOV ZLOO EH MXGJHG &OHDUO\ WKHVH FULWHULD VKRXOG EH FORVHO\ UHODWHG WR WKH REMHFWLYH IXQFWLRQ XVHG WR WUDLQ WKH PRGHOV 2EMHFWLYH )XQFWLRQ $OO RI WKH PRGHO GHILQLWLRQV FRQVLGHUHG DUH IRU SUHGLFWRUV LH WKH\ VROYH WKH JHQHUDOn L]H QRQOLQHDU UHJUHVVLRQ SUREOHP 0RGHO WUDLQLQJ ZLOO WKHUHIRUH IDOO LQWR WKH FDWHJRU\ RI VXSHUYLVHG OHDUQLQJ ZKHUH WKHUH LV D NQRZQ GHVLUHG UHVSRQVH IRU WKH PRGHO GWf H $OWKRXJK WKHUH DUH PDQ\ REMHFWLYH IXQFWLRQ ZKLFK FDQ EH DSSOLHG WR WKLV SUREOHP WKH RUGLQDU\ PHDQVTXDUHGHUURU REMHFWLYH IXQFWLRQ LV E\ IDU WKH PRVW FRPPRQ DQG VXFFHVVIXO 7KH 06( REMHFWLYH IXQFWLRQ ZLOO EH JLYHQ E\ A=.ff 1 f

PAGE 116

M\O ZKHUH \Wf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n DEOH FDQGLGDWH IRU RXU fEHVWf PHWULF 7KHUH DUH D IHZ SUREOHPV ZLWK FRPSDULQJ PRGHOV EDVHG RQ 06( KRZHYHU f ,W LV GLIILFXOW WR GHULYH PHDQLQJ IURP WKH 06( YDOXH DVVRFLDWHG ZLWK WHVWLQJ DQ LQGLYLGXDO PRGHO f ,W FDQQRW EH XVHG WR FRPSDUH WKH SHUIRUPDQFH EHWZHHQ LQGLYLGXDO RXWn SXWV EHFDXVH WKH 06( LV VLJQLILFDQWO\ DIIHFWHG E\ WKH SRZHU RI WKH LQGLn YLGXDO RXWSXW YDULDEOHV f ,W FDQQRW EH XVHG WR FRPSDUH WKH SHUIRUPDQFH RI PRGHOV WHVWHG RQ VHSDn UDWH GDWDVHWV VLQFH WKH SRZHU RI HDFK YDULDEOH ZLOO EH KLJKO\ YDULDEOH DFURVV GDWDVHWV 6LQFH DOO RI WKH PRGHOV ZLOO EH WHVWHG DJDLQVW D FRPPRQ GDWDVHW WKH WKLUG SUREOHP ZLOO QRW EH D IDFWRU KHUH 7KH PRVW FRPSHOOLQJ UHDVRQ QRW WR XVH 06( DV D SHUIRUPDQFH FULWHn ULRQ LV VLPSO\ RQH RI LQWHUSUHWDWLRQ ,W LV YHU\ GLIILFXOW WR WHOO KRZ ZHOO D PRGHO LV SHUIRUPn LQJ UHODWLYH WR WKH SURFHVV :H ZLOO LQVWHDG LQWURGXFH WZR UHODWHG SHUIRUPDQFH FULWHULD ZKLFK ZLOO QRW RQO\ VHUYH DV D PHWULF ZLWK ZKLFK WR FRPSDUH PRGHOV DJDLQVW RQHDQRWKHU EXW ZKRVH YDOXHV SRVVHVV VLPSOH SK\VLFDO LQWHUSUHWDWLRQ 1RUPDOL]HG 0HDQ6TXDUHG (UURU 7KH ILUVW PHWULF LV D FORVH UHODWLYH WR WKH 06( FDOOHG WKH QRUPDOL]HG PHDQVTXDUHG HUURU 106(f 7KLV PHWULF QRUPDOL]HV WKH 06( UHODWLYH WR ZKDW LV VRPHWLPHV FDOOHG WKH

PAGE 117

fWULYLDO SUHGLFWRUf 7KLV SUHGLFWRU LV VLPSO\ DQ HVWLPDWH IRU WKH SURFHVV VWDWLVWLFDO PHDQ DQG WKH 106( LV JLYHQ E\ 7 106(\ ; W 02f OE72ff f ZKHUH SI LV WKH L WK HOHPHQW RI WKH YHFWRU IWA ZKLFK LV WKH VWDWLVWLFDO PHDQ RI WKH GHVLUHG UHVSRQVH 1RWLFH WKDW WKH 106( LV EDVHG RQ LQGLYLGXDO RXWSXW YDULDEOHV ZKLFK DOORZV XV WR FRPSDUH WKH PRGHO SHUIRUPDQFH IRU VSHFLILF PRGHO RXWSXWV 7KH PDLQ DGYDQWDJHV WR WKH 106( RYHU WKH 06( DUH f 'LIIHUHQFHV LQ WKH SRZHU RI LQGLYLGXDO RXWSXWV LV QRUPDOL]HG RXW f 6LPSOH LQWHUSUHWDWLRQ FDQ EH XVHG WR SURYLGH D IHHO IRU PRGHO SHUIRUn PDQFH EDVHG RQ WKH 106( YDOXH DORQH 7KH EHVW SHUIRUPDQFH D PRGHO FDQ KDYH ZRXOG EH WR PDWFK WKH GHVLUHG UHVSRQVH SUHFLVHO\ ZKLFK ZRXOG UHVXOW LQ D 106( 6LPLODUO\ D 106( LQGLFDWHV WKDW WKH PRGHO LV GRLQJ QR EHWWHU WKDQ SUHGLFWLQJ WKH SURFHVV PHDQ :KLOH WKLV LV FOHDUO\ QRW WKH ZRUVW SHUIRUPDQFH D PRGHO FDQ KDYH LW LV FHUWDLQO\ FDXVH IRU VSHFXODWLRQ &RUUHODWLRQ %RWK WKH 06( DQG 106( PHWULFV UHSUHVHQW WKH PRGHOfV HUURU ZLWK UHVSHFW WR WKH GHVLUHG UHVSRQVH 7KH HUURU RI D PRGHO GHVFULEHV KRZ ZHOO LW PDWFKHV WKH DFWXDO YDOXH RI WKH GHVLUHG UHVSRQVH ,W LV YHU\ SRVVLEOH IRU WKH PRGHO WR KDYH D ODUJH HUURU DQG VWLOO FRQn WDLQ YDOXDEOH LQIRUPDWLRQ DERXW WKH SURFHVV 7KLV LV SDUWLFXODUO\ WUXH ZKHQ XVLQJ PRGHOV IRU RSWLPL]DWLRQ DQG FRQWURO DSSOLFDWLRQV &RQVLGHU WKH H[DPSOH LOOXVWUDWHG LQ )LJXUH ,W LV FOHDU WKDW WKH 106( DQG 06( EHWZHHQ YDULDEOH G DQG PRGHO P ZLOO EH ORZHU WKDQ WKH UHVSHFWLYH HUURUV EHWZHHQ G DQG P ‘ ,W LV DOVR FOHDU WKDW P KDV FDSWXUHG PRUH RI WKH G\QDPLF LQIRUPDWLRQ RI G

PAGE 118

7KLV H[DPSOH LOOXVWUDWHV WKH QHHG IRU DQ DGGLWLRQDO PHWULF :H DUH EXLOGLQJ PRGHOV ZKLFK DUH QRW JRLQJ WR EH XVHG IRU WKHLU DELOLW\ WR IRUHFDVW WKH DFWXDO YDOXH RI D YDULDEOH EXW UDWKHU WR H[SODLQ WKH FDXVHDQGHIIHFW UHODWLRQVKLSV EHWZHHQ WKH PRGHOfV LQSXW DQG RXWn SXW YDULDEOHV )RU VXFK DSSOLFDWLRQV FRUUHODWLRQ SURYLGHV D EHWWHU PHWULF 7KH FRUUHODWLRQ RI D PRGHO FDQ EH FDOFXODWHG DV IROORZV ZKHUH f 7 "\ A:2Qn;UW29f f W $ 5a VXJJHVW WKDW WKHUH LV QR OLQHDU UHODWLRQVKLS EHWZHHQ WKH PRGHOfV RXWSXW DQG \c WKH GHVLUHG UHVSRQVH ZKLOH 5a VXJJHVWV WKDW WKH\ DUH LGHQWLFDO LQ D OLQHDU VHQVHf %RWK 106( DQG 5 SURYLGH XVHIXO PHWULF ZLWK ZKLFK WR MXGJH PRGHO SHUIRUPDQFH ,Q IDFW XVHG WRJHWKHU WKH\ SURYLGH GLUHFW LQVLJKW LQWR WKH ELDVYDULDQFH GLOHPPD RI PRGHO GHYHORSPHQW 7KHVH ZLOO EH WKH SULPDU\ PHWULF XVHG WR DVVHV PRGHO SHUIRUPDQFH LQ WKLV VHFWLRQ

PAGE 119

/HDUQLQJ $OJRULWKP 7KH OHDUQLQJ DOJRULWKP RU DOJRULWKPV IRU D QHXUDOQHWZRUNEDVHG FRQWURO DSSOLFDWLRQ PXVW EH DEOH WR GHDO ZLWK WKH IROORZLQJ FRQILJXUDWLRQV f 7UDLQLQJ ZKHUH WKH LQLWLDO SDUDPHWHUV RI PRGHOV DQGRU FRQWUROOHUV DUH GHWHUPLQHG IRU WKH ILUVW WLPH f 5HWXQLQJ ZKHUH WKH PRGHO DQGRU FRQWUROOHU SDUDPHWHUV DUH DGMXVWHG RQOLQH WR DOORZ WKH V\VWHP WR FRPSHQVDWH IRU FKDQJHV LQ LWV RSHUDWLQJ HQYLURQPHQW f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f VHFRQGRUGHU DSSUR[LPDWLRQ PHWKRGV UHTXLULQJ OLQH VHDUFKHV 7KLV FRQn FHUQV SDUWLFXODUO\ WKH FRQMXJDWH JUDGLHQW PHWKRGV DQG 3RZHOOfV DOJRn ULWKP 7KHVH DOJRULWKPV DUH QRW DEOH WR NHHS WKH VHDUFK GLUHFWLRQV FRQMXJDWH ZLWKRXW OLQH VHDUFKHV f PRVW JOREDO RSWLPL]DWLRQ PHWKRGV ([FHSW IRU VLPXODWHG DQQHDOLQJ DQG HYROXWLRQDU\ DOJRULWKPV JOREDO RSWLPL]DWLRQ DOJRULWKPV UHTXLUH WKH FRVW IXQFWLRQ WR EH HYDOXDWHG DW YDULRXV SRLQWV RI WKH VWDWH VSDFH ZLWK WKH UHVXOWV FRPSDUHG

PAGE 120

7KH FRPSXWDWLRQDO FRPSOH[LW\ ORVW E\ QRW XVLQJ VHFRQGRUGHU DSSUR[LPDWLRQV LV E\ LWVHOI UHDVRQ HQRXJK WR UHVWULFW RXU OHDUQLQJ WR EDWFK VFKHPHV )RU ILUVWRUGHU PHWKRGV WKHUH DUH QR HVWLPDWHV RI FRQYHUJHQFH HYHQ IRU H[DFW TXDGUDWLF IXQFWLRQV ,Q RWKHU ZRUGV WKH\ FDQ FRQYHUJH DUELWUDULO\ VORZO\ 6HFRQGRUGHU PHWKRGV FDQ EH FRPSDUHG E\ WKH QXPn EHU RI FRVW IXQFWLRQ HYDOXDWLRQV WKDW DUH QHFHVVDU\ WR UHDFK WKH PLQLPXP RI D TXDGUDWLF IXQFWLRQ +U\FHV >@ GHPRQVWUDWHV KRZ WKH FRQMXJDWH JUDGLHQWV DOJRULWKP FDQ EH DSSOLHG WR VSHHG XS ERWK WKH EDFNSURSDJDWLRQ DQG EDFNSURSDJDWLRQ WKURXJK WLPH DOJRULWKPV 7KHVH DFFHOHUDWLRQV XVH RQO\ Q. f FRVW IXQFWLRQ HYDOXDWLRQV ZKHUH Q LV WKH QXPEHU RI IUHH SDUDPHWHUV DQG LV WKH QXPEHU RI HYDOXDWLRQV UHTXLUHG E\ WKH OLQH VHDUFK ,W VKRXOG EH QRWHG WKDW WKH ILUVWRUGHU JUDGLHQW GHVFHQW LV QRW WKH ODVW UHVRUW IRU LQFUHn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n ]DWLRQ 7KH UREXVWQHVV RI D QHXURFRQWUROOHU VROXWLRQV ZLOO EH SUHGRPLQDWHO\ D IXQFWLRQ RI LWV SHUIRUPDQFH GXULQJ WUDLQLQJ UHWXQLQJ DQG RSWLPL]DWLRQ &RQVLGHU WKH FRQVHTXHQFHV RI f 6KRZLQJ XS DQG WU\LQJ WR ZLQ RSHUDWRU FRQILGHQFH ZLWK SRRUO\ WUDLQHG PRGHOV

PAGE 121

,OO f +DYLQJ PRGHOV ZKLFK KDG ZRUNHG SURSHUO\ EHLQJ SODFHG LQ FORVHGORRS FRQWURO RI D SRZHU SODQW ZKLFK VXGGHQO\ VWRS ZRUNLQJ IROORZLQJ D UHWUDLQLQJ f $ FRQWUROOHU QRW DEOH WR VHH DQ RSWLPDO FRQWURO WUDMHFWRU\ EHFDXVH RI WKH QRQFRQYH[ QDWXUH WR WKH SHUIRUPDQFH VXUIDFH ZKLFK H[LVW EHWZHHQ WZR GLVWLQFW RSHUDWLQJ PRGHV RI WKH XQLW 6LQFH WKH OHDUQLQJ VFKHPHV GHVFULEHG DERYH ZLOO EH DSSOLHG WR KLJKO\ FRPSOH[ QRQOLQn HDU V\VWHPV DQG WKXV KLJKO\ QRQFRQYH[ SHUIRUPDQFH VXUIDFHV JOREDO RSWLPL]DWLRQ ZLOO EH FRQVLGHUHG D UHTXLUHPHQW 7KH QRWDWLRQ $UJ0LQ^-` VKDOO EH XVHG WR UHIHU WR WKH JORn EDO RSWLPL]DWLRQ RI FRVW IXQFWLRQDO ZLWK UHVSHFW WR WKH YDULDEOHV 7KH JOREDO RSWLPL]Dn WLRQ DOJRULWKPV XVHG WKURXJKRXW WKH UHVW RI WKLV ZRUN DUH GHWDLOHG QH[W *OREDO 2SWLPL]DWLRQ $OJRULWKP $Q H[WUHPXP RI D FRPSOHWHO\ DUELWUDU\ IXQFWLRQ RI UHDOYDOXHG DUJXPHQWV LV QRW FRPn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n QHUV RI D K\SHUFXEH
PAGE 122

)RU WUDLQLQJ DQG UHWXQLQJ OHDUQLQJ WDVNV WKHVH PHWKRGV DUH VWLOO IDU WRR FRPSXWDWLRQn DOO\ EXUGHQVRPH )RU WKHVH FRQILJXUDWLRQV ZH ZLOO KDYH WR EH VDWLVILHG ZLWK VLPSO\ JHQHUn DWLQJ ,9V VWDUWLQJ YHFWRUV XQLIRUPO\ )RU RSWLPL]DWLRQ OHDUQLQJ WDVNV KRZHYHU WKHUH DUH IDU IHZHU IUHH SDUDPHWHUV RQH IRU HDFK 09 )RU WKLV FRQILJXUDWLRQ JOREDO RSWLPL]DWLRQ VKDOO EH GHWHUPLQHG E\ UXQQLQJ 1n ORFDO RSWLPL]DWLRQV IURP VWDUWLQJ YHFWRUV [ ZLWK HOHPHQWV L M L RWKHUZLVH f /RFDO 2SWLPL]DWLRQ $OJRULWKP 7KH FRQMXJDWH JUDGLHQWV DOJRULWKPV XVHG IRU DOO ORFDO RSWLPL]DWLRQ LV JLYHQ E\ *LYHQ D VHW RI ZHLJKWV :H GHWHUPLQH D VHDUFK GLUHFWLRQ :H XVLQJ WKH 3RON5LEHUH DOJRULWKP SUHVHQWHG LQ 6HFWLRQ ,, 2SWLPL]H WKH REMHFWLYH IXQFWLRQ :H O:Hf ZLWK UHVSHFW WR WKH VFDn ODU YDULDEOH XVLQJ WKH OLQH VHDUFK DOJRULWKP SUHVHQWHG LQ 6HFWLRQ ,,, 8SGDWH WKH ZHLJKWV DFFRUGLQJ WR O3nH ? ,9 5HSHDW XQWLO HLWKHU H HfD[ RU :Hf -^AH ?f $IQ ZKHUH H LV WKH LQGH[ RI LWHUDWLRQ HnD[ LV WKH PD[LPXP QXPEHU RI HSRFKV DQG $ Mf LV WKH PLQLPXP SHUIRUPDQFH LPSURYHPHQW IRU HDUO\ VWRSSLQJ

PAGE 123

9DULDEOH 3UXQLQJ $SSHQGL[ LGHQWLILHG WKH LQLWLDO WDJ VHWV WR EH XVHG DV 09V '9V DQG 69V 7KH JRDO RI WKLV VHFWLRQ LV WR GHYHORS D PHWKRGRORJ\ IRU SUXQLQJ WKHVH WDJ VHWV LQWR WKH VPDOOHVW VHW RI YDULDEOHV UHTXLUHG WR SURGXFH DFFXUDWH PRGHOV IRU WKH PRGHO GHILQLWLRQV SUHVHQWHG LQ 6HFn WLRQ $W WKLV SRLQW ZH KDYH H[KDXVWHG RXU ILUVWSULQFLSOHV XQGHUVWDQGLQJ RI WKH SURFHVV ZKLFK ZDV XVHG WR FUHDWH WKH (VVHQWLDO 7DJ /LVW 7KH YDULDEOH SUXQLQJ PHWKRGRORJ\ ZLOO WKHUHIRUH FRQVLVW RI WUDLQLQJ PXOWLSOH PRGHOV ZLWK GLIIHUHQW FRPELQDWLRQV RI LQSXWV DQG VHOHFWLQJ WKH EHVW PRGHOV EDVHG RQ WKH SHUIRUPDQFH FULWHULD DERYH 7KH PHWKRGRORJ\ LV DQ HGXFDWHG GLUHFW VHDUFK DSSURDFK 7KHUH DUH D IHZ UXOHV DQG DVVXPSWLRQV WKDW ZLOO EH XVHG WR OLPLW WKH VHDUFK f 2QO\ RQH YDULDEOH IURP HDFK JURXS ZLOO EH DOORZHG IRU HDFK W\SH f 2SHUDWRUV GR QRW ZDQW WR PDQLSXODWH PRUH WKDQ f 9DULDEOHV IRU WKH EHVW VWHDG\VWDWH PRGHO ZLOO EH WKH EHVW YDULDEOHV IRU G\QDPLF PRGHOLQJ f 9DULDEOHV IRU WKH EHVW 0/3f PRGHO ZLOO EH WKH EHVW YDULDEOHV IRU DOO PRGHO DUFKLWHFWXUHV FRQVLGHUHG +HUH 0/3f LV GHILQHG DV D 0/3 ZLWK KLGGHQ OD\HUV ZLWK SURFHVVLQJ HOHn PHQWV LQ WKH ILUVW OD\HU DQG LQ WKH VHFRQG 7KHVH DVVXPSWLRQV DUH LQKHUHQWO\ IODZHG *LYHQ OLPLWHG FRPSXWDWLRQDO UHVRXUFHV KRZHYHU WKH\ DUH UHDVRQDEOH DVVXPSWLRQV 7KH YDULDEOH SUXQLQJ PHWKRGRORJ\ ZLOO FRQVLVW RI WZR DOJRULWKPV %RWK DOJRULWKPV XVH D 0/3f XVLQJ WKH VWHDG\VWDWH &9 PRGHO GHILQLWLRQV ZLWK RQH PRGLILFDWLRQ WKH &2 YDULDEOH KDV EHHQ PRYHG IURP WKH 69 VHW WR WKH &9 VHW LH IURP WKH PRGHOV LQSXW WR LWV RXWSXW 7KLV PRGLILFDWLRQ ZDV PDGH WR IRUFH WKH SUXQLQJ DOJRULWKPV WR FRQVLGHU HDFK YDULn DEOHV HIIHFW RQ ERWK 12[ DQG &2

PAGE 124

*URXS 3UXQLQJ $OJRULWKP 7KH JRDO RI WKH JURXS SUXQLQJ DOJRULWKP LV WR LGHQWLI\ ZKLFK YDULDEOH JURXSV LQ WKH (VVHQWLDO 7DJ /LVW KDYH WKH JUHDWHVW LPSDFW RQ WKH SHUIRUPDQFH RI WKH PRGHO 7KH JURXS SUXQLQJ DOJRULWKP ZDV LPSOHPHQWHG DV IROORZV 7UDLQ WKH 0/3f Afn WLPHV IRU HDFK XQLTXH JURXS LQ WKH (VVHQn WLDO 7DJ /LVW ,, )RU HDFK JURXS PRGHO VHOHFW WKH fEHVW JURXSf PRGHO DV WKDW ZLWK WKH ORZHVW FURVVYDOLGDWLRQ 06( ,,, &DOFXODWH WKH FRUUHODWLRQ RI HDFK fEHVW JURXSf PRGHO DFURVV WKH EOLQG WHVWLQJ GDWDVHW DQG GHILQH WKH fUHPRYDO JURXSf PRGHO DV WKH fEHVW JURXSf PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ ,9 5HPRYH DOO YDULDEOHV IURP WKH (VVHQWLDO 7DJ /LVW WKDW DUH LQ WKH JURXS DVVRFLDWHG ZLWK WKH fUHPRYDO JURXSf PRGHO 9 5HSHDW WKLV VWHSV ,9 XQWLO WKHUH DUH RQO\ 1VnRXSV XQLTXH JURXSV UHPDLQLQJ 7KLV DOJRULWKP ZDV UXQ ZLWK 1XV DQG @?IJnrLSV 1RWLFH WKDW WKHUH DUH D WRWDO RI XQLTXH JURXS PRGHOV LJQRULQJ 12[ DQG &2f UHVXOWLQJ LQ PRGHO WUDLQLQJV IRU WKH ILUVW LWHUDWLRQ 7KLV UHVXOWV LQ fEHVW JURXSf PRGHOV 7KH WRWDO QXPEHU RI PRGHO WUDLQLQJV UHTXLUHG E\ WKLV DOJRULWKP IRU DOO LWHUDWLRQV LV WKHUHIRUH f f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

PAGE 125

UHVHDUFK WRSLF KRZHYHU DQG QRW WKH UHVHDUFK WRSLF FKRVHQ IRU WKLV ZRUN 7KH fEUXLW IRUFHf PHWKRG KDV EHHQ VHOHFWHG EHFDXVH RI LWV UREXVWQHVV WR SDUDPHWHUL]DWLRQ 7KH FRPSXWDWLRQDO FRPSOH[LWLHV RI WKLV DOJRULWKP DQG WKH DOJRULWKPV GLVFXVVHG LQ VXEn VHTXHQW VHFWLRQV DUH PDQDJHG XVLQJ D GLVWULEXWHG FRPSXWLQJ HQYLURQPHQW 'LVWULEXWLQJ WKH WUDLQLQJ RI D QHXUDO QHWZRUN LV \HW DQRWKHU UHVHDUFK SURMHFW 5DWKHU WKDW GLVWULEXWLQJ WKH WUDLQLQJ RI D VLQJOH WUDLQLQJ UXQ WKH PXOWLSOH WUDLQLQJ UXQV ZHUH GLVWULEXWHG DFURVV D QHWn ZRUN RI FRPSXWHUV 7KH DXWKRU KDG HYHQLQJ DFFHVV WR PRUH WKDQ LQWHO EDVHG ZRUNVWDn WLRQV FRQQHFWHG WKURXJK D ORFDO DUHD QHWZRUN 7KHVH PDFKLQHV ZHUH XVHG LQ SDUDOOHO WR UXQ WKLV DQG VXEVHTXHQW DOJRULWKPV 7KLV DOJRULWKP IRU H[DPSOH ZDV UXQ RQ D VLQJOH 6DWXUn GD\ ZKHQ WKH PDFKLQHV ZHUH QRW LQ XVH 7KH UHVXOWV RI WKLV DOJRULWKP DUH SURYLGHG LQ )LJXUH 7KH DOJRULWKP ZDV DEOH WR UHPRYH JURXSV IURP WKH PRGHO ZLWKRXW GHJUDGLQJ LWV SHUIRUPDQFH ,Q IDFW WKHUH LV DQ

PAGE 126

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n JDWH WKHLU LPSDFW IXUWKHU 7KH\ DUH D JRRG FDQGLGDWH IRU IXWXUH H[SDQVLRQ RI RXU FRQWUROn OHUV KRZHYHU 5HSUHVHQWDWLRQ 3UXQLQJ $OJRULWKP 7KH JRDO RI WKH UHSUHVHQWDWLRQ SUXQLQJ DOJRULWKP LV WR LGHQWLI\ WKH EHVW YDULDEOH UHSUHn VHQWDWLRQ ZLWKLQ HDFK UHPDLQLQJ JURXS 7KH UHSUHVHQWDWLRQ SUXQLQJ DOJRULWKP ZDV LPSOHn PHQWHG DV IROORZV 6HOHFW WKH ILUVW JURXS LQ WKH UHGXFHG (VVHQWLDO 7DJ /LVW ,, ,I WKHUH LV RQO\ RQH XQLTXH UHSUHVHQWDWLRQ LQ WKH JURXS WKHQ VHOHFW WKH QH[W JURXS DQG JRWR VWHS ,, ,,, )RU HDFK XQLTXH UHSUHVHQWDWLRQ LQ WKH VHOHFWHG JURXS WUDLQ WKH 0/3f WLPHV ZLWK WKH YDULDEOHV IRU WKLV UHSUHVHQWDWLRQ DV WKH RQO\ LQSXW YDULDEOHV IURP WKH VHOHFWHG JURXS 7KH UHVXOWLQJ PRGHOV ZLOO EH FDOOHG fUHSUHVHQWDWLRQf PRGHOV

PAGE 127

,9 )RU HDFK fUHSUHVHQWDWLRQf PRGHO WKH WUDLQLQJ UHVXOW ZLWK WKH ORZHVW FURVVYDOLGDWLRQ 06( LV VHOHFWHG DV WKH fEHVW UHSUHVHQWDWLRQf PRGHO 9 &DOFXODWH WKH FRUUHODWLRQ RI HDFK fEHVW UHSUHVHQWDWLRQf PRGHO DFURVV WKH EOLQG WHVWLQJ GDWDVHW DQG UHPRYH DOO YDULDEOHV DVVRFLDWHG ZLWK HYHU\ fUHSUHVHQWDWLRQ PRGHOf H[FHSW IRU WKH fEHVW UHSUHVHQWDWLRQf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

PAGE 128

UHFDOO IURP &KDSWHU WKDW H[FHVV JURVV DLUIORZ LV D SULPDU\ FDXVH RI 12[ IRUPDWLRQ 7KH 09V DOVR FRQWDLQ RYHUILUH DLU 2)$f DQG JDV UHFLUFXODWLRQ *5f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n RSHG KDV EHHQ VLJQLILFDQW :H ODWWHU YHULILHG WKLV LPSDFW E\ PDQLSXODWLQJ WKHVH YDULDEOHV DW WKH SODQW $UFKLWHFWXUH 6HOHFWLRQ 7KH ILQDO YDULDEOH VHOHFWLRQV LGHQWLILHG WKURXJK WKH SUXQLQJ DOJRULWKPV LQ WKH ODVW VHFn WLRQ ZLOO EH FRQVLGHUHG WKH EHVW YDULDEOHV IRU DOO PRGHO DUFKLWHFWXUHV 7KH DXWKRU UHFRJn QL]HV WKDW WKH EHVW YDULDEOHV FRXOG EH D IXQFWLRQ RI WKH PRGHO DUFKLWHFWXUH 7KH FRPSXWDWLRQDO FRPSOH[LW\ LQYROYHG LQ YDULDEOH SUXQLQJ KRZHYHU PDNHV ILQGLQJ WKH WRSRORJ\ VSHFLILF YDULDEOHV LPSUDFWLFDO

PAGE 129

7KHLU DUH D QXPEHU RI DUFKLWHFWXUH VSHFLILF SDUDPHWHUV WKDW ZLOO UHTXLUH VHOHFWLRQ IRU HDFK DUFKLWHFWXUH KRZHYHU ,GHQWLI\LQJ WKH EHVW DUFKLWHFWXUH SDUDPHWHUV ZLOO HPSOR\ D VLPLODU PHWKRGRORJ\ WR WKDW XVHG IRU YDULDEOH SUXQLQJ LH XVLQJ D GLUHFW VHDUFK RI WKH IHDn VLEOH SDUDPHWHU VSDFH 7KH IROORZLQJ SUHVHQWV D GHWDLOHG GHVLJQ IRU WKH PRGHO DUFKLWHFWXUHV LQYHVWLJDWHG LQ WKH VWXG\ DORQJ ZLWK WKH UHVXOWV XVHG WR LGHQWLI\ WKHLU EHVW SDUDPHWHUL]Dn WLRQ $50$ 0RGHO 7KH $50$ PRGHO FRQVLGHUHG LQ WKLV VWXG\ LV JLYHQ E\ 1@A7 1P 7 3 < DQ \WaQf < L f Q P A 1X 1 ZKHUH ^PY GY` H DUH WKH PRGHO LQSXWV DQ H \ 9Q DUH WKH DXWRUHJUHVVLYH f§A 1 FRHIILFLHQWV EP H 9P DUH WKH PRYLQJ DYHUDJH FRHIILFLHQWV 1Q LV WKH QXPEHU RI 1 DXWRUHJUHVVLYH WDSV 1P LV WKH QXPEHU RI PRYLQJ DYHUDJH WDSV DQG W H L5 Y LV D FRQVWDQW YHFWRU 3DUDPHWHU VHOHFWLRQ *LYHQ WKH EHVW VHW RI 09V DQG '9V IURP YDULDEOH SUXQLQJ DQG RQO\ D VLQJOH &9 WKHUH DUH RQO\ WZR SDUDPHWHUV WR EH LGHQWLILHG IRU WKH $50$ PRGHO 1Q DQG 1P 7KH DOJRULWKP XVHG WR GHWHUPLQH 1 ZDV DV IROORZV 6HW 1Q ,, 7UDLQ $50$1Q f ,,, ,QFUHPHQW 1Q E\ DQG UHSHDW ,, XQWLO 1Q

PAGE 130

,9 &DOFXODWH WKH FRUUHODWLRQ RI HDFK PRGHO LQ ^$50$1Q f`9L9 DFURVV r WKH EOLQG WHVWLQJ GDWDVHW DQG DVVLJQ 1Q WKH RSWLPXP QXPEHU DXWRn UHJUHVVLYH WDSV DFFRUGLQJ WR WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ )LJXUH 5HVXOWV RI DXWRUHJUHVVLYH WDSV VHDUFK DOJRULWKP IRU WKH 12[ $50$ 0RGHO )LJXUH LOOXVWUDWHV WKH UHVXOWV RI WKLV DOJRULWKP IRU WKH G&90RGHO )RU WKH G&9 0RGHO GHILQLWLRQ $50$ f ZDV FKRVHQ DV WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ WDNn LQJ LQWR FRQVLGHUDWLRQ RXU GHVLUH WR PLQLPL]H WKH QXPEHU RI WDSV 7KLV DOJRULWKP ZDV UHSHDWHG IRU WKH G690RGHO GHILQLWLRQ DQG WKH FRPELQHG UHVXOWV DUH VXPPDUL]HG LQ 7DEOH 1RWLFH WKDW RQO\ WKH G\QDPLF PRGHO GHILQLWLRQV KDYH EHHQ FRQVLGHUHG KHUH VLQFH WKH $50$ LV D G\QDPLF PRGHO

PAGE 131

7DEOH 5HVXOWV RI DXWRUHJUHVVLYH WDS VHDUFK DOJRULWKP IRU DOO G\QDPLF PRGHOV 0RGHO 3(V 7UDLQ 106( 7UDLQ 5 G&90RGHO G690RGHO *LYHQ WKH RSWLPXP QXPEHU RI DXWRUHJUHVVLYH WDSV 1Q IRU HDFK PRGHO GHILQLWLRQ D r VLPLODU DOJRULWKP ZDV XVHG WR GHWHUPLQH WKH RSWLPDO QXPEHU RI PRYLQJ DYHUDJH WDSV 1P 7KLV DOJRULWKP FDQ IRUPDOO\ EH VWDWHG DV IROORZV 6HW 1P ,, 7UDLQ WKH 0/31Q 1Pf ,,, ,QFUHPHQW 1P E\ DQG UHSHDW ,, XQWLO 1P r ,9 &DOFXODWH WKH FRUUHODWLRQ RI HDFK PRGHO LQ ^$50$1Q1Qcf`?M1P r DFURVV WKH EOLQG WHVWLQJ GDWDVHW DQG DVVLJQ 1P WKH RSWLPXP QXPEHU RI PRYLQJ DYHUDJH WDSV DFFRUGLQJ WR WKH PRGHO ZLWK WKH KLJKHVW FRUUHn ODWLRQ )LJXUH 5HVXOWV RI WKH PRYLQJ DYHUDJH WDS VHDUFK DOJRULWKP IRU 12[ $50$ 0RGHO

PAGE 132

)LJXUH LOOXVWUDWHV WKH UHVXOWV RI WKLV DOJRULWKP IRU WKH G&90RGHO )RU WKH G&9 0RGHO GHILQLWLRQ $50$ f SURGXFHG WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ 2QFH DJDLQ WKH DOJRULWKP ZDV UHSHDWHG IRU HDFK RI WKH G\QDPLF PRGHO GHILQLWLRQV DQG WKH UHVXOWV DUH VXPPDUL]HG LQ 7DEOH 7KH PRGHOV VXPPDUL]HG LQ WKLV WDEOH ZLOO EH WDNHQ DV WKH EHVW $50$ PRGHOV IRU WKHLU FRUUHVSRQGLQJ PRGHO GHILQLWLRQV 7DEOH 5HVXOWV RI PRYLQJ DYHUDJH WDS VHDUFK DOJRULWKP IRU DOO G\QDPLF PRGHOV 0RGHO 3(V 7UDLQ 106( 7UDLQ 5 G&90RGHO G690RGHO 0XOWLOD\HU 3HUFHSWURQ 7KH PXOWLOD\HU SHUFHSWURQ 0/3f DUFKLWHFWXUH FRQVLGHUHG LQ WKLV VWXG\ ZLOO FRQVLVW RI KLGGHQ OD\HUV ZLWK WDQK SURFHVVLQJ HOHPHQWV 3(Vf DQG DQ RXWSXW OD\HU ZLWK D OLQHDU 3( )RUPDOO\ WKH DUFKLWHFWXUH LV JLYHQ E\ \ DnD• 6nnf EKOf E? f Y 1X VE 1K [ 1 ZKHUH X ^PY GY` H DUH WKH PRGHO LQSXWV : H LV WKH PDWUL[ RI LW 1 ZHLJKWV IRU WKH ILUVW KLGGHQ OD\HU E H n DUH WKH ELDV YDOXHV IRU WKH ILUVW KLGGHQ OD\HU e9f $L ; AK? f K 1f f 5 nLV WKH PDWUL[ RI ZHLJKWV IRU WKH VHFRQG KLGGHQ OD\HU E H + DUH WKH ELDV YDOXHV IRU WKH VHFRQG KLGGHQ OD\HU H \ LV WKH PDWUL[ RI ZHLJKWV IRU WKH 0 VHFRQG KLGGHQ OD\HU E H DUH WKH ELDV YDOXHV IRU WKH VHFRQG KLGGHQ OD\HU 1X LV WKH WRWDO QXPEHU RI 09V DQG '9V LQ WKH LQSXW OD\HU 1K@ LV WKH QXPEHU RI 3(V LQ WKH ILUVW KLG

PAGE 133

GHQ OD\HU 1K LV WKH QXPEHU RI 3(V LQ WKH VHFRQG KLGGHQ OD\HU 1 LV WKH QXPEHU RI 3(V LQ WKH RXWSXW OD\HU DQG D LV WKH WDQK ORJLVWLF IXQFWLRQ 3DUDPHWHU VHOHFWLRQ *LYHQ WKH EHVW VHW RI 09V DQG '9V IURP YDULDEOH SUXQLQJ DQG RQO\ D VLQJOH &9 WKHUH DUH RQO\ WZR SDUDPHWHUV WR EH LGHQWLILHG IRU WKH 0/3 1K DQG 1O ‘ 7KH DOJRULWKP XVHG WR GHWHUPLQH 9 ZDV DV IROORZV 6HW9 ,, 7UDLQ WKH 0/31K[ f WLPHV ,,, 6HOHFW 0/3r1K[ f DV WKH PRGHO ZLWK WKH ORZHVW FURVVYDOLGDWLRQ 06( IURP WKH WUDLQLQJ UHVXOWV ,9 ,QFUHPHQW 1K[ E\ DQG UHSHDW ,, XQWLO 1K[ 9 &DOFXODWH WKH FRUUHODWLRQ RI HDFK PRGHO LQ ^0/3r1K[ f`9$ r DFURVV WKH EOLQG WHVWLQJ GDWDVHW DQG DVVLJQ 1K c WKH RSWLPXP QXPEHU RI KLGGHQ 3(V LQ WKH ILUVW OD\HU DFFRUGLQJ WR WKH PRGHO ZLWK WKH KLJKn HVW FRUUHODWLRQ

PAGE 134

f§Df§7UDLQ 5 f§‘f§7HVW 5 3(V )LJXUH 5HVXOWV RI KLGGHQ OD\HU 3( VHDUFK DOJRULWKP IRU WKH 12[ 0/3 0RGHO )LJXUH LOOXVWUDWHV WKH UHVXOWV RI WKLV DOJRULWKP IRU WKH VV&90RGHO )RU WKH VV&9 0RGHO GHILQLWLRQ 0/3 f SURGXFHG WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ RQ WKH EOLQG WHVW VHW 7KLV DOJRULWKP ZDV UHSHDWHG IRU HDFK RI WKH VWHDG\VWDWH PRGHO GHILQLWLRQV DQG WKH UHVXOWV DUH VXPPDUL]HG LQ 7DEOH 1RWLFH WKDW RQO\ WKH VWHDG\VWHDG\ PRGHO GHILn QLWLRQV KDYH EHHQ FRQVLGHUHG KHUH VLQFH WKH 0/3 LV D VWDWLF PRGHO 7DEOH 5HVXOWV RI KLGGHQ OD\HU 3( VHDUFK DOJRULWKP IRU DOO VWHDG\VWDWH PRGHOV 0RGHO 3(V 7UDLQ 106( 7UDLQ 5 7HVW 106( 7HVW 5 VV&90RGHO VV690RGHO VV,690RGHO VV,090RGHO r *LYHQ WKH RSWLPXP QXPEHU RI KLGGHQ SURFHVVLQJ HOHPHQWV LQ WKH ILUVW OD\HU 1K IRU HDFK PRGHO GHILQLWLRQ D VLPLODU DOJRULWKP ZDV XVHG WR GHWHUPLQH WKH RSWLPDO QXPEHU RI r KLGGHQ 3(V LQ WKH VHFRQG KLGGHQ OD\HU 1K 7KLV DOJRULWKP FDQ IRUPDOO\ EH VWDWHG DV IROn ORZV

PAGE 135

, 6HW 1K ,, 7UDLQ WKH 0/31K[ 1KOf WLPHV r ,,, 6HOHFW 0/3r1K[ 1KOf DV WKH PRGHO ZLWK WKH ORZHVW FURVVYDOLGDWLRQ 06( IURP WKH WUDLQLQJ UHVXOWV ,9 ,QFUHPHQW 1KO E\ DQG UHSHDW ,, XQWLO 1K r 9 &DOFXODWH WKH FRUUHODWLRQ RI HDFK PRGHO LQ ^0/3r 1K M 1Kf` YWYL r DFURVV WKH EOLQG WHVWLQJ GDWDVHW DQG DVVLJQ 1KO WKH RSWLPXP QXPEHU RI KLGGHQ 3(V LQ WKH VHFRQG OD\HU DFFRUGLQJ WR WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ )LJXUH 5HVXOWV RI WKH KLGGHQ OD\HU 3( VHDUFK DOJRULWKP IRU 12[ 0/3 0RGHO )LJXUH LOOXVWUDWHV WKH UHVXOWV RI WKLV DOJRULWKP IRU WKH VV&90RGHO )RU WKH VV&9 0RGHO GHILQLWLRQ 0/3r f SURGXFHG WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ RQ WKH EOLQG WHVW VHW 2QFH DJDLQ WKH DOJRULWKP ZDV UHSHDWHG IRU HDFK RI WKH PRGHO GHILQLWLRQV DQG WKH UHVXOWV DUH VXPPDUL]HG LQ 7DEOH 7KH PRGHOV VXPPDUL]HG LQ WKLV WDEOH ZLOO EH WDNHQ DV WKH EHVW 0/3 PRGHOV IRU WKHLU FRUUHVSRQGLQJ PRGHO GHILQLWLRQV

PAGE 136

7DEOH 5HVXOWV RI KLGGHQ OD\HU 3( VHDUFK DOJRULWKP IRU DOO VWHDG\VWDWH PRGHOV 0RGHO 3(V 7UDLQ 106( 7UDLQ 5 7HVW 106( 7HVW 5 VV&90RGHO VV690RGHO VV,690RGHO VV,090RGHO 7LPH'HOD\HG 1HXUDO 1HWZRUN $V SUHVHQWHG LQ &KDSWHU WKH PRVW FRPPRQO\ DSSOLHG WLPHGHOD\HG QHXUDO QHWZRUN 7'11f LV VLPSO\ D 0/3 ZLWK D WDSSHGGHOD\OLQH 7'/f SUHSURFHVVRU DW LWV LQSXW 7KLV DUFKLWHFWXUH FDQ EH IRUPDOO\ SUHVHQWHG DV IROORZV S W9RMIUn*MIUOQ7'/œf 6nf Enf E\ f 1X 1X[17 ZKHUH 7'/n\L f f§} LV WKH 7'/ PDSSLQJ DQG 17 LV QXPEHU RI WDSV 3DUDPHWHU VHOHFWLRQ 7KH 7'11 DUFKLWHFWXUH WKHUHIRUH DGGV D VLQJOH DGGLWLRQDO SDUDPHWHU WR WKH 0/3 DUFKLWHFWXUH 17 &OHDUO\ WKH 0/3 SDUDPHWHUV $A DQG 1KO ZLOO EH IXQFWLRQ RI 1U 2QFH DJDLQ KRZHYHU 17 ZLOO DOVR EH IXQFWLRQ RI WKHVH SDUDPHWHUV DQG WKLV VWXG\ ZLOO VLPSOLI\ WKLV SUREOHP E\ DVVXPLQJ WKDW WKH EHVW SDUDPHWHU IRU WKH 0/3 ZLOO DOVR EH RSWLPDO IRU WKH 7'11 r *LYHQ WKH RSWLPXP QXPEHU RI KLGGHQ SURFHVVLQJ HOHPHQWV IRU WKH 0/3 1K DQG r 1L IRU HDFK PRGHO GHILQLWLRQ WKH DOJRULWKP XVHG WR GHWHUPLQH 17 ZDV LPSOHPHQWHG DV IROORZV 6HW 17 ,, 7UDLQ WKH 7'111K M 1KO 17f WLPHV

PAGE 137

,,, 6HOHFW 7'11r1A 1O 17f DV WKH PRGHO ZLWK WKH ORZHVW FURVVYDOn LGDWLRQ 06( IURP WKH WUDLQLQJ UHVXOWV ,9 ,QFUHPHQW 17 E\ DQG UHSHDW ,, XQWLO 1U 9 &DOFXODWH WKH FRUUHODWLRQ RI HDFK PRGHO LQ r r ^7'11r 1K @ 1D 1Wf`AQ DFURVV WKH EOLQG WHVWLQJ GDWDVHW DQG r DVVLJQ 1W WKH RSWLPXP QXPEHU RI WDSV DFFRUGLQJ WR WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ )LJXUH 5HVXOWV RI WDS VHDUFK DOJRULWKP IRU 12[ 7'11 0RGHO )LJXUH LOOXVWUDWHV WKH UHVXOWV RI WKLV DOJRULWKP )RU WKH G&90RGHO GHILQLWLRQ 7'11r f SURGXFHG WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ RQ WKH EOLQG WHVW VHW 7KH DOJRULWKP ZDV UHSHDWHG IRU HDFK RI WKH G\QDPLF PRGHO GHILQLWLRQV DQG WKH UHVXOWV DUH VXPPDUL]HG LQ 7DEOH 7KH PRGHOV VXPPDUL]HG LQ WKLV WDEOH ZLOO EH WDNHQ DV WKH EHVW 7'11 PRGHOV IRU WKHLU FRUUHVSRQGLQJ PRGHO GHILQLWLRQV

PAGE 138

7DEOH 5HVXOWV RI WDS VHDUFK DOJRULWKP IRU DOO G\QDPLF PRGHOV 0RGHO 7DSV 7UDLQ 106( 7HVW 106( 7UDLQ 5 7HVW 5 G&90RGHO G690RGHO *DPPD 1HXUDO 1HWZRUN 7KH *DPPD 1HXUDO 1HWZRUN *11f LV VLPLODU WR WKH 7'11 FRQVLGHUHG DERYH ZLWK WZR IXQGDPHQWDO GLIIHUHQFHV f WKH 7'/ IHHGIRUZDUG PHPRU\ PHFKDQLVP LV UHSODFHG ZLWK DQ +5 *DPPD )LOWHU *)f DQG f WKH PHPRU\ LV HPEHGGHG ZLWKLQ WKH KLGGHQ OD\HUV RI WKH DUFKLWHFWXUH 7KLV VWXG\ ZLOO FRQVLGHU DQ DUFKLWHFWXUH FRQWDLQLQJ WZR *)V RQH LQ WKH LQSXW OD\HU DQG WKH VHFRQG LQ WKH ILUVW KLGGHQ OD\HU 7KLV DUFKLWHFWXUH FDQ EH IRUPDOO\ SUHn VHQWHG DV IROORZV M! -Yn J-3Ana*cM-9n *XXf E ff E f E\ f ZKHUH *X??f§! 6?1X ; 1Jf LV WKH LQSXW OD\HU *) f§! L5An ; 1DK LV WKH KLGGHQ OD\HU *) DQG WKHVH PHPRULHV KDYH 1*X DQG 1*K PHPRU\ WDSV UHVSHFWLYHO\ 3DUDPHWHU VHOHFWLRQ 7KH *11 WKHUHIRUH DGGV WZR DGGLWLRQDO SDUDPHWHU WR WKH 0/3 DUFKLWHFWXUH 1*X DQG 1FK :H DJDLQ VLPSOLI\ WKH SDUDPHWHU VHOHFWLRQ SUREOHP E\ DVVXPLQJ WKDW WKH EHVW SDUDPHWHU IRU WKH 0/3 ZLOO DOVR EH RSWLPDO IRU WKH *11 r *LYHQ WKH RSWLPXP QXPEHU RI KLGGHQ 3(V IRU WKH 0/3 IRU HDFK PRGHO GHILQLWLRQ 1K r DQG 1O WKH DOJRULWKP XVHG WR GHWHUPLQH 1&X ZDV LPSOHPHQWHG DV IROORZV 6HW 1*X

PAGE 139

,, 7UDLQ D *01K 1K 1*X f WLPHV r r ,,, 6HOHFW *0r1K[ 1K 1&X f DV WKH PRGHO ZLWK WKH ORZHVW FURVV YDOLGDWLRQ 06( IURP WKH WUDLQLQJ UHVXOWV ,9 ,QFUHPHQW 1*X E\ DQG UHSHDW ,, XQWLO 1&X 9 &DOFXODWH WKH FRUUHODWLRQ RI HDFK PRGHO LQ r r ^*0r1c 1L 1*X f`9&Z DFURVV WKH EOLQG WHVWLQJ GDWDVHW DQG r DVVLJQ 1*X WKH RSWLPXP QXPEHU RI WDSV DFFRUGLQJ WR WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ $ $ r £ $ $ $ r $ f§Df§ 7UDLQ 5 ‘ 7HVW5 3(V )LJXUH 5HVXOWV RI WDSV VHDUFK DOJRULWKP IRU 12[ *11 PRGHO )LJXUH LOOXVWUDWHV WKH UHVXOWV RI WKLV DOJRULWKP )RU WKH G&90RGHO GHILQLWLRQ *0r f SURGXFHG WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ RQ WKH EOLQG WHVW VHW 7KLV DOJRULWKP ZDV UHSHDWHG IRU HDFK RI WKH G\QDPLF PRGHO GHILQLWLRQV DQG WKH UHVXOWV DUH VXPPDUL]HG LQ 7DEOH

PAGE 140

7DEOH 5HVXOWV RI WDSV VHDUFK DOJRULWKP IRU DOO G\QDPLF PRGHOV 0RGHO 7DSV 7UDLQ 106( 7HVW 106( 7UDLQ 5 7HVW 5 G&90RGHO G690RGHO *LYHQ WKH RSWLPXP QXPEHU RI PHPRU\ WDSV LQ WKH LQSXW OD\HU 1&X WKH RSWLPDO QXPn EHU RI PHPRU\ WDSV IRU WKH ILUVW KLGGHQ OD\HU ZHUH GHWHUPLQHG DV IROORZV 6HW 1FK ,, 7UDLQ WKH *01KO 1KO 1&X 1FKf WLPHV r r r ,,, 6HOHFW *0r1K[ 1O 1*X 1*Kf DV WKH PRGHO ZLWK WKH ORZHVW FURVV YDOLGDWLRQ 06( IURP WKH WUDLQLQJ UHVXOWV ,9 ,QFUHPHQW 1&K E\ DQG UHSHDW ,, XQWLO 1*K 9 &DOFXODWH WKH FRUUHODWLRQ RI HDFK PRGHO LQ r r r ^*0r 1K M 1L 1*X 1*Kf` YYF DFURVV WKH EOLQG WHVWLQJ GDWDVHW DQG r DVVLJQ 1*K WKH RSWLPXP QXPEHU RI PHPRU\ WDSV LQ WKH ILUVW OD\HU DFFRUGLQJ WR WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ )LJXUH 5HVXOWV RI KLGGHQ WDSV VHDUFK DOJRULWKP IRU 12[ *11 PRGHO

PAGE 141

)LJXUH LOOXVWUDWHV WKH UHVXOWV RI WKLV DOJRULWKP IRU WKH G&90RGHO )RU WKH G&9 r r 0RGHO GHILQLWLRQ *0r1KO 1K f SURGXFHG WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ RQ WKH EOLQG WHVW VHW 2QFH DJDLQ WKH DOJRULWKP ZDV UHSHDWHG IRU HDFK RI WKH PRGHO GHILQLn WLRQV DQG WKH UHVXOWV DUH VXPPDUL]HG LQ 7DEOH 7KH PRGHOV VXPPDUL]HG LQ WKLV WDEOH ZLOO EH WDNHQ DV WKH EHVW *11 PRGHOV IRU WKHLU FRUUHVSRQGLQJ PRGHO GHILQLWLRQV 7DEOH 5HVXOWV RI KLGGHQ WDSV VHDUFK DOJRULWKP IRU DOO G\QDPLF PRGHOV 0RGHO 7DSV 7UDLQ 106( 7HVW 106( 7KDLQ6OLO 7HVW 5 G&90RGHO G690RGHO 1RQOLQHDU 6WDWH6SDFH 0RGHO $V SUHVHQWHG LQ &KDSWHU WKH QRQOLQHDU VWDWHVSDFH 1/66f PRGHO DFWXDOO\ FRQVLVWV RI WZR VHSDUDWH PRGHOV D VWDWH HYROXWLRQ PRGHO DQG D RXWSXW REVHUYDWLRQ PRGHO +HUH ZH FRQVLGHU WKH FDVH ZKHUH ERWK RI WKHVH PRGHOV DUH 0/3V HDFK ZLWK D VLQJOH KLGGHQ OD\HU )RUPDOO\ WKH 1/66 PRGHO FRQVLGHUHG IRU WKLV VWXG\ LV JLYHQ E\ rIf D[. 6Zf 6 f M! 7IDA\.Wf EK\f E\ f ZKHUH UWf ^XWf[W f` H DQG VWf ^XWf [Wf` H 1r DUH VWDWH K[ A1U [ 1f 1Uf sK[ f1KU [ f$7 [ $rU s[ fM9U UHSUHVHQWDWLRQ YHFWRUV ,)HL5 6HL5 : H A5 DQG E H DUH WKH ZHLJKWV RI WKH VWDWH HYROXWLRQ PRGHO : H \ B P1K\[1X 1[f AK\ A \1K\ A;ADQG 3H;r \ DUH WKH ZHLJKWV RI WKH RXWSXW REVHUYDWLRQ PRGHO 1 LV WKH WRWDO QXPEHU RI 09V DQG '9V LQ WKH LQSXW OD\HU 1K[ LV WKH QXPEHU RI KLGGHQ 3(V LQ WKH

PAGE 142

VWDWH HYROXWLRQ PRGHO 1K\ LV WKH QXPEHU RI KLGGHQ 3(V LQ WKH RXWSXW REVHUYDWLRQ 1 LV WKH QXPEHU RI &9V LQ WKH RXWSXW DQG FW LV WKH WDQK ORJLVWLF IXQFWLRQ 7KHUH DUH WKHUHIRUH SDUDPHWHUV ZKLFK QHHG WR EH GHWHUPLQHG IRU WKH 1/66 PRGHO f WKH QXPEHU RI KLGGHQ VWDWHV 1[ f WKH QXPEHU RI KLGGHQ 3(V LQ WKH VWDWH HYROXWLRQ PRGHO 1K[ DQG f WKH QXPEHU RI KLGGHQ 3(V LQ WKH RXWSXW REVHUYDWLRQ PRGHO 1K\ 2QFH DJDLQ WKHVH SDUDPHWHUV ZLOO EH GHWHUPLQHG XVLQJ DQ H[KDXVWLYH VHDUFK PHWKRGRORJ\ :H EHJLQ E\ IL[LQJ WKH QXPEHU RI KLGGHQ 3(V LQ ERWK PRGHOV WR DQG GHWHUPLQH WKH QXPEHU RI KLGn GHQ VWDWHV 'HILQLQJ 1/661; 1K[ 1K\f DV D 1/66 PRGHO DV GHVFULEHG DERYH WKH QXPn EHU RI KLGGHQ VWDWH FDQ EH GHWHUPLQHG DV IROORZV 6HW 1[ ,, 7UDLQ WKH 1/661; f WLPHV ,,, 6HOHFW 1/66r1; f DV WKH PRGHO ZLWK WKH ORZHVW FURVVYDOLGDWLRQ 06( IURP WKH WUDLQLQJ UHVXOWV ,9 ,QFUHPHQW 1[ E\ DQG UHSHDW ,, XQWLO 1[ 9 &DOFXODWH WKH FRUUHODWLRQ RI HDFK PRGHO LQ ^1/66r 1[ f` 91 r DFURVV WKH EOLQG WHVWLQJ GDWDVHW DQG DVVLJQ 1 WKH RSWLPXP QXPEHU RI KLGGHQ VWDWHV DFFRUGLQJ WR WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ

PAGE 143

A f§f§ } D 7UDLQ 5 ‘ 7HVW 5 6WDWHV )LJXUH 5HVXOWV RI KLGGHQ VWDWHV VHDUFK DOJRULWKP IRU 12[ 1/66 PRGHO )LJXUH LOOXVWUDWHV WKH UHVXOWV RI WKLV DOJRULWKP )RU WKH G&90RGHO GHILQLWLRQ 1/66r f SURGXFHG WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ RQ WKH EOLQG WHVW VHW 7KH DOJRULWKP ZDV UHSHDWHG IRU HDFK G\QDPLF PRGHO GHILQLWLRQ DQG WKH UHVXOWV DUH VXPPDUL]HG LQ 7DEOH 7DEOH 5HVXOWV RI KLGGHQ VWDWHV VHDUFK DOJRULWKP IRU DOO G\QDPLF PRGHOV 0RGHO 6WDWHV 7UDLQ 106( 7HVW 106( 7UDLQ 5 7HVW 5 G&90RGHO G690RGHO r *LYHQ WKH RSWLPDO QXPEHU RI KLGGHQ VWDWHV 1[ WKH RSWLPDO QXPEHU RI KLGGHQ 3(V LQ WKH VWDWH HYROXWLRQ PRGHO FDQ EH GHWHUPLQHG DV IROORZV 6HW 1K[ r ,, 7UDLQ WKH 1/661; 1K[ f WLPHV

PAGE 144

,,, 6HOHFW 1/66r1; 1K[ f DV WKH PRGHO ZLWK WKH ORZHVW FURVVYDOLGDn WLRQ 06( IURP WKH WUDLQLQJ UHVXOWV ,9 ,QFUHPHQW 1K[ E\ DQG UHSHDW ,, XQWLO 1OO[ 9 &DOFXODWH WKH FRUUHODWLRQ RI HDFK PRGHO LQ ^1/66r 1[ 1K[ f` 9YOW r DFURVV WKH EOLQG WHVWLQJ GDWDVHW DQG DVVLJQ 1K[ DFFRUGLQJ WR WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ =f§f§r nr nf§r r f§,)f§ D 7UDLQ 5 f§ff§7HVW 5 6WDWH +LGGHQ 3(V )LJXUH 5HVXOWV RI VWDWH KLGGHQ 3(V VHDUFK DOJRULWKP IRU 12[ 1/66 PRGHO )LJXUH LOOXVWUDWHV WKH UHVXOWV RI WKLV DOJRULWKP )RU WKH G&90RGHO GHILQLWLRQ 1/66r f SURGXFHG WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ RQ WKH EOLQG WHVW VHW 7KH DOJRULWKP ZDV UHSHDWHG IRU HDFK RI WKH PRGHO GHILQLWLRQV DQG WKH UHVXOWV DUH VXPPDUL]HG LQ 7DEOH 7DEOH 5HVXOWV RI VWDWH KLGGHQ 3(V VHDUFK DOJRULWKP IRU DOO G\QDPLF PRGHOV 0RGHO 6WDWHV 3(V 7UDLQ 106( 7HVW 106( 7UDLQ 5 7HVW 5 G&90RGHO G690RGHO

PAGE 145

)LQDOO\ WKH RSWLPDO QXPEHU RI KLGGHQ VWDWHV 1[ DQG KLGGHQ 3(V LQ WKH VWDWH HYROXWLRQ r PRGHO 1K[ WKH RSWLPDO QXPEHU RI KLGGHQ 3(V LQ WKH RXWSXW REVHUYDWLRQ PRGHO FDQ EH GHWHUPLQHG DV IROORZV 6HW 1K\ ,, 7UDLQ WKH 1/661; 1K[ 1K\f WLPHV ,,, 6HOHFW 1/66r1; 1K[ 1K\f DV WKH PRGHO ZLWK WKH ORZHVW FURVVYDOLn GDWLRQ 06( IURP WKH WUDLQLQJ UHVXOWV ,9 ,QFUHPHQW 1K E\ DQG UHSHDW ,, XQWLO 1 9 &DOFXODWH WKH FRUUHODWLRQ RI HDFK PRGHO LQ r r ^1/66r1; 1K[ DFURVV WKH EOLQG WHVWLQJ GDWDVHW DQG r DVVLJQ 1K DFFRUGLQJ WR WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ $ $ r $ r $ $ $ n f§Df§7UDLQ 5 f§Pf§7HVW 5 6WDWH 2XWSXW 3(V )LJXUH 5HVXOWV RI RXWSXW KLGGHQ 3(V VHDUFK DOJRULWKP IRU 12[ 1/66 PRGHO )LJXUH LOOXVWUDWHV WKH UHVXOWV RI WKLV DOJRULWKP )RU WKH G&90RGHO GHILQLWLRQ 1/66r f SURGXFHG WKH PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ RQ WKH EOLQG WHVW VHW 7KH

PAGE 146

DOJRULWKP ZDV UHSHDWHG IRU HDFK RI WKH PRGHO GHILQLWLRQV DQG WKH UHVXOWV DUH VXPPDUL]HG LQ 7DEOH 7DEOH 5HVXOWV RI RXWSXW KLGGHQ 3(V VHDUFK DOJRULWKP IRU DOO G\QDPLF PRGHOV 0RGHO 2XWSXW 3(V 7UDLQ 106( 7HVW 106( 7UDLQ 5 7HVW 5 G&90RGHO G690RGHO $QDO\VLV )LJXUH VXPPDUL]HV WKH ILQDO PRGHOLQJ UHVXOWV DIWHU DUFKLWHFWXUH VHOHFWLRQ 5HFDOO WKDW WKH REMHFWLYH RI WKLV VHFWLRQ ZDV WR ILQG WKH PRVW DFFXUDWH PRGHO IRU WKH VL[ PRGHO GHIn LQLWLRQV UHTXLUHG WR LPSOHPHQW RXU IRXU FRQWURO GHVLJQV 7KH EHVW PRGHOV LGHQWLILHG IRU HDFK GHILQLWLRQ DUH f 6WHDG\6WDWH 69 0RGHO 0/3r f f 6WHDG\6WDWH &9 0RGHO 0/3r f f 6WHDG\6WDWH ,69 0RGHO 0/3r? f f 6WHDG\6WDWH ,09 0RGHO 0/3r f f '\QDPLF 69 0RGHO *0r f f'\QDPLF &9 0RGHO *0r f

PAGE 147

$50$ 0/3 7'11 *0 1/66 ’ &9 0RGHO ‘ 69 0RGHO ’ ,69 0RGHO ’ ,09 0RGHO )LJXUH %HVW PRGHOV IRU DOO PRGHO GHILQLWLRQV E\ DUFKLWHFWXUH 7KH QH[W FKDSWHU ZLOO LPSOHPHQW HDFK FRQWURO GHVLJQ XVLQJ WKHVH PRGHOV

PAGE 148

&+$37(5 &21752//(5 ,03/(0(17$7,216 7KLV VHFWLRQ ZLOO LPSOHPHQW WKH IRXU FRQWURO GHVLJQV SUHVHQWHG LQ &KDSWHU XVLQJ WKH fEHVWf UHIHUHQFH PRGHOV GHYHORSHG LQ &KDSWHU DQG LGHQWLI\ WKH fEHVWf FRQWURO GHVLJQ IRU PHHWLQJ WKH REMHFWLYHV RXWOLQHG LQ 6HFWLRQ 3HUIRUPDQFH &ULWHULD 7KH fEHVWf FRQWUROn OHU ZLOO EH GHWHUPLQHG DV IROORZV f 8VLQJ G\QDPLF PRGHOV DV VLPXODWRUV IRU WKH SODQW FRQWUROOHU SHUIRUn PDQFH LV TXDQWLILHG RIIOLQH f $OORZLQJ WKH FRQWUROOHUV WR PDQLSXODWH DFWXDO SODQW YDOXHV FRQWUROOHU SHUIRUPDQFH LV TXDQWLILHG RQOLQH 2IIOLQH 4XDQWLILFDWLRQ 7R TXDQWLI\ WKH SHUIRUPDQFH RI WKH IRXU FRQWURO GHVLJQV RIIOLQH D FRPPRQ GDWDVHW DQG SODQW VLPXODWRU QHHG WR EH VHOHFWHG 7KH GDWDVHW VKRXOG QRW LQFOXGH GDWD WKDW ZDV XVHG GXUn LQJ WUDLQLQJ RI WKH FRQWUROOHUV RU WKHLU XQGHUO\LQJ UHIHUHQFH PRGHOV 5HFDOO WKDW D EOLQG ZHHN GDWDVHW ZDV VHW DVLGH GXULQJ PRGHOLQJ IRU WHVWLQJ DQG WKDW QRQH RI WKH PRGHOV KDYH HYHU VHHQ WKLV GDWDVHW GXULQJ WUDLQLQJ ,Q DGGLWLRQ WKH G\QDPLF SURFHVV PRGHOV FKRVHQ IRU RIIOLQH TXDQWLILFDWLRQ VKRXOG QRW EH PRGHOV WKDW ZHUH XVHG DV UHIHUHQFH PRGHOV IRU WKH DQ\ RI WKH FRQWURO GHVLJQV 6LQFH WKH EHVW PRGHOV GHYHORSHG LQ &KDSWHU ZHUH XVHG DV UHIHUn HQFH PRGHOV LH WKH PRGHOV ZLWK WKH ORZHVW FURVVYDOLGDWLRQ 06( IURP VHSDUDWH WUDLQn LQJV WKH RIIOLQH TXDQWLILFDWLRQ ZLOO XVH WKH VHFRQG EHVW PRGHOV DV WKH G\QDPLF SURFHVV PRGHOV

PAGE 149

2IIOLQH TXDQWLILFDWLRQ ZLOO DSSO\ HDFK FRQWUROOHU WR WKH G\QDPLF SURFHVV PRGHOV DFURVV WKH WHVW GDWDVHW DQG FDOFXODWH WKH DYHUDJH 12[ UHGXFWLRQ DORQJ ZLWK WKH DYHUDJH &2 SURn GXFWLRQ DERYH WKH PD[LPXP &2 FRQVWUDLQW )RUPDOO\ WKH DYHUDJH 12[ SHUFHQW UHGXFWLRQ ZLOO EH UHSRUWHG DFFRUGLQJ WR $ 12[ f§f§  12[r^Wf12[^Wf f 7Ia 7nW W ZKHUH 7c LV WKH VWDUW RI WKH WHVW GDWDVHW 7M LV LWV HQG 12[Wf LV WKH YDOXH RI 12[ DW WLPH W SUHGLFWHG E\ WKH G\QDPLF SURFHVV PRGHOV LQ UHVSRQVH WR WKH DFWXDO 09 VHWSRLQWV PYWf DQG 12[rWf LV WKH SUHGLFWHG YDOXH RI 12[ LQ UHVSRQVH WR WKH FRQWUROOHUV RSWLPDO 09 VHW SRLQWV LQY Wf 6LPLODUO\ WKH DYHUDJH &2 DERYH WKH PD[LPXP FRQVWUDLQW ZLOO EH UHSRUWHG DFFRUGLQJ WR UFRL L 7I U 7I 7  OW 7 &2rWf&2 Y &2rWf!&2 HOVH f ZKHUH &2rWf LV WKH YDOXH RI &2 DW WLPH W SUHGLFWHG E\ WKH G\QDPLF SURFHVV PRGHOV LQ UHVSRQVH WR WKH FRQWUROOHUV RSWLPDO 09 VHWSRLQWV PYrWf DQG &2nD[ LV WKH PD[ OLPLW VHW IRU &2 $OO RI WKH UHVXOWV SUHVHQWHG DUH IRU D &2 PD[LPXP RI SSP 7KH RSWLPDO 09 69 DQG &9 WUDMHFWRULHV DFURVV WKH WHVW GDWDVHW SURGXFHG E\ HDFK FRQn WUROOHU DUH HVWLPDWHG E\

PAGE 150

, 6HW RSWLPDO 09 69 DQG &9 WUDMHFWRULHV WR WKHLU DFWXDO YDOXHV Lf PYWf VYrWf VYWf FYr Wf FYWf f ,, )RU W 7c WR 7Mf§ A ‘7n L *HW WKH RSWLPDO 09V IURP WKH FRQWUROOHU PY Wf 1 cf LL 8VH WKH 69 PRGHO WR FDOFXODWH WKH UHVXOWLQJ 69V VY W f $ A LLL 8VH WKH &9 PRGHO WR FDOFXODWH WKH UHVXOWLQJ &9V FY W f LY &RS\ IRUZDUG WKH 09V WR FDUU\ IRUZDUG WKH SURFHVV VWDWH WR LQLWLDOL]H WKH QH[W VWHS f 9 1H[W W 7KH UHVXOWV RI DSSO\LQJ WKLV DOJRULWKP WR HDFK FRQWUROOHU DUH LOOXVWUDWHG EHORZ )LJXUH SUHVHQWV WKH DYHUDJH 12[ UHGXFWLRQ $12[ ZKLOH )LJXUH LOOXVWUDWHV WKH DYHUDJH &2 DERYH LWV PD[ I &2a? 7KHVH DUH QRW SDUWLFXODUO\ HQFRXUDJLQJ UHVXOWV :KLOH HDFK FRQWUROn OHU GLG PDQDJH WR UHGXFH 12[ WKH UHGXFWLRQV ZHUH TXLWH VPDOO )XUWKHUPRUH WKH FRQWUROn OHUV VHHPHG WR KDYH HYHQ OHVV HIIHFW RQ WKH &2 DERYH SSP 7KH NH\ TXHVWLRQ DW WKLV VWDJH LV WR GHWHUPLQH ZKHWKHU WKHVH UHVXOWV UHSUHVHQW f DOO RI WKH SRWHQWLDO 12[ UHGXFWLRQV LQKHUHQW WR WKH SURFHVV f D SUREOHP ZLWK WKH FRQWURO GHVLJQ RU f D SUREOHP ZLWK WKH UHIHUHQFH PRGHOV IRU WKH XQGHUO\LQJ SURFHVV

PAGE 151

%DVHOLQH 6WHDG\6WDWH 03& 6WHDG\6WDWH 0,& '\QDPLF 03& '\QDPLF 05$& Â’ $YJ 12[ 5HGXFWLRQ )LJXUH $YHUDJH 12[ UHGXFWLRQ RYHU WHVWLQJ GDWDVHW 7R KHOS DQVZHU WKLV TXHVWLRQ D VHFRQG UXQ RI WKH RIIOLQH TXDQWLILFDWLRQ DOJRULWKP ZDV UXQ 7KLV WLPH KRZHYHU WKH WHVW PRGHOV ZHUH UHSODFHG ZLWK WKH UHIHUHQFH PRGHOV XVHG WR GHYHORS HDFK FRQWUROOHU 7KHVH UHVXOWV DUH SUHVHQWHG LQ )LJXUHV DQG )LJXUH $YHUDJH &2 DERYH PD[ RYHU WHVWLQJ GDWDVHW

PAGE 152

$YJ 12[ 5HGXFWLRQ %DVHOLQH 6WHDG\6WDWH 0 3& 6WHDG\6WDWH 0,& '\QDPLF 03& '\QDPLF 05$& ’ 7UDLQ 0RGHOV ‘ 7HVW 0RGHOV )LJXUH $YHUDJH 12[ UHGXFWLRQ RYHU WHVWLQJ GDWDVHW XVLQJ WUDLQ DQG WHVW PRGHOV &OHDUO\ WKHUH LV D SUREOHP 7KH IROORZLQJ REVHUYDWLRQ FDQ EH PDGH f WKH FRQWUROOHUV DSSHDU WR EH ZRUNLQJ ILQH DQG f WKH UHIHUHQFH PRGHOV DQG WKH WHVW PRGHOV DUH SURYLGLQJ LQFRQVLVWHQW NQRZOHGJH DERXW WKH SURFHVV $YJ &2 $ERYH 0D[ )LJXUH $YHUDJH &2 DERYH PD[ RYHU WHVWLQJ GDWDVHW XVLQJ WUDLQ DQG WHVW PRGHOV

PAGE 153

&KDSWHU DGGUHVVHV WKHVH LVVXHV LQ PRUH GHWDLO EXW IRU QRZ WKH TXDQWLILFDWLRQ RI WKH FXUUHQW FRQWUROOHUV FRQWLQXHV 2QOLQH 4XDQWLILFDWLRQ 2QOLQH TXDQWLILFDWLRQ ZLOO EH UHVWULFWHG WR PHDVXULQJ WKH FRQWUROOHUV DELOLW\ WR DIIHFW WKH VWHDG\VWDWH SHUIRUPDQFH RI WKH XQLW 4XDQWLI\LQJ RQOLQH SHUIRUPDQFH ZLOO UHTXLUH UXQQLQJ RQOLQH H[SHULPHQWV ZKHUH 09V DUH PRYHG DQG WKH UHVXOWLQJ &9V PHDVXUHG (DFK H[SHULn PHQW ZLOO KDYH WR EH RI UHODWLYHO\ VKRUW GXUDWLRQ VLQFH WKH ORQJHU DQ H[SHULPHQW WDNHV WKH OHVV OLNHO\ LW LV WKDW VWHDG\VWDWH FRQGLWLRQV ZLOO EH PDLQWDLQHG ,Q RUGHU WR FRPSDUH WKH SHUn IRUPDQFH RI GLIIHUHQW FRQWUROOHUV ZKLFK ZLOO LQYDULDEOH KDYH WR SHUIRUP WKHLU DFWLRQV XQGHU GLIIHUHQW FRQGLWLRQV EDVHOLQH FRQGLWLRQV ZLOO EH HVWDEOLVKHG SULRU WR HDFK H[SHULn PHQW 7KHVH EDVHOLQH FRQGLWLRQV ZLOO EH VWDWHG LQ WHUPV RI 09V VLQFH ZH KDYH QR GLUHFW FRQWURO RYHU '9V 7R DFFRXQW IRU GLIIHUHQFHV EHWZHHQ WKH '9V EHWZHHQ LQGLYLGXDO H[SHULn PHQWV WKH 09V ZLOO EH UHWXUQHG WR WKH EDVHOLQH FRQGLWLRQV SULRU WR HDFK H[SHULPHQW (DFK RQOLQH H[SHULPHQW ZLOO IROORZ WKH IROORZLQJ SURWRFRO 7KH XQLW RSHUDWRU LV DVNHG WR EULQJ WKH XQLW WR VWHDG\VWDWH FRQGLWLRQV LH KROGLQJ DOO 09 VHWSRLQWV FRQVWDQW $IWHU WKH XQLW KDV UHDFKHG VWHDG\VWDWH PHDVXUHPHQWV DUH WDNHQ WR HVWDEOLVK EDVHOLQH FRQGLWLRQV WREDVH AEDVH PY DQG FY 7KH FRQWUROOHU LV WKHQ TXHULHG IRU 09 VHWSRLQWV PY Wf ZKLFK ZLOO EH DSSOLHG E\ WKH XQLW RSHUDWRU 1HZ VHWSRLQWV ZLOO EH UHSHDWHGO\ TXHULHG DQG DSSOLHG XQWLO WKH XQLW KDV RQFH DJDLQ UHWXUQHG WR VWHDG\VWDWH 7KH IUHTXHQF\ ZLWK ZKLFK DQ RSHUDWRU FDUULHV RXW WKLV SURFHVV ZLOO EH DW WKHLU GLVFUHWLRQ ,I WKH RSHUDWRU IHHOV WKDW LQGLYLGXDO 09 VHWSRLQWV FDQ QRW EH PDGH WKHQ WKHVH VHWSRLQWV DUH FRQVWUDLQHG DQG WKH FRQWUROOHU LV TXHULHG IRU DQRWKHU VHW RI

PAGE 154

09V :KHQ QRQH RI WKH 09 VHWSRLQWV FDQ EH PDGH RU WKH FRQWUROOHU KDV QR QHZ DGYLFH LH LW KDV VDWXUDWHG DJDLQVW LWV FRQVWUDLQWV WKHQ WKH XQLW LV DW VWHDG\VWDWH DQG WKH H[SHULn PHQW LV WHUPLQDWHG AF[S 7KH &9V DUH RQFH DJDLQ PHDVXUHG WR HVWDEOLVK H[SHULPHQW FRQGLWLRQV FY 7KH GLIn IHUHQFH EHWZHHQ WKH H[SHULPHQW DQG EDVHOLQH FRQGLWLRQV ZLOO EH FDOOHG WKH H[SHULPHQW GHOWD $AS 7KH RSHUDWRUV DUH WKHQ DVNHG WR UHWXUQ WKH 09 VHWSRLQWV EDFN WR WKHLU EDVHOLQH FY sWf&O6 FRQGLWLRQV PY DQG D WKLUG PHDVXUHPHQW LV WDNHQ WR HVWDEOLVK YDOLGDWLRQ FRQGLWLRQV s9&OO FY 7KH GLIIHUHQFH EHWZHHQ WKH H[SHULPHQW DQG YDOLGDWLRQ FRQGLWLRQV ZLOO EH FDOOHG WKH YDOLGDWLRQ GHOWD $AI 1RWLFH WKDW LI WKH YDOLGDWLRQ FRQGLWLRQV PDWFK WKH EDVHOLQH FRQGL &9 WLRQV WKHQ WKH H[SHULPHQW DQG YDOLGDWLRQ GHOWDV ZLOO KDYH LGHQWLFDO PDJQLWXGH DQG VLJQ &RQWUROOHUV DUH WKHQ FRPSDUHG EDVHG RQ WKHLU DELOLW\ WR DIIHFW WKH &9V UHODWLYH WR WKHVH EDVHOLQH YDOXHV DFURVV PXOWLSOH H[SHULPHQWV .H\ WR WKH DFFXUDF\ RI WKHVH H[SHULPHQWV ZLOO EH KRZ WKHVH PHDVXUHPHQWV DUH UHSRUWHG 7KH QH[W VHFWLRQV RXWOLQH WKH PHDVXUHPHQW PHWKRGRORJ\ DSSOLHG GXULQJ RIIOLQH TXDQWLILFDWLRQ 0HDVXUHPHQW 0HWKRGRORJ\ 2QH RI WKH PRVW FULWLFDO DVSHFW WR TXDQWLI\LQJ WKH DIIHFW WKDW GLIIHUHQW FRQWURO VWUDWHJLHV KDYH RQ D UHDO SURFHVV ZLOO EH LQ GHWHUPLQLQJ WKH VLJQLILFDQFH RI RXU UHVXOWV UHODWLYH WR SURFHVV QRLVH DQG FKDQJLQJ VWHDG\VWDWH FRQGLWLRQV 7KH IROORZLQJ RXWOLQHV WKH VWDWLVWLFDO PHDVXUHPHQW PHWKRGRORJ\ IROORZHG LQ WKLV VWXG\

PAGE 155

f§sEDVH (DFK PHDVXUHPHQW HJ FY WDNHQ RI WKH XQLWV VWHDG\VWDWH FRQGLWLRQ IRU D YDULDEOH DW WLPH WD HJ FYWDf ZLOO LPSOHPHQW WKH IROORZLQJ &ROOHFW 7 REVHUYDWLRQV ^FYWf`Wr WR ,, &DOFXODWH WKH VDPSOH PHDQ WQ7 ,,, &DOFXODWH WKH VDPSOH VWDQGDUG GHYLDWLRQ f /7 EDVH e FY$ f f W WQ (DFK GHOWD FDOFXODWHG HJ $AS RI FKDQJHV WR WKH XQLWV VWHDG\VWDWH FRQGLWLRQ ZLOO LPSOHPHQW WKH IROORZLQJ f§sEDVH f§sYDO *LYHQ PHDVXUHPHQWV IRU FY DQG FY ,, &DOFXODWH WKH ORZHU ERXQG RI D Df b FRQILGHQFH LQWHUYDO IRU WKH GLIIHUHQFH EHWZHHQ WZR VDPSOH PHDQV DFFRUGLQJ WR IeDVHDYB"nff§ KEH An >@ YUFY QFY n 7 9 FY FY / f 5HVXOWV *LYHQ WKH UHVXOWV REWDLQHG LQ 6HFWLRQ 2IIOLQH 4XDQWLILFDWLRQ WKHUH VHHPV OLWWOH KRSH RI FRQGXFWLQJ VXFFHVVIXO RQOLQH H[SHULPHQWV :KLOH RQOLQH WHVWV PLJKW EH XVHIXO LQ FRQILUPLQJ WKH RIIOLQH UHVXOWV WKHUH DUH FRQVLGHUDEOH FRVWV DVVRFLDWHG ZLWK FRQGXFWLQJ WKHP ,Q DGGLWLRQ WR WKH WLPH DQG UHVRXUFH UHTXLUHPHQWV WKHUH LV WKH LQYDOXDEOH FDSLWDO RI EX\LQ IURP RSHUDWLRQV DQG HQJLQHHULQJ WR EH FRQVLGHUHG 0RVW RI WKHVH RSHUDWRUV DQG

PAGE 156

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f§Af§ ([SHULPHQW f§ff§ 9DOLGDWLRQ $ $YHUDJH ([SHULPHQW )LJXUH &KDQJH LQ 12[ IRU VWHDG\VWDWH FRQWUROOHU H[SHULPHQWV )LJXUHV VKRZV WKH PHDVXUHG 12[ FKDQJH EHWZHHQ WKH EDVHOLQH DQG H[SHULPHQW DORQJ ZLWK WKH FRUUHVSRQGLQJ FKDQJH EHWZHHQ WKH H[SHULPHQW DQG YDOLGDWLRQ UHJLRQV :H FDQ VHH WKDW WKHUH VHHPV WR EH D VOLJKW GHFUHDVH LQ 12[ 7KH DYHUDJH 12[ SHUFHQW FKDQJH

PAGE 157

IRU DOO H[SHULPHQWV ZDV b 7KH FKDQJHV ZHUH VR VPDOO ZLWK UHVSHFW WR WKH VWHDG\VWDWH 12[ YDULDQFH KRZHYHU WKDW IHZ RI WKH H[SHULPHQWV SURYHG VLJQLILFDQW )LJXUHV VKRZV WKH FRUUHVSRQGLQJ &2 FKDQJHV +HUH LWV KDUG WR VHH DQ\ WUHQG 7KH DYHUDJH &2 SHUFHQW FKDQJH IRU DOO H[SHULPHQWV ZDV b 2QFH DJDLQ WKHVH FKDQJHV DUH VR VPDOO WKDW IHZ RI WKH H[SHULPHQWV SURYHG VLJQLILFDQW )LJXUH )LQDO &2 OHYHO IRU VWHDG\VWDWH FRQWUROOHU H[SHULPHQWV

PAGE 158

&+$37(5 3$5$0(7(5,=$7,21 352%/(0 7KH ZRUN WKXV IDU KDV UHVXOWHG LQ VRPH XQH[SHFWHG UHVXOWV $SSO\LQJ DFFHSWHG QHXUDO QHWZRUN PRGHOLQJ WHFKQLTXHV KDV SURGXFHG VHYHUDO PRGHOV IRU D SURFHVV WKDW FRQWDLQ GUDVn WLFDOO\ GLIIHUHQW fNQRZOHGJHf DERXW WKH SURFHVV +HUH NQRZOHGJH UHIHUV WR WKH FDXVHDQG HIIHFW UHODWLRQVKLSV EHWZHHQ 09V DQG &9V ZKLFK LV ZKDW HDFK RI WKH FRQWURO GHVLJQV GHSHQG XSRQ &RQWUDVW WKLV ZLWK RXU PRGHOLQJ UHVXOWV IURP 6HFWLRQ $QDO\VLV ZKHUH WKHVH VDPH PRGHOV GHPRQVWUDWHG FRQVLVWHQW DQG UREXVW fNQRZOHGJHf RI WKH SURFHVV ZKHUH NQRZOHGJH ZDV FRQVLGHUHG WR EH WKH PRGHOV DELOLW\ WR SUHGLFW RU IRUHFDVW D EOLQG GDWDVHW $OVR QRWLFH WKDW WKLV VDPH EOLQG WHVW GDWDVHW ZDV XVHG WR SHUIRUP WKH RIIOLQH TXDQWLILFDWLRQ RI WKH FRQWUROOHUV 7KLV SUREOHP LV DQDORJRXV WR WKH SDUDPHWHUL]DWLRQ SUREOHP LQ FODVVLFDO DGDSWLYH FRQn WURO WKHRU\ $Q DGDSWLYH FRQWURO V\VWHP FHQWHUV DURXQG WKH LGHD WKDW D SURFHVV LV GHVFULEHG DV D PDWKHPDWLFDO IXQFWLRQ ZLWK SDUDPHWHUV ,W PLJKW WKHUHIRUH EH H[SHFWHG WKDW WKH ZD\ SDUDPHWHUV DUH HVWLPDWHG LV HVVHQWLDO WR WKH VXFFHVV RI DQ DGDSWLYH FRQWUROOHU ,W LV XVHIXO WR YLHZ SDUDPHWHU HVWLPDWLRQ LQ WKH EURDGHU FRQWH[W RI V\VWHP LGHQWLILFDWLRQ 7KH NH\ HOHn PHQWV RI V\VWHP LGHQWLILFDWLRQ DUH WKH VHOHFWLRQ RI PRGHO DUFKLWHFWXUH H[SHULPHQW GHVLJQ SDUDPHWHU HVWLPDWLRQ DQG YDOLGDWLRQ 6LQFH V\VWHP LGHQWLILFDWLRQ LV H[HFXWHG DXWRPDWLn FDOO\ LQ DGDSWLYH V\VWHPV LW LV HVVHQWLDO WR KDYH D JRRG XQGHUVWDQGLQJ RI DOO DVSHFWV RI WKH SUREOHP 7KH HOHPHQWV RI V\VWHP LGHQWLILFDWLRQ DUH NQRZQ WR EH IXQGDPHQWDO LVVXHV LQ DGDSWLYH FRQWURO WKHRU\

PAGE 159

(YHQ WKRXJK WKLV LV D SUREOHP IURP FODVVLFDO FRQWUROV WKHUH DUH QR FODVVLFDO VROXWLRQV ZKLFK DSSO\ WR QHXURFRQWURO GHVLJQV 7KH SULPDU\ UHDVRQ IRU WKLV LV WKH XVH RI QHXUDO QHWn ZRUNV DV SURFHVV PRGHOV 1HXUDO QHWZRUNV DUH IXQGDPHQWDOO\ GLIIHUHQW WKDQ OLQHDU RU ILUVW SULQFLSOHVEDVHG SURFHVV PRGHOV LQ WKDW WKH\ DUH QRQSDUDPHWULF 7KH LQGLYLGXDO SDUDPHWHUV KDYH QR SK\VLFDO LQWHUSUHWDWLRQ LH WKH\ DUH PHDQLQJOHVV FRHIILFLHQWV LQ D EODFN ER[ :H DUH WKHUHIRUH VHHNLQJ D PHWKRG WR YDOLGDWH WKH SDUDPHWHUL]DWLRQ RI D QRQSDUDPHWULF V\VWHP ,Q FODVVLFDO DGDSWLYH FRQWURO WKHRU\ SDUDPHWHUL]DWLRQ LV D GHVLJQWLPH LVVXH WKDW LV W\SLn FDOO\ GHDOW ZLWK DQDO\WLFDOO\ :H FDQ H[SHFW WKDW WKH VROXWLRQ IRU QHXURFRQWURO WKHRU\ ZLOO OLNH PRVW RWKHU DVSHFWV RI QHXUDOQHWZRUNEDVHG V\VWHP GHVLJQ KDYH WR EH GHDOW ZLWK HPSLULFDOO\ 6HDUFK IRU D 9DOLGDWLRQ 0HWULF /HWfV EHJLQ RXU VHDUFK IRU D PHWKRGRORJ\ WR YDOLGDWH WKH SDUDPHWHUL]DWLRQ RI QHXUDO QHWZRUN PRGHOV E\ UHYLHZLQJ WKH PHWULFV XVHG WR DVVHUW WKH SHUIRUPDQFH RI WKHVH PRGHOV )LJXUH SUHVHQWV WKUHH RI WKHVH PHWULFV IRU WKH VWHDG\VWDWH &9 0/3 PRGHO 7KH SULPDU\ PHWULF RI PRGHO TXDOLW\ XVHG LQ WKLV VWXG\ KDV EHHQ FRUUHODWLRQ 7KH FRUUHODWLRQ EHWZHHQ WKH DFWXDO XQLW 12[ HPLVVLRQ DQG WKH 0/3fV SUHGLFWLRQ RI 12[ ZDV 7KLV PHWULF LQGLFDWHV WKDW WKH 0/3 PRGHO XQGHUVWDQGV D FRQVLGHUDEOH DPRXQW DERXW WKH YDULDWLRQ LQ 12[ SURGXFWLRQ )XUWKHUPRUH QRWLFH WKDW WKH 0/3 SURGXFHG WKLV SUHGLFWLRQ E\ REVHUYLQJ WKH 09V '9V DQG 69V RQO\ LPSO\LQJ WKDW WKH 0/3 XQGHUVWDQGV D FRQVLGHUDEOH DPRXQW DERXW WKH UHODWLRQVKLS EHWZHHQ WKHVH YDULDEOHV DQG 12[ IRUPDWLRQ 7KH QRUPDOL]HG PHDQVTXDUHG HUURU 106(f ZDV DOVR XVHG WR DVVHV WKH ZRUWKLQHVV RI HDFK PRGHO 5HFDOO WKDW D 106( JUHDWHU RU HTXDO WR LQGLFDWHV WKDW WKH PRGHO LV SHUIRUPn LQJ QR EHWWHU WKDQ D WULYLDO SUHGLFWLRQ ZKLFK VLPSO\ SUHGLFWV WKH VWDWLVWLFDO PHDQ +HUH WKH

PAGE 160

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n UHODWLRQ DQG 106( UHVXOWV SUHVHQWHG LQ )LJXUH FDQQRW GLIIHUHQWLDWH EHWZHHQ WKHVH WZR PRGHOV EHFDXVH WKH\ UHODWH WR RQO\ RQH RI WKHVH PRGHOV ,W LV FOHDU WKDW WKH YDOLGDWLRQ PHWULF WKDW ZH DUH VHHNLQJ VKRXOG SURYLGH DQ LQGLFDWLRQ DERXW KRZ D SDUWLFXODU PRGHO FRPSDUHV ZLWK WKH HQVHPEOH RI SRVVLEOH PRGHOV ZKLFK FRXOG

PAGE 161

EH GHYHORSHG IRU WKH SURFHVV )RU QRQOLQHDU PRGHOV WUDLQHG ZLWK D JUDGLHQW GHVFHQW OHDUQn LQJ DOJRULWKP WKLV PXVW LQFOXGH VRPH XQGHUVWDQGLQJ DERXW WKH VWUXFWXUH RI WKH SHUIRUn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Â’ 106( _ M )LJXUH 106( DQG 5 IRU DOO WUDLQLQJ UHVXOWV IRU 0/3 &9 PRGHO ,Q DGGLWLRQ WR FRUUHODWLRQ DQG 106( )LJXUH SUHVHQWHG D PHWULF ODEHOHG fVWG HUU YDUf 7KH UHDOL]DWLRQ WKDW ZH DUH ORRNLQJ IRU D PHWULF FDSDEOH RI GLIIHUHQWLDWLQJ EHWZHHQ D VLQJOH PRGHO DQG WKH HQVHPEOH RI SRVVLEOH PRGHOV OHDGV QDWXUDOO\ WR FRQVLGHULQJ VWDQGDUG

PAGE 162

HUURUV 5HFDOO WKDW 7LEVKLUDQL >@ SURSRVHG WKH fERRWVWUDSf PHWKRG IRU HVWLPDWLQJ WKH VWDQGDUG HUURUV RI D 0/3fV SUHGLFWLRQV +DYLQJ HVWLPDWHV IRU WKH VWDQGDUG HUURUV ZLOO SURn YLGH LQVLJKW LQWR YDULDELOLW\ LQ WKH SUHGLFWLRQV EHWZHHQ PRGHOV GUDZQ IURP WKH HQVHPEOH 7KH fVWG HUUYDUf FROXPQ LQ )LJXUH SUHVHQWV WKH DYHUDJH VWDQGDUG HUURU FDOFXODWHG XVLQJ WKH PRGHO WUDLQLQJV DQG WKH fERRWVWUDSf PHWKRG DFURVV WKH WHVW GDWDVHW GLYLGHG E\ WKH PHDVXUHG 12[ YDULDQFH DFURVV WKLV GDWDVHW 7KH UHDVRQ IRU GLYLGLQJ E\ WKH YDULDQFH LV WR SURYLGH DQ LQWXLWLYH IHHO IRU WKH UHSRUWHG YDOXH RI WKLV PHWULF $ VWDQGDUG HUURU HTXDO WR WKLV YDULDQFH LQGLFDWHV DOPRVW QR FRQILGHQFH LQ WKH PRGHO SUHGLFWLRQV ZKLOH D VWDQGDUG HUURU VLJQLILFDQWO\ VPDOOHU WKDQ WKH YDULDQFH LQGLFDWHV FRQILGHQFH WKDW WKH PRGHOV XQGHUVWDQG PRUH WKDW WKH QDWXUDO YDULDWLRQ RI WKH YDULDEOH 7KH fVWG HUUYDUf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n SHQ f 7KH 09V DQG '9V DUH QRW LQGHSHQGHQW :KHQ WKH FRQWUROOHU PRYHV D 09 LW DVVXPHV WKDW WKH RWKHU 09V DQG '9V UHPDLQ FRQVWDQW f 7KH 69 PRGHOV DUH QRW DFFXUDWHO\ PRGHOLQJ WKH LPSDFW RI 09 PRYHV f 7KH 09V DUH KLJKO\ FRUUHODWHG 7KXV FKDQJLQJ D 09 LQ WKH WHVW GDWDVHW EUHDNV RXU DVVXPSWLRQ WKDW WKH FRUUHODWLRQ UHODWLRQVKLSV ZLWKLQ WKH WUDLQn LQJ DQG WHVW GDWDVHWV DUH WKH VDPH ,W LV XQOLNHO\ WKDW FDVH f LV WKH FDXVH RI RXU WURXEOHV )RU LI WKH 09V DQG '9V DUH QRW LQGHSHQGHQW WKHQ WKHUH LV QR ZD\ IRU DQ\ RI WKH PRGHOV WR KDYH GHWHUPLQHG WKLV *LYHQ

PAGE 163

WKDW DOO RI WKH PRGHOV KDYH EHHQ GHYHORSHG IURP WKH VDPH 09 DQG '9 YDULDEOH VHWV ,I LQ IDFW WKHVH YDULDEOHV DUH QRW LQGHSHQGHQW WKHQ WKH LPSDFW RI WKLV RQ RXU FRQWURO GHVLJQV ZRXOG EH WKH VDPH DV LI WKHVH YDULDEOHV DUH FRUUHODWHG ZKLFK ZLOO EH FRQVLGHUHG LQ FDVH f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n WUROOHU VV690RGHO VV&90RGHO ’ &9 ‘ 69&9 )LJXUH 6XPPDU\ RI YDOLGDWLRQ PHWULFV IRU FRPELQHG 69&9 PRGHO

PAGE 164

&DVH f LV D YDOLG FDXVH IRU RXU SDUDPHWHUL]DWLRQ SUREOHP :KHQ PRGHOLQJ DQ LQGXVn WULDO SURFHVV FRUUHODWHG YDULDEOHV DUH JRLQJ WR EH D IDFW RI OLIH 8QIRUWXQDWHO\ WKH DYDLODEOH WRROV LQ WKH OLWHUDWXUH IRU GHDOLQJ ZLWK VXFK VLWXDWLRQV DUH YHU\ OLPLWHG 7KH IROORZLQJ VHFn WLRQV ZLOO GHPRQVWUDWH WKDW FRUUHODWLRQ KDV D WUHPHQGRXV LPSDFW RQ WKH GHVLJQ RI PRGHO EDVHG FRQWUROOHUV &RUUHODWLRQ 3DUDGR[ &RUUHODWHG YDULDEOHV DUH D QDWXUDO SKHQRPHQRQ DV WZR YDULDEOHV DUH FRUUHODWHG ZKHQ WKH\ DUH UHODWHG WKURXJK VRPH SK\VLFDO ODZV RU SURFHVV ,Q IDFW WKH PLVVLRQ VWDWHPHQW IRU WKH HPSLULFDO LQYHVWLJDWRU LV WR LQIHU RU OHDUQ WKHVH SK\VLFDO ODZV IURP REVHUYDWLRQV RI SURFHVV GDWD +HQFH FRUUHODWLRQ LV D GRXEOHHGJHG VZRUG :LWKRXW LW OHDUQLQJ LV QRW SRVVLn EOH DV WKHUH LV QRWKLQJ WR LQIHU ZKLOH XQDQWLFLSDWHG FRUUHODWLRQ PDNHV LW GLIILFXOW LI QRW LPSRVVLEOH WR XQGHUVWDQG ZKDW KDV EHHQ OHDUQHG RU LQIHUUHG ,QGXVWULDO SURFHVV SODQWV WKURXJK FHQWUDOL]HG FRQWURO V\VWHPDWLFDOO\ FRUUHODWH WKHLU SURFHVV YDULDEOHV 7KH SODQWfV GLVWULEXWHG FRQWURO V\VWHP '&6f PDLQWDLQV D ODUJH QXPEHU WKRXVDQGV WR WHQV RI WKRXVDQGVf RI IHHGIRUZDUG DQG IHHGEDFN FRQWURO ORRSV IURP D UHODn WLYHO\ VPDOO QXPEHU WHQV WR KXQGUHGVf RI RSHUDWRU VHWSRLQWV 7KH '&6 LV GHVLJQHG WR EDWFK FRQWURO RYHU DV PDQ\ VXEV\VWHPV DV SRVVLEOH IURP WKH VPDOOHVW SRVVLEOH QXPEHU RI RSHUDn WRU VHWSRLQWV ,I LW ZHUH IHDVLEOH WKH RSHUDWRU ZRXOG RQO\ KDYH D VLQJOH VHWSRLQW FDOOHG GHPDQG $V D UXOH PRVW YDULDEOHV ZLWKLQ DQ LQGXVWULDO SODQW ZLOO EH KLJKO\ FRUUHODWHG WR SODQW GHPDQG 7KH HIIHFWV RI FRUUHODWLRQ KDYH EHHQ ZHOO GRFXPHQW IRU PRGHOLQJ DSSOLFDWLRQV OLNH V\Vn WHP LGHQWLILFDWLRQ >@ DQG UHJUHVVLRQ >@ 7KH PRVW VLJQLILFDQW DWWULEXWH WKDW WKHVH DSSOLn FDWLRQV KDYH LQ FRPPRQ LV WKDW WKH\ GHDO ZLWK V\VWHPV WKDW DUH HLWKHU OLQHDU RU KDYH D

PAGE 165

UHODWLYHO\ VLPSOH SDUDPHWULF QRQOLQHDU IRUP 7KLV IHDWXUH DOORZV DQ LQYHVWLJDWRU WR DSSO\ DQDO\WLFDOO\ DVVHVV WKH VLJQLILFDQFH RI PRGHO SDUDPHWHUV ,Q IDFW WKH FRUUHODWLRQ HIIHFWV DUH DXWRPDWLFDOO\ DFFRXQWHG IRU LQ VLJQLILFDQFH WHVWLQJ 7KXV WKH YDOLGDWLRQ UHTXLUHG WR DVVHVV WKH SDUDPHWHUL]DWLRQ RI D PRGHO LV DFFRPSOLVKHG WKURXJK VLJQLILFDQFH WHVWLQJ 7R LOOXVWUDWH WKH HIIHFW RI FRUUHODWLRQ FRQVLGHU WKH SK\VLFDO SURFHVVLQJ SODQW WKDW SURn GXFHV RXWSXW \ IURP LQSXWV X [ DQG X $Q LQYHVWLJDWRU EXLOGLQJ D PRGHO IRU WKLV SODQW LV SUHVHQWHG ZLWK GDWD IRU \ X c DQG X ZLWKRXW DQ\ LQIRUPDWLRQ DERXW WKH SK\VLFDO V\VWHP 7KH LQYHVWLJDWRU K\SRWKHVL]HV WKDW WKH GDWD ZDV JHQHUDWHG DFFRUGLQJ WKH UHJUHVVLRQ PRGHO \ 3MmL 3Z Wf f ZKHUH 3 DQG 3 DUH XQNQRZQ FRHIILFLHQWV DQG U_ LV D ]HURPHDQ XQFRUUHODWHG GLVWXUn EDQFH WHUP $SSO\LQJ WKH PHWKRG RI OHDVW VTXDUHV /6f 5DPDQDWKDQ >@ VKRZV WKDW WKH FRUUHVSRQGLQJ QRUPDO HTXDWLRQV DUH JLYHQ E\ 3L=L S=ZP f 3O=fO03O=f
PAGE 166

DQG YDULDQFHV 9DU3f I f eOO aUf 9DU Sf I f ;ZO aUf &RY Sf f§‘ D nn f ZKHUH Df )DUU_f DQG U LV WKH FRUUHODWLRQ FRHIILFLHQW EHWZHHQ 8M DQG m 6XSSRVH Z DQG Z DUH KLJKO\ FRUUHODWHG U LV QHDU s ,W LV HYLGHQW IURP HTXDWLRQV f DQG f WKDW WKH YDULDQFHV DQG KHQFH WKH VWDQGDUG HUURUV RI S DQG S" ZLOO EH YHU\ ODUJH ZKHQ U LV FORVH WR $ ODUJH YDULDQFH PHDQV SRRU SUHFLVLRQ DQG D ORZ VWXGHQW WVWDWLVWLF ZKLFK UHVXOWV LQ LQVLJQLILFDQFH ,Q DGGLWLRQ ZH FDQ VHH IURP HTXDWLRQ f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f ,I WZR RU PRUH H[SODQDWRU\ YDULDEOHV LQ WKH PXOWLSOH LQSXW 0,f PRGHO DUH H[DFWO\ OLQHDUO\ UHODWHG WKHQ WKH PRGHO FDQQRW EH HVWLPDWHG

PAGE 167

f ,I VRPH H[SODQDWRU\ YDULDEOHV DUH QHDUO\ OLQHDUO\ UHODWHG WKHQ 2/6 DQG 0/( DQG KHQFH DUH XQELDVHG HIILFLHQW DQG FRQVLVWHQW f 7KH HIIHFW RI FRUUHODWLRQ DPRQJ LQSXW YDULDEOHV LV WR LQFUHDVH WKH VWDQn GDUG HUURUV RI WKH SDUDPHWHUV DQG UHGXFH WKH WVWDWLVWLFV WKXV PDNLQJ WKHVH SDUDPHWHUV OHVV VLJQLILFDQW DQG SRVVLEO\ HYHQ LQVLJQLILFDQWf 7KH WHVWV RI K\SRWKHVHV DUH KRZHYHU YDOLG f 7KH FRYDULDQFH EHWZHHQ WKH SDUDPHWHUV RI D SDLU RI KLJKO\ FRUUHODWHG YDULDEOHV ZLOO EH YHU\ KLJK LQ DEVROXWH YDOXH WKXV PDNLQJ LW GLIILFXOW WR LQWHUSUHW LQGLYLGXDO FRHIILFLHQWV f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n ZRUNV KDYH EHHQ DSSOLHG WR GDWH 7KH DXWKRU RIIHUV WKH IROORZLQJ REVHUYDWLRQV f 'HVFHQW EDVHG OHDUQLQJ ZLOO DOZD\V DUULYH DW D VROXWLRQV UHJDUGOHVV RI WKH GHJUHH RI FRUUHODWLRQ ZLWKLQ WKH LQSXW VSDFH 7KH VROXWLRQ KRZHYHU LV UDUHO\ XQLTXH QRU JOREDOO\ RSWLPDO f 7KH YDVW PDMRULW\ RI DSSOLFDWLRQV IRU QHXUDO QHWZRUNV UHO\ SXUHO\ RQ WKHLU DELOLW\ WR IRUHFDVW )RU WKLV UHDVRQ VWDQGDUG HUURUV IRU D QHWZRUNV SUHGLFWLRQV KDV EHHQ D UHFHQW WRSLF RI UHVHDUFK f ,I WKH FRUUHODWLRQ UHODWLRQVKLSV LQ WKH LQSXW VSDFH RI DOO WHVWLQJ GDWDVHWV DUH LGHQWLFDO WR WKRVH LQ WKH WUDLQLQJ GDWDVHW WKHQ WKH FRUUHODWHG LQSXWV ZLOO QRW GHJUDGH WKH QHWZRUNfV SUHGLFWLRQV 7KLV VLWXDWLRQ ZKLFK ZLOO RIWHQ EH WKH FDVH IRU SUHGLFWLRQ DSSOLFDWLRQV LV XQOLNHO\ LQ FRQWURO DSSOLn FDWLRQV VLQFH WKH FRQWUROOHU ZLOO LQGHSHQGHQWO\ PRYH WKHVH LQSXWV

PAGE 168

f 7KH QHWZRUN FRHIILFLHQWV LWV ZHLJKWV KDYH QR SDUDPHWULF LQWHUSUHWDWLRQ 7KHLU KDV EHHQ OLWWOH UHDVRQ HYHQ LI ZH NQHZ KRZ WR WR FDOFXODWH WKHLU VWDQGDUG HUURUV DQG VLJQLILFDQFH %HFDXVH RI WKH ZD\ LQ ZKLFK QHXUDO QHWZRUNV KDYH EHHQ DSSOLHG WKH YDVW PDMRULW\ RI ZRUN GRQH RQ WKH LVVXH RI FRUUHODWHG LQSXWV GHDOV DOPRVW HQWLUHO\ ZLWK LWV HIIHFW RQ WKH G\QDPLFV RI OHDUQLQJ &RUUHODWLRQ LQ WKH LQSXW VSDFH GUDVWLFDOO\ UHGXFHV WKH UDWH RI FRQYHUn JHQFH IRU GHVFHQWEDVHG OHDUQLQJ DOJRULWKPV >@ 7R FRXQWHU WKLV HIIHFW WKH PRVW FRPn PRQ QHXUDO QHWZRUN VROXWLRQ WR FRUUHODWLRQ LV WR SURMHFWLRQ WKH LQSXW VSDFHV WR D ORZHU GLPHQVLRQDO VXEVSDFH ZLWK RUWKRJRQDO EDVHV 0RVW RI WKHVH SURMHFWLRQ RSHUDWRUV DUH OLQHDU DQG EDVHG RQ WKH HQHUJ\ RU LQIRUPDWLRQ FRQWHQW RI WKH YDULDEOHV 7KH PRVW FRPPRQ VXFK SURMHFWLRQ LV SULQFLSOH FRPSRQHQW DQDO\VLV 7KH UHVXOW RI XVLQJ VXFK D SURMHFWLRQ LV WR SUHn FHGH WKH QHXUDO QHWZRUN ZLWK D VLPSOH PDWUL[ PXOWLSO\ 7KXV WKH DFWXDO LQSXWV WR WKH QHXUDO QHWZRUN DUH FRPSOHWHO\ RU QHDUO\ FRPSOHWHO\ XQFRUUHODWHG IHDWXUHV DQG WKXV DFFHOHUDWH WKH FRQYHUJHQFH RI WKH OHDUQLQJ DOJRULWKP $OWKRXJK YHU\ XVHIXO WKHVH WHFKQLTXHV GR QRW SURYLGH D VROXWLRQ WR WKH SDUDPHWHUL]Dn WLRQ SUREOHP 7KHVH WHFKQLTXHV FDQ EH XVHG WR LPSURYH WKH FRQVLVWHQF\ RI WKH QHXUDO QHWn ZRUN SRUWLRQ RI WKH PRGHO EXW GR QRW LPSURYH WKH FRQVLVWHQF\ RI WKH FRPELQHG PRGHOV FRQWDLQLQJ WKH WUDQVIRUP VWDJH ZLWK WKH QHXUDO QHWZRUN 7KLV SRLQW PD\ EH FRQIXVLQJ QRZ EXW WKH IROORZLQJ VHFWLRQV VKRXOG KHOS WR FODULI\ LW ,Q VXPPDU\ QHXUDO QHWZRUNV KDYH IRXQG DQ DSSOLFDWLRQV QLFKH ZKHUH WKLV UREXVW SUHn GLFWRU KDV GHPRQVWUDWHG WKH DELOLW\ WR RXW IRUHFDVW PRUH WUDGLWLRQDO PHWKRGV 2QH RI WKH SUREOHPV ZLWK PRYLQJ QHXUDO QHWZRUNV IURP DQ DFDGHPLF LQWHUHVW WR DQ DFFHSWHG PRGHOn LQJ PHWKRGRORJ\ LV WKH ODFN RI VWDQGDUGL]HG UHSRUWLQJ 7KH YDVW PDMRULW\ RI QHXUDO QHWZRUN DSSOLFDWLRQV WR GDWH DSSO\ WKH LQIHUHQFHV RI D PRGHO ZLWKRXW NQRZOHGJH RI WKHLU VWDWLVWLFDO

PAGE 169

VLJQLILFDQFH 5HFHQW ZRUN LQ VWDQGDUG HUURUV IRU WKHVH SUHGLFWLRQV KDYH WDNHQ DQ LPSRUWDQW VWHS WRZDUGV VROYLQJ SDUW RI WKLV SUREOHP >@ (IIHFWV RI 0XOWLFROOLQHDULW\ RQ 0RGHO%DVHG &RQWURO (DFK RI WKH FRQWUROOHUV SUHVHQWHG LQ &KDSWHU EHORQJ WR WKH JHQHUDO IDPLO\ RI PRGHO EDVHG FRQWURO LH WKH\ HDFK H[SOLFLWO\ XVH D UHIHUHQFH PRGHO GXULQJ RIIOLQH WUDLQLQJ RU RQOLQH FRQWURO 7KH SURFHVV NQRZOHGJH SURYLGHG E\ WKH PRGHO LQ HDFK RI WKHVH FRQWURO GHVLJQV LV GHY GPY f ZKHUH WKH YHFWRUV FY DQG PY DUH WKH &9V DQG 09V UHVSHFWLYHO\ 5HFDOO RXU H[DPSOH LQ 6HFWLRQ &RUUHODWLRQ 3DUDGR[ ZKHUH WKH PRGHO ZDV JLYHQ E\ \ 0L 3m 7 f ,I WKLV PRGHO ZDV XVHG DV D UHIHUHQFH PRGHO IRU D PRGHOEDVHG FRQWURO VFKHPH FY ^\` DQG PY ^X[X` FRXOG EH FRQVLGHUHG WR EH WKH &9V DQG 09V UHVSHFn WLYHO\ +HQFH WKH SURFHVV NQRZOHGJH SURYLGHG E\ RXU PRGHO LV WKH VHW ^S c 3` %XW DV ZH KDYH VHHQ ZKHQ WKH 09V DQG X DUH KLJKO\ FRUUHODWHG WKH VWDQGDUG HUURUV IRU 3 DQG 3" ZLOO EH YHU\ ODUJH $ VLWXDWLRQ WKDW ZLOO TXLFNO\ UHQGHU RXU SURFHVV NQRZOHGJH LQVLJQLIn LFDQW &OHDUO\ WKLV ZLOO KDYH D WUHPHQGRXV LPSDFW RQ WKH SHUIRUPDQFH DQG UREXVWQHVV RI RXU FRQWUROOHU

PAGE 170

9DOLGDWLRQ 0HWULF 1HXUDO QHWZRUNV GR QRW KDYH WKH VDPH FRQYHQLHQW LQWHUSUHWDWLRQ IRU WKHLU FRHIILFLHQWV 7KH\ GR KRZHYHU LQIHU UHODWLRQVKLSV EHWZHHQ WKH 09V DQG &9V ZKLFK LV XVHG GLUHFWO\ E\ WKH FRQWUROOHUV 7KH PHWKRG RI H[WUDFWLQJ WKLV UHODWLRQVKLS IURP D QHXUDO QHWZRUN PRGHO LV FRPPRQO\ UHIHUUHG WR DV VHQVLWLYLW\ DQDO\VLV 7KH 09 VHQVLWLYLWLHV GFYGPY FDQ EH FDOFXODWHG GLUHFWO\ XVLQJ EDFNSURSDJDWLRQ 7KHVH VHQVLWLYLWLHV ZLOO EH D IXQFWLRQ RI WKH XQLWfV RSHUDWLQJ VWDWH ^PY GY` 5HFDOO WKDW ZH DUH WU\LQJ WR JHW D IHHO IRU KRZ WKHVH VHQVLn WLYLWLHV YDU\ DFURVV WKH HQVHPEOH RI SRVVLEOH PRGHOV )LJXUH LOOXVWUDWHV WKH 09 VHQVLWLYn LWLHV IRU D W\SLFDO RSHUDWLQJ VWDWH DFURVV WKH VWHDG\VWDWH 0/3 PRGHOV UHSUHVHQWLQJ D VDPSOLQJ IURP WKLV HQVHPEOH D R D R G R D R )LJXUH 6HQVLWLYLW\ UHVXOWV IRU DOO WUDLQLQJ UHVXOWV IRU 12[ &9 PRGHO )RU WKH ILUVW WLPH ZH FDQ FOHDUO\ VHH RXU SDUDPHWHUL]DWLRQ SUREOHP (DFK PRGHO LV SURn YLGLQJ VLJQLILFDQWO\ GLIIHUHQW SURFHVV NQRZOHGJH WR WKH FRQWUROOHU 0RUH H[SOLFLWO\ RXU EHVW DQG VHFRQGEHVW PRGHOV ZLWK UHVSHFW WR 106( 5 DQG 6( UHSUHVHQWHG E\ WUDLQLQJ

PAGE 171

UXQV QXPEHU DQG ORRN DV LI WKH\ DUH PRGHOLQJ WZR FRPSOHWHO\ GLIIHUHQW SURFHVVHV $ODV WKHUHfV KRSH )LJXUH SURYLGHV WKH DYHUDJH VHQVLWLYLW\ UHSRUWHG IURP WKH HQVHPEOH RI PRGHOV DORQJ ZLWK WKH FRUUHVSRQGLQJ b FRQILGHQFH LQWHUYDOV 7KLV UHVXOW GHPRQVWUDWHV WKDW ZH DUH QRW DEOH WR DVVHUW WKH GLUHFWLRQDO VHQVLWLYLW\ ZLWK DQ\ GHJUHH RI FRQILGHQFH ,W LV HDV\ WR VHH ZK\ WKH FRQWUROOHUV KDG VR PXFK WURXEOH )LJXUH 12[ &9 PRGHO VHQVLWLYLW\ ZLWK b FRQILGHQFH LQWHUYDOV 5HYLVHG 5HSUHVHQWDWLRQ 3UXQLQJ $OJRULWKP 5HFDOO WKDW WKH JRDO RI WKH UHSUHVHQWDWLRQ SUXQLQJ DOJRULWKP LV WR LGHQWLI\ WKH EHVW YDULn DEOH UHSUHVHQWDWLRQV ZLWKLQ HDFK JURXS ,Q &KDSWHU WKH fEHVWf YDULDEOH UHSUHVHQWDWLRQ ZDV GHWHUPLQHG E\ VHOHFWLQJ WKH UHSUHVHQWDWLRQ ZKLFK SURGXFHG D PRGHO ZLWK WKH KLJKHVW FRUUHODWLRQ RQ D EOLQG WHVW GDWDVHW ,Q KLQGVLJKW LW LV FOHDU WKDW WKH GHILQLWLRQ RI EHVW PXVW EH DSSOLFDWLRQVSHFLILF ,I WKH PRGHOV ZHUH EHLQJ DSSOLHG WR D IRUHFDVWLQJ DSSOLFDWLRQ WKHQ WKH UHSUHVHQWDWLRQ SUXQLQJ DOJRULWKP ZRXOG KDYH VXLWHG RXU SXUSRVHV )RU FRQWURO DSSOLFD

PAGE 172

WLRQV KRZHYHU LW LV FOHDU WKDW WKH GHILQLWLRQ RI EHVW ZLOO KDYH WR EH PRGLILHG 1RWH WKDW WKH UREXVWQHVV RI WKH SUHGLFWRUV SURGXFHG E\ WKLV DOJRULWKP VKRXOG VWLOO EH LQ GRXEW ,I WKH IRUHFDVWV DUH DOZD\V PDGH IRU D GDWDVHW ZLWK LGHQWLFDO FRUUHODWLRQV FKDUDFWHULVWLFV WR WKH WUDLQLQJ GDWDVHW WKHQ WKH PRGHOfV IRUHFDVW ZRXOG EH ILQH ,I KRZHYHU WKH FRUUHODWLRQ LV OHVV SK\VLFDO WKHQ WKHUH LV VWLOO FDXVH IRU FRQFHUQ ,QSXW 6HQVLWLYLW\ 6WDQGDUG (UURUV 5HFDOO WKH fERRWVWUDSf DSSURDFK WR FDOFXODWLQJ WKH SUHGLFWLRQ VWDQGDUG HUURUV IRU D QHXn UDO QHWZRUN SUHVHQWHG LQ 6HFWLRQ %RRWVWUDS PHWKRGV 7KH YDOLGDWLRQ PHWULF WR GLIIHUHQWLDWH PRGHOV WR EH XVHG IRU RXU PRGHOEDVHG FRQWURO GHVLJQV ZLOO VLPLODUO\ QHHG WR HYDOXDWH WKH VWDQGDUG HUURU RI WKH 09 VHQVLWLYLWLHV 7KH IROORZLQJ DOJRULWKP KDV EHHQ GHYHORSHG WR DSSUR[LPDWH WKH 09 VWDQGDUG HUURUV XVLQJ D fERRWVWUDSf PHWKRGRORJ\ *HQHUDWH 1 GDWDVHWV HDFK RQH RI VL]H 1r GUDZQ ZLWK UHSODFHPHQW ,, )RU HDFK ERRWVWUDS GDWDVHW E H > LY @ ILQG $UJ0LQK^-GESXE 7Yff` f ,,, (VWLPDWH WKH VWDQGDUG HUURU RI WKH L WK LQSXW VHQVLWLYLW\ DV f ZKHUH f

PAGE 173

&RQFHSWXDOO\ WKLV DOJRULWKP VLPSO\ FDOFXODWHV WKH YDULDQFH RI WKH LQSXW VHQVLWLYLWLHV DFURVV PXOWLSOH PRGHOV DOO WUDLQHG XVLQJ GLIIHUHQW UDQGRP LQLWLDO FRQGLWLRQV DQG LQGHSHQn GHQWO\ VDPSOHG fERRWVWUDSf GDWDVHWV &OHDUO\ WKH YDULDQFH RI WKH VHQVLWLYLW\ FDOFXODWLRQV SUHVHQWHG LQ )LJXUH ZLOO EH KLJK ,I DQ DOJRULWKP FDQ EH IRXQG WKDW LV FDSDEOH RI UHGXFn LQJ WKHVH VWDQGDUG HUURUV WKH RIIOLQH FRQWUROOHU TXDQWLILFDWLRQ UHVXOWV VKRXOG LPSURYH $OJRULWKP 7KH UHYLVHG UHSUHVHQWDWLRQ SUXQLQJ DOJRULWKP XVHV WKH VWDQGDUG HUURU HVWLPDWHV IRU WKH PRGHOfV LQSXW VHQVLWLYLW\ DFURVV D GDWDVHW DV SUHVHQWHG LQ 6HFWLRQ ,QSXW 6HQVLWLYLW\ 6WDQGDUG (UURUV WR GHWHUPLQH fEHVWf DV IROORZV 3HUIRUP WKH JURXS VHOHFWLRQ SUXQLQJ DOJRULWKP WR WKH (VVHQWLDO 7DJ /LVW LQ WKH DSSHQGL[ ,, 6HOHFW WKH ILUVW 09 JURXS LQ WKH UHGXFHG (VVHQWLDO 7DJ /LVW ,,, ,I WKHUH LV RQO\ RQH XQLTXH fUHSUHVHQWDWLRQf LQ WKH JURXS WKHQ VHOHFW WKH QH[W 09 JURXS DQG JRWR VWHS ,,, ,9 )RU HDFK XQLTXH UHSUHVHQWDWLRQ LQ WKH VHOHFWHG JURXS WUDLQ D 0/3f WLPHV ZLWK WKH YDULDEOHV IRU WKLV UHSUHVHQWDWLRQ DV WKH RQO\ LQSXW YDULDEOHV IURP WKH VHOHFWHG JURXS WKH PRGHOV UHVXOWLQJ IURP HDFK WUDLQLQJ ZLOO EH FDOOHG fUHSUHVHQWDWLRQf PRGHO VHW 9 &DOFXODWH WKH VWDQGDUG HUURU IRU WKH LQSXW VHQVLWLYLWLHV IRU HDFK 09 9, &DOFXODWH WKH DYHUDJH QRUPDOL]HG LQSXW VHQVLWLYLW\ VWDQGDUG HUURU IRU HDFK fUHSUHVHQWDWLRQf PRGHO VHW DFFRUGLQJ WR f

PAGE 174

9,, 6HOHFW WKH EHVW fUHSUHVHQWDWLRQf PRGHO VHW DV WKH VHW ZLWK WKH ORZHVW LQSXW VHQVLWLYLW\ QRUPDOL]HG DYHUDJH HUURU DQG UHPRYH DOO YDULDEOHV DVVRFLDWHG ZLWK HYHU\ fUHSUHVHQWDWLRQf PRGHO VHW H[FHSW IRU WKH EHVW fUHSUHVHQWDWLRQf PRGHO VHW 9,,,6HOHFW WKH QH[W JURXS DQG JRWR VWHS ,,, )LJXUH 5HVXOWV RI UHYLVHG UHSUHVHQWDWLRQ SUXQLQJ DOJRULWKP 7KH UHVXOWV DIWHU UXQQLQJ WKH UHYLVHG SUXQLQJ DOJRULWKP IRU HDFK RI WKH JURXSV DUH LOOXVWUDWHG LQ )LJXUH 1RWLFH WKDW XQOLNH WKH LQLWLDO SUXQLQJ DOJRULWKP WKLV DOJRULWKP VLJn QLILFDQWO\ GHJUDGHG WKH ILGHOLW\ RI WKH PRGHO DV LW SUXQHG HDFK UHSUHVHQWDWLRQ JURXS 7KH YDULDEOH VHOHFWLRQV DIWHU WKH UHYLVHG UHSUHVHQWDWLRQ SUXQLQJ DOJRULWKPV DUH SURYLGHG LQ 7DEOH

PAGE 175

7DEOH )LQDO YDULDEOH VHOHFWLRQV DIWHU UHYLVHG SUXQLQJ 0DQLSXODWHG 9DULDEOHV 'LVWXUEDQFH 9DULDEOHV 6WDWH 9DULDEOHV &RQWURO 9DULDEOHV 2)$ 'DPSHU %LDV $PELHQW $LU 3UHVV 6HF $LU7HPS 6LGH $ &(0 12[ 2)$ 'DPSHU %LDV $PELHQW $LU 7HUQS 6HF $LU 7HPS 6LGH % )' )DQ %LDV %QU $WP 6WP 3UHVV 2(0 &2 c 7ULP %QU $WP 6WP 7HUQ S *HQHUDWHG 0 : *5 )DQ $ ,QOHW 'PSU %LDV 2(0 %DURPHWULF 3UHVVXUH : LQGER[ 3UHVVXUH *5 )DQ % ,QOHW 'PSU %LDV &RQG %DFN 3UHV 6LGH $ *5 )DQ +SSU'PSU$ %LDV &RQG %DFN 3UHV 6LGH % *5 )DQ +SSU'PSU% %LDV )XHO *DV )ORZ ,QGLFDWLRQ )XHO 2LO )ORZ ,QGLFDWLRQ ]M )XHO 7HPS )LUHG )XUQDFH 3UHVVXUH 9DULDEOH 5HSUHVHQWDWLRQ 7KH UHVXOWV RI WKH UHYLVHG UHSUHVHQWDWLRQ SUXQLQJ DOJRULWKP DUH YHU\ LQWHUHVWLQJ DQG SURYLGH VRPH OHVVRQV DERXW YDULDEOH UHSUHVHQWDWLRQ FKRLFHV )RU HDFK 09 JURXS WKH UHYLVHG DOJRULWKP FKRVH WKH RSHUDWRU ELDV UDWKHU WKDQ WKH SRVLWLRQV VHOHFWHG E\ WKH LQLWLDO DOJRULWKP 5HFDOO WKH H[DPSOH JLYHQ LQ 6HFWLRQ 9DULDEOH UHSUHVHQWDWLRQ ZKHUH WDJV ZHUH LGHQWLILHG WKDW UHSUHVHQWHG WKH VLQJOH SURFHVV YDULDEOH RI JURVV DLUIORZ $V ZH KDYH VHHQ WKLV PDQ\WRRQH PDSSLQJ EHWZHHQ WDJV DQG YDULDEOHV LV YHU\ FRPPRQ LQ LQGXVWULDO SODQWV 7KH WDJV VHOHFWHG IRU YDULDEOH UHSUHVHQWDWLRQ VKRXOG KDYH WKH IROORZLQJ FKDUDFWHULVWLFV f 8QFRUUHODWHG 7KH WDJ RU WDJV FKRVHQ WR UHSUHVHQW WKH YDULDEOH VKRXOG EH XQFRUUHn ODWHG ZLWK HDFK RWKHU DQG ZLWK WDJV FKRVHQ WR UHSUHVHQW WKH RWKHU LQSXW YDULDEOHV f 5HSUHVHQWDWLYH 7KH WDJV FKRVHQ VKRXOG SURYLGH D FRPSOHWH UHSUHVHQWDWLRQ RI WKH G\QDPLFV RI WKH YDULDEOH LH WKH\ VKRXOG EH DV FRUUHODWHG DV SRVVLEOH WR WKH SURn FHVV YDULDEOH f '\QDPLF 7KH WDJV FKRVHQ VKRXOG FRQWDLQ DV PXFK RI WKH UHDO YDULDWLRQ LQ WKH SURFHVV YDULDEOH DV SRVVLEOH ,I WKH YDULDEOH LV FKDQJLQJ IRU DQ\ UHDVRQ HYHQ LI WKH YDULDWLRQ LI QRW LQWHQGHG RU QRW GHVLUDEOH WKLV LQIRUPDWLRQ VKRXOG EH UHSUH

PAGE 176

VHQWHG LQ WKH FKRVHQ WDJV )URP WKH GLVFXVVLRQ DERYH LW LV HDV\ WR VHH WKDW DOO WDJV H[FHSW )' IDQ WULP ZLOO EH KLJKO\ FRUUHODWHG WR SODQW GHPDQG $V QRWHG SUHYLRXVO\ PRVW WDJV ZLOO EH FRUUHODWHG ZLWK SODQW GHPDQG 7KLV ZRXOG PDNH WKHVH IRXU WDJV D SRRU FKRLFH IRU UHSUHVHQWLQJ )' IDQ DLUn IORZ )' IDQ WULP RQ WKH RWKHU KDQG LV QRW FRUUHODWHG WR GHPDQG 7KH WULP LV D ELDV WDJ WKDW RQO\ PRYHV ZKHQ WKH RSHUDWRU ZLVKHV WR DOWHU WKH IXHODLU PL[WXUH UHSUHVHQWHG E\ )' IDQ GHPDQG WKDW ZDV GHVLJQHG LQWR WKH '&6 7KH ELDV LV DOVR DSSHDOLQJ IURP DQ RSWLPL]Dn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n IORZ DFWXDOO\ FKDQJHG DQG KDV DQ LPSDFW RQ WKH FRPEXVWLRQ SURFHVV $OWKRXJK QRW LQWHQGHG WKLV YDULDWLRQ SURYLGHV ULFK GDWD IRU OHDUQLQJ 6R ZH FDQ HLWKHU FKRRVH D UHSUHVHQWDWLRQ ULFK ZLWK G\QDPLF LQIRUPDWLRQ DERXW WKH SURn FHVV EXW KLJKO\ FRUUHODWHG WR GHPDQG RU D UHSUHVHQWDWLRQ FRPSOHWHO\ XQFRUUHODWHG WR GHPDQG WKDW LV UDUHO\ PRYHG 3UXQLQJ WKH YDULDEOH UHSUHVHQWDWLRQV EDVHG RQ WKH ILGHOLW\ RI WKH UHVXOWLQJ PRGHO ZLOO VHOHFW WDJV ULFK LQ G\QDPLF LQIRUPDWLRQ UHJDUGOHVV RI FRUUHODWLRQ 7KLV LV DQ LQWHUHVWLQJ IDFW VLQFH )' IDQ WULP WXUQHG RXW WR EH RQH RI WKH VLJQLILn FDQW OHYHUV RYHU RSWLPL]LQJ 12[

PAGE 177

SUXQLQJ EDVHG RQ VWDQGDUG HUURUV RI WKH NQRZOHGJH H[WUDFWHG DERXW WKH FDXVHDQGHIIHFW UHODWLRQVKLSV ZLOO SULRULWL]H LQGHSHQGHQW WDJV 7KH DQVZHU WR WKLV GLOHPPD LGHQWLILHG LQ WKLV ZRUN UHTXLUHV WZR VROXWLRQV f ZLWK VXIILFLHQW SDUDPHWULF WHVWLQJ VWUXFWXUHG PRYHPHQW RI WKH 09V WKH PRGHOV ZLOO EH DEOH WR H[WUDFW HQRXJK NQRZOHGJH DERXW WKH SURFHVV IRU WKH FRQWUROOHU WR IXQFWLRQ SURSHUO\ DQG f ZKHQ WKH FRQWUROOHUV DUH FRQQHFWHG WR WKH SODQW WKH\ ZLOO PRYH WKH 09V DQG PRGHO UHWXQLQJ ZLOO FRQWLQXH WR H[WUDFW PRUH NQRZOHGJH DERXW WKH SURFHVV )LJXUH LOOXVWUDWHV WKH 09 VHQVLWLYLWLHV IRU WKH VDPH RSHUDWLQJ VWDWH LOOXVWUDWHG LQ )LJn XUH DFURVV WKH ILUVW RI WKH UHYLVHG VWHDG\VWDWH 0/3 PRGHOV )LJXUH 6HQVLWLYLW\ UHVXOWV IRU DOO WUDLQLQJ UHVXOWV IRU UHYLVHG 12[ &9 PRGHO 7KHUH LV D YLVLEOH GLIIHUHQFH LQ WKH VLJQLILFDQFH LH RXU FRQILGHQFH LQ WKHVH PRGHOV DV FRPSDUHG WR RULJLQDO PRGHOV 7KHVH PRGHOV KDYH H[WUDFWHG FRQVLVWHQW SURFHVV NQRZOHGJH ZLWK UHVSHFW WR WKH UHODWLRQVKLSV EHWZHHQ WKH 09V DQG &9V $ FRPSDULVRQ EHWZHHQ WKH

PAGE 178

VWDQGDUG HUURU FDOFXODWLRQV EHWZHHQ WKHVH PRGHOV DQG WKH LQLWLDO PRGHOV LV SUHVHQWHG LQ )LJXUH 7KLV UHVXOW FOHDUO\ YDOLGDWHV RXU REVHUYDWLRQV DERXW )LJXUH 2)$ 'DPSHU %LDV 2)$ 'DPSHU %LDV )' )DQ %LDV 7ULP *5 )DQ $ ,QOHW 'PSU %LDV *5 )DQ % ,QOHW 'PSU %LDV *5 )DQ +SSU 'PSU $ %LDV *5 )DQ +SSU 'PSU % %LDV ’ %HIRUH ‘ $IWHU )LJXUH 5HYLVHG 12[ &9 PRGHO VHQVLWLYLW\ ZLWK b FRQILGHQFH LQWHUYDOV 0RGHOLQJ )LJXUH SUHVHQWV WKH UHVXOWV RI WUDLQLQJ HDFK RI WKH EHVW PRGHO DUFKLWHFWXUHV IURP &KDSWHU ZLWK WKH QHZ YDULDEOH GHILQLWLRQV SUHVHQWHG LQ 7DEOH +HUH HDFK PRGHO DUFKLn WHFWXUH ZDV WUDLQHG WLPHV WKH WUDLQLQJ UHVXOW ZLWK WKH ORZHVW FURVVYDOLGDWLRQ HUURU ZDV VHOHFWHG DQG WKH VHOHFWHG UHVXOW ZDV WHVWHG DJDLQVW WKH EOLQG WHVW VHW 1RWLFH WKDW WKLV SURn FHVV LV HTXLYDOHQW WR WKH WUDLQLQJ SURFHVV IRU WKH SUHYLRXV PRGHOV &RPSDULQJ )LJXUH ZLWK )LJXUH WKH QHZ PRGHOV KDYH ORVW VRPH ILGHOLW\ ZLWK UHVSHFW WR WKHLU DELOLW\ WR SUHGLFW D EOLQG WHVW GDWDVHW 7KH GHJUDGDWLRQ LQ ILGHOLW\ KRZHYHU LV TXLWH VPDOO ZKHQ FRPSDUHG ZLWK WKH LQFUHDVH LQ FRQILGHQFH ZLWK UHVSHFW WR WKH 09 VHQn VLWLYLWLHV WKDW WKH QHZ PRGHOV KDYH )LJXUH f

PAGE 179

XXXX 0/3 7'11 *0 1/66 ’ &9 0RGHO ‘ 69 0RGHO ’ ,69 0RGHO ’ ,09 0RGHO )LJXUH %HVW UHYLVHG PRGHOV IRU DOO PRGHO GHILQLWLRQV E\ DUFKLWHFWXUH &RQWURO ,PSOHPHQWDWLRQ 7KH DVVXPSWLRQ LV WKDW PRGHOV ZLWK KLJKHU FRQILGHQFH LQ WKH 09 VHQVLWLYLWLHV ZLOO SURn YLGH EHWWHU UHIHUHQFH PRGHOV IRU RXU PRGHOEDVHG FRQWURO GHVLJQV (DFK FRQWURO GHVLJQ ZLOO QRZ EH LPSOHPHQWHG ZLWK WKHVH QHZ PRGHOV DQG WKHLU SHUIRUPDQFH UHTXDQWLILHG 2IIOLQH 4XDQWLILFDWLRQ 7KH RIIOLQH SHUIRUPDQFH IRU WKH QHZ FRQWUROOHU ZLOO IROORZ WKH VDPH PHWKRGRORJ\ SUHn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

PAGE 180

$YJ 12[ 5HGXFWLRQ )LJXUH $YHUDJH 12[ UHGXFWLRQ RYHU WHVWLQJ GDWDVHW XVLQJ ROG DQG UHYLVHG PRGHOV $YJ &2 $ERYH 0D[ f UU P f§L ‘f§L M OL E %DVHOLQH 6WHDG\6WDWH 0 3& 6WHDG\6WDWH 0,& '\QDPLF 03& '\QDPLF 05$& ’ %HIRUH D $IWHU )LJXUH $YHUDJH &2 DERYH PD[ RYHU WHVWLQJ GDWDVHW XVLQJ ROG DQG UHYLVHG PRGHOV 5HFDOO WKDW LW ZDV DW WKLV SRLQW LQ GHYHORSLQJ WKH FRQWUROOHUV WKH ILUVW WLPH WKDW WKH SUREn OHP ILUVW VXUIDFHG &OHDUO\ WKH QHZ FRQWUROOHUV DUH SHUIRUPLQJ PXFK EHWWHU WKDQ WKH RULJLn QDO RQHV ZKRVH UHVXOWV KDYH EHHQ LQFOXGHG LQ )LJXUHV DQG IRU UHIHUHQFH

PAGE 181

2QOLQH 4XDQWLILFDWLRQ 7KH RIIOLQH TXDQWLILFDWLRQ ILUVW LGHQWLILHG D SUREOHP ZLWK WKH FRQWURO GHVLJQV DQG QRZ LQGLFDWHV WKDW LW KDV EHHQ VROYHG 7KH SURRI KRZHYHU LV DOZD\V LQ WKH SXGGLQJ RU DW OHDVW LQ WKH UHVSRQVH RI WKH DFWXDO SODQW ,W VWDQGV WR UHDVRQ WKDW LI ZH ZHUH QRW DEOH WR GHYHORS PRGHOV ZLWK FRQVLVWHQW NQRZOHGJH RI WKH SURFHVV WKHQ WKHUH LV OLWWOH KRSH WKDW WKH UHVXOWLQJ FRQWUROOHUV ZRXOG EH DEOH WR LPSURYH WKH SURFHVV &RQVLVWHQW PRGHOV RI WKH SURFHVV KRZn HYHU GR QRW QHFHVVDULO\ LPSO\ WKDW WKH UHVXOWLQJ FRQWUROOHUV ZLOO EH DEOH WR LPSURYH WKH SURFHVV 7R TXDQWLI\ WKH DELOLW\ RI WKH FRQWUROOHUV WR LPSURYH WKH SURFHVV WKH PHWKRGRORJ\ RXWn OLQHG LQ 6HFWLRQ 2QOLQH 4XDQWLILFDWLRQ LV DSSOLHG 2QFH DJDLQ H[SHULPHQWV ZHUH FRQGXFWHG ZLWK WKH VWHDG\VWDWH RSWLPL]HU 7KH UHVXOWV WKLV WLPH KRZHYHU DUH YHU\ HQFRXUn DJLQJ )LJXUH VKRZV WKH PHDVXUHG 12[ FKDQJH EHWZHHQ WKH EDVHOLQH DQG H[SHULPHQW DORQJ ZLWK WKH FRUUHVSRQGLQJ FKDQJH EHWZHHQ WKH H[SHULPHQW DQG YDOLGDWLRQ UHJLRQV 7KHUH LV D VLJQLILFDQW GHFUHDVH LQ 12[ 7KH DYHUDJH 12[ SHUFHQW FKDQJH IRU DOO H[SHULn PHQWV LV b )LJXUH VKRZV WKH FRUUHVSRQGLQJ &2 FKDQJHV +HUH WKH FRQWUROOHU ZDV DEOH WR KDYH D VLJQLILFDQW LPSDFW RQ &2 7KH DYHUDJH &2 IRU ERWK EDVHOLQH DQG H[SHULPHQW LV DSSUR[Ln PDWHO\ SSP EXW WKH YDULDQFH RI WKH EDVHOLQH GDWD LV SSP ZKLOH WKH H[SHULPHQW VWDQGDUG GHYLDWLRQ LV

PAGE 182

)LJXUH &KDQJH LQ 12[ IRU UHYLVHG VWHDG\VWDWH FRQWUROOHU H[SHULPHQWV r %DVHOLQH A ([SHULPHQW f§ff§9DOLGDWLRQ ([SHULPHQW )LJXUH )LQDO &2 OHYHO IRU UHYLVHG VWHDG\VWDWH FRQWUROOHU H[SHULPHQWV 7KLV RIIOLQH TXDQWLILFDWLRQ SURFHGXUH ZDV UHSHDWHG IRU HDFK RI WKH IRXU FRQWURO GHVLJQV DQG WKH 12[ DQG &2 UHVXOWV DUH VXPPDUL]HG LQ )LJXUHV DQG ,QFOXGHG LQ WKLV ILJXUHV DUH WKH HVWLPDWHG UHVXOWV WKDW HDFK FRQWUROOHU WKRXJKW LW ZRXOG DFKLHYH DFURVV WKH VDPH

PAGE 183

GDWDVHW WKLV ZDV DFFRPSOLVKHG E\ XVLQJ D VLPXODWRU IRU WKH SODQW DV GHVFULEHG LQ 6HFWLRQ 2IIOLQH 4XDQWLILFDWLRQ )LJXUH $YHUDJH SHUFHQW 12[ UHGXFWLRQ IRU RQOLQH H[SHULPHQWV 7KHUH DUH VRPH LQWHUHVWLQJ REVHUYDWLRQV DERXW WKHVH UHVXOWV f 7KH SHUIRUPDQFH RI WKH 0,& FRQWUROOHU ZDV VXEVWDQWLDOO\ SRRUHU WKDQ WKH RWKHU FRQWURO GHVLJQV LQ ERWK RIIOLQH DQG RQOLQH DQDO\VLV f 7KHUH RQOLQH SHUIRUPDQFH RI WKH VWHDG\VWDWH RSWLPL]HU ZDV EHWWHU WKDQ HLWKHU RI WKH G\QDPLF FRQWUROOHUV f 7KH G\QDPLF FRQWUROOHUV SHUIRUPHG VLJQLILFDQWO\ EHWWHU LQ RIIOLQH DQDO\n VLV WKDW WKH\ GLG IRU RQOLQH DQDO\VLV 7KH SRRU SHUIRUPDQFH RI WKH 0,& FRQWUROOHU FDQ EH H[SODLQHG E\ LWV WUHDWPHQW RI FRQn VWUDLQWV 5HFDOO WKDW WKH 09 FRQVWUDLQWV ZHUH WUHDWHG DV SHQDOW\ IXQFWLRQV IRU WKH RQOLQH RSWLPL]HU DQG WKDW WKH &2 PD[LPXP FRQVWUDLQW ZDV DGGUHVVHG E\ SODFLQJ &2 LQ WKH &9 VHW DQG IL[LQJ LW WR SSP 7KLV OHDYHV RQO\ D VLQJOH GHJUHH RI IUHHGRP IRU WKH RQOLQH RSWLPL]HU ZKLFK LV WKH VLQJOH &9 RI 12[ 7KH SUREOHP LV WKDW ZKHQ WKH RQOLQH RSWLPL]HU VWDUWV PRYLQJ 12[ WR PHHW LWV REMHFWLYH IXQFWLRQ LW ZLOO KDYH WR VWRS DV VRRQ DV RQH RI WKH

PAGE 184

09 FRQVWUDLQWV H[HUFLVH WKHLU SHQDOW\ IXQFWLRQV 7KHUH DUH QR RWKHU GHJUHH RI IUHHGRP IRU WKH RSWLPL]HU WR H[SORUH RWKHU WKDQ WR VLPSO\ VWRS PRYLQJ 12[ 1RWLFH WKDW DOWKRXJK SHQDOW\ IXQFWLRQ DUH DOVR HPSOR\HG LQ WKH VWHDG\VWDWH RSWLPL]HU WKLV SUREOHP GRHV QRW H[LVW +HUH WKH RQOLQH RSWLPL]HU KDV 1n GHJUHHV RI IUHHGRP WKH QXPEHU RI 09V VR WKDW ZKHQ D 09 FRQVWUDLQWfV SHQDOW\ IXQFWLRQ LV H[HUFLVHG LW FDQ VLPn SO\ VWRS PRYLQJ WKDW 09 )XUWKHUPRUH ZKHQ D 69 RU &9 FRQVWUDLQWfV SHQDOW\ IXQFWLRQ LV H[HUFLVHG WKH RSWLPL]HU FDQ H[SORUH RWKHU FRPELQDWLRQV RI 09V WR PDLQWDLQ WKLV FRQn VWUDLQW ZKLOH VWLOO WU\LQJ WR ORZHU 12[ )LJXUH $YHUDJH SHUFHQW &2 UHGXFWLRQ DERYH SSP IRU RQOLQH H[SHULPHQWV 7KH REVHUYDWLRQ WKDW WKH VWHDG\VWDWH RSWLPL]HU RXWSHUIRUPHG WKH G\QDPLF FRQWUROOHUV LV PRVW OLNHO\ D IXQFWLRQ RI WKH RQOLQH WHVW SURFHGXUH 5HFDOO WKDW WKH FRQWUROOHUV ZHUH GHSOR\HG RQOLQH LQ DQ DGYLVRU\ PRGH +HUH WKH FRQWUROOHUV RSWLPDO 09 VHWSRLQWV DUH IRUn ZDUGHG WR DQ RSHUDWRU ZKR LV WKHQ UHVSRQVLEOH IRU PDQLSXODWLQJ WKH DFWXDO SODQW VHWSRLQWV 2QH KDV WR UHPHPEHU WKDW WKH RSHUDWRUV DUH UHVSRQVLEOH IRU WKH XQLW ZKLFK LV ZRUWK VHYHUDO

PAGE 185

KXQGUHG PLOOLRQ GROODUV 2SHUDWRUV RSHUDWH WKH XQLW EDVHG RQ \HDUV RI H[SHULHQFH DQG GRJPD 6RUWLQJ RXW ZKLFK LV ZKLFK LV QRW DQ HDV\ WDVN DQG LQYDULDEO\ WKH RSHUDWRU ZLOO DQG VKRXOG PDNH WKH ILQDO FDOO DERXW ZKHWKHU WKH UHTXHVWHG VHWSRLQWV DUH VDIH 7KH UHVXOW ZDV WKDW 09 VHWSRLQWV ZHUH QRW PDGH RQ UHJXODU LQWHUYDOV DQG ZKDW WKH RSWLPL]HU UHFRPn PHQGHG DV VLPXOWDQHRXV 09 PRYHV PLJKW KDYH EHHQ LPSOHPHQWHG RQH DW D WLPH RYHU D SHULRG DV ORQJ DV PLQXWHV )XUWKHUPRUH UHFDOO WKDW WKH PHDVXUHPHQW PHWKRGRORJ\ SUHn VHQWHG LQ 6HFWLRQ 0HDVXUHPHQW 0HWKRGRORJ\ UHTXLUHG RSHUDWLQJ WKH XQLW XQGHU VWHDG\VWDWH FRQGLWLRQV &OHDUO\ DQ\ G\QDPLF DGYDQWDJH WKDW WKHVH FRQWUROOHUV PLJKW KDYH KDG ZDV SUREDEO\ ORVW 7KH REVHUYDWLRQ WKDW WKH G\QDPLF FRQWUROOHUV SHUIRUPHG VLJQLILFDQWO\ EHWWHU LQ RIIOLQH DQDO\VLV WKDW WKH\ GLG IRU RQOLQH DQDO\VLV FDQ DOVR EH FKDONHG XS WR WKH RQOLQH WHVW SURFHn GXUH ,Q IDFW WKLV UHVXOW UHLQIRUFHV WKLV FRQFOXVLRQ 7KH IDFW WKDW WKH RIIOLQH DQDO\VLV GHPn RQVWUDWHG D VLJQLILFDQW EHQHILW WR XVLQJ G\QDPLF FRQWUROOHUV SURYLGHV LQFHQWLYH WR FRQWLQXH ZRUN WRZDUG D FORVHGORRS FRQILJXUDWLRQ

PAGE 186

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f DUH VWUDLJKWIRUZDUG WR LPSOHPHQW f DFFRXQW IRU GHSHQGHQW LQWHUQDO SURFHVV VWDWHV DQG f FDQ GHDO ZLWK FRUUHODWHG SURFHVV YDULDEOHV &RQWULEXWLRQV $SSOLFDWLRQ%DVHG 1HXURFRQWURO ,PSOHPHQWDWLRQ 0HWKRGRORJ\ $ QHXURFRQWURO LPSOHPHQWDWLRQ PHWKRGRORJ\ ZDV GHYHORSHG ZKHUHE\ D SURFHVV HQJLn QHHU ZLWK UHDVRQDEOH NQRZOHGJH DERXW WKH SURFHVV YDULDEOHV FDQ GHYHORS DGYDQFHG QHXURn FRQWUROOHUV 7KH SURFHVV HQJLQHHU ZDV DEOH WR VLPSO\ FODVVLI\ SURFHVV YDULDEOHV DV 09V '9V 69V DQG &9V DQG WKH PHWKRGRORJ\ ZDV DEOH WR DXWRPDWH WKH GHYHORSPHQW RI HDFK VWDJH RI WKH FRQWUROOHU $ VHW RI VXSSRUWLQJ DOJRULWKPV ZDV SURSRVHG LPSOHPHQWHG DQG YDOLGDWHG

PAGE 187

6WDWH6SDFH 1HXURFRQWURO 'HVLJQV 6HYHUDO QHXURFRQWURO GHVLJQV ZHUH LPSOHPHQWHG UDQJLQJ IURP SXUHO\ LQSXWRXWSXW WR IXOO VWDWHVSDFH &RQVLVWHQW ZLWK WKH OLWHUDWXUH RQ QHXUDO QHWZRUN PRGHOLQJ >@ WKH IXOO VWDWHVSDFH DUFKLWHFWXUH ZDV GLIILFXOW WR WUDLQ DQG ZDV WKXV QRW WKH EHVW SHUIRUPHU 7KH EHVW SHUIRUPHU KRZHYHU ZDV D SDUWLDO VWDWHVSDFH PRGHO 7KLV ZRUN GHPRQVWUDWHG WKDW DYDLOn DEOH SURFHVV NQRZOHGJH FDQ EH XVHG WR FUHDWH SDUWLDO VWDWHVSDFH PRGHOV ZKLFK FDQ VLJQLIn LFDQWO\ LPSURYH WKH RYHUDOO FRQWUROOHU SHUIRUPDQFH 0HWKRGV IRU 'HDOLQJ ZLWK &RUUHODWLRQ 7KH SULPDU\ OLPLWDWLRQV RI QHXURFRQWURO LGHQWLILHG GXULQJ WKLV VWXG\ DUH GLUHFWO\ UHODWHG WR SUREOHPV WKDW DULVH ZKHQ PRGHOLQJ ZLWK FRUUHODWHG LQSXW YDULDEOHV ,Q IDFW WKH RYHUn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n WLRQV LV D GDWD PLQLQJ DSSURDFK IRU DSSOLFDWLRQV ZKHUH PDQ\ UHSUHVHQWDWLRQV DUH DYDLODEOH RI WKH VDPH XQGHUO\LQJ SURFHVV YDULDEOHV 7KLV DSSURDFK ZLOO QRW ZRUN IRU DSSOLFDWLRQV ZKHUH PXOWLSOH FKRLFHV IRU D SURFHVV YDULDEOH DUH QRW DYDLODEOH DQG WKH DXWKRU SURSRVHV XVLQJ WKLV QHZ PHWULF DV DQ REMHFWLYH IRU WKH OHDUQLQJ UXOH DV DQ H[WHQVLRQ WR WKLV ZRUN

PAGE 188

$FFXUDWH &RPEXVWLRQ 0RGHOV 1HXUDO QHWZRUNV KDYH IRXQG DQ DSSOLFDWLRQV QLFKH ZKHUH WKLV UREXVW SUHGLFWRU KDV GHPRQVWUDWHG WKH DELOLW\ WR RXW IRUHFDVW PRUH WUDGLWLRQDO PHWKRGV 2QH RI WKH SUREOHPV ZLWK PRYLQJ QHXUDO QHWZRUNV IURP DQ DFDGHPLF LQWHUHVW WR DQ DFFHSWHG PRGHOLQJ PHWKRGn RORJ\ LV WKH ODFN RI VWDQGDUGL]HG UHSRUWLQJ 7KH YDVW PDMRULW\ RI QHXUDO QHWZRUN DSSOLFDn WLRQV WR GDWH DSSO\ WKH LQIHUHQFHV RI D PRGHO ZLWKRXW NQRZOHGJH RI WKHLU VWDWLVWLFDO VLJQLILFDQFH 7KH ODFN RI VWDQGDUGL]HG UHSRUWLQJ KDV QRW FORXGHG WKH VXFFHVV RI QHXUDO QHWn ZRUN EHFDXVH PRVW RI WKHVH DSSOLFDWLRQV UHO\ VROHO\ RQ WKH PRGHOfV DELOLW\ WR IRUHFDVW ZLWKRXW UHJDUG IRU ZKDW UHODWLRQVKLSV WKH PRGHO KDV LQIHUUHG IURP WKH XQGHUO\LQJ SURFHVV $V FOHDUO\ GHPRQVWUDWHG LQ WKLV ZRUN WKH VDPH FDQQRW EH VDLG IRU QHXURFRQWURO KRZn HYHU 7KH UHVXOWV SUHVHQWHG KHUH UDLVH VLJQLILFDQW TXHVWLRQV DERXW WKH PHWULFV EHLQJ XVHG WR HYDOXDWH PRGHO SHUIRUPDQFH LQ WKH ILHOG RI QHXUDO QHWZRUNV 0HWULFV ZKLFK HYDOXDWH WKH SUHGLFWLRQV PDGH E\ D PRGHO GR QRW LPSO\ WKDW WKH PRGHO KDV OHDUQLQJ WKH IXQGDPHQWDO FDXVHDQGHIIHFW UHODWLRQVKLSV ZLWKLQ WKH SURFHVV $ QHZ PHWULF ZDV SURSRVHG DQG GHPRQn VWUDWHG WR RYHUFRPH WKLV OLPLWDWLRQ 1RYHO &RPEXVWLRQ &RQWUROOHU $Q RQOLQH 12[ RSWLPL]HU ZDV GHYHORSHG DQG HYDOXDWHG 7KLV RSWLPL]HU GHPRQVWUDWHG D FRQVLVWHQW b UHGXFWLRQ LQ WKH RYHUDOO 12[ HPLVVLRQV IURP WKH SRZHU SODQW $IWHUZRUG 7KH EHQHILWV GHPRQVWUDWHG LQ WKLV VWXG\ DW WKH &DQDO (OHFWULF JHQHUDWLQJ VWDWLRQ KDYH EHHQ PDLQWDLQHG IRU PRUH WKDQ WZR \HDUV 6LQFH WKHVH LQLWLDO UHVXOWV QHXURFRQWUROOHUV KDYH

PAGE 189

EHHQ GHSOR\HG RQ JHQHUDWLQJ XQLWV WKURXJKRXW WKH 8QLWHG 6WDWHV ,Q DGGLWLRQ WR EHLQJ LQVWDOOHG RQ PRUH XQLWV WKH QHXURFRQWUROOHUV KDYH EHHQ H[WHQGHG ZLWK f FORVHGORRS LPSOHPHQWDWLRQV f D VRIWZDUH SURJUDP WKDW DXWRPDWHV WKH QHXURFRQWURO GHVLJQ PHWKRGRORn JLHV f RQOLQH DXWRPDWLRQ RI WKH YDULDEOH DQG DUFKLWHFWXUH SUXQLQJ PHWKRGRORn JLHV XVLQJ HYROXWLRQDU\ FRPSXWLQJ WHFKQLTXHV DQG )XWXUH 'LUHFWLRQ 7KH DXWKRU RIIHUV WKUHH LPSRUWDQW GLUHFWLRQV IRU H[WHQGLQJ WKLV ZRUN f GHYHORSPHQW RI OHDUQLQJ DOJRULWKPV WKDW H[SOLFLWO\ DFFRXQW IRU 09 VHQn VLWLYLW\ VWDQGDUG HUURUV f DOJRULWKPV WR VXSSRUW DXWRPDWLF GHWHUPLQDWLRQ RI YDULDEOH W\SH DQG f RQOLQH DGDSWDWLRQ RI WKH PRGHOUHIHUHQFH DGDSWLYH FRQWUROOHU XVLQJ UHHQn IRUFHPHQW OHDUQLQJ VWUDWHJLHV

PAGE 190

$33(1',; 7DEOH (VVHQVLDO WDJ OLVW 'HVFULSWRU 7\SH 8QLW 2)$ 0DVWHU %LDV 09 8QLW 2)$ 0DVWHU 2XW 09 8QLW 2)$ 0DVWHU %LDV 09 8QLW 2)$ 0DVWHU 2XW 09 $3), $ $LU ,Q 7HPS 09 $3+ % $LU ,Q 7HPS 09 )XHO 7HPS )LUHG 09 %RLOHU (IILFLHQF\ DEEU /RVVHVf &9 %QU $WP 6WP 3UHVV 09 ([LW %6LGH 69 ([LW $6LGH 69 )OXH *DV &2 %6LGH 69 )OXH *DV &2 $6LGH 69 *HQHUDWHG 0: 69 3$6$ 3RV 09 3$6% 3RV 09 3$6$ 3RV 09 3$6 % 3RV 09 3$6 $ 3RV 09 3$6% 3RV 09 3$6 $ 3RV 09 3$6% 3RV 09 3$6 $ 3RV 09 3$6% 3RV 09 3$6 $ 3RV 09 3$6% 3RV 09 3$6% 3RV 09 3$6 $ 3RV 09 3$6% 3RV 09 3$6 $ 3RV 09 8 3$6 0DVW *DV %LDV 09 8&(0 &2 69 )XHO 2LO )ORZ 09 )8(/ *$6 35(6 ,1',&$7,21 69 )8(/ *$6 )/2: ,1',&$7,21 69 ,')% ,1/ 9$1( 09 ,') $,1/ 9$1( 09 'HVFULSWRU 7\SH 3$6 $ %LDV 09 3$6% %LDV 09 3$6 $ %LDV 09 3$6 % %LDV 09 3$6 $ %LDV 09 3$6 % %LDV 09 3$6 $ %LDV 09 3$6% %LDV 09 3$6 $ %LDV 09 3$6% %LDV 09 3$6 $ %LDV 09 3$6 % %LDV 09 3$6 % %LDV 09 3$6 $ %LDV 09 3$6% %LDV 09 3$6 $ %LDV 09 +HDW 5DWH &9 *5 )DQ +SSU 'PSU % 3RV 09 *5 )DQ % ,QOHW 'PSU 3RV 09 *5 )DQ $ ,QOHW 'PSU 3RV 09 5HODWLYH (IILFLHQF\ &9 :LQGER[ 3UHVVXUH 69 &(0 %DURPHWULF 3UHVVXUH 69 8QLW )' )DQ %LDV 09 8QLW +HDW 5DWH &9 &(0 1R[ &9 %RLOHU 7ULP 8QLW 09 &(0 2SDFLW\ 8QLW &9 &(0 6WDFN 7HPS 69 $,5 35(+75 $ *$6 287 7(03 69 $,5 35(+75 % *$6 287 7(03 69 8QLW 6$6 0DVWHU 69 8QLW 6$6 0DVWHU %LDV 69 8QLW 3$ 6KURXG 0DVW 09 8QLW 3$6 0DVWHU %LDV 09 8 3$ 6KURXG 0DVW *DV 09

PAGE 191

5()(5(1&(6 >@ $QGHUVRQ & : /HDUQLQJ WR FRQWURO DQ LQYHUWHG SHQGXOXP XVLQJ QHXUDO QHWZRUNV ,((( &RQWURO 6\VWHPV 0DJD]LQH SDJHV >@ $VWURP t 0F$YR\ 7 ,QWHOOLJHQW FRQWURO DQ RYHUYLHZ DQG HYDOXDWLRQ +DQGERRN RI ,QWHOOLJHQW &RQWURO 9DQ 1RVWUDQG 5HLKKROG 1HZ @ $VWURP t :LWWHQPDUN % $GDSWLYH &RQWURO $GGLVRQ:HVOH\ 5HDGLQJ 0$ >@ $UFKHWWL ) $ VDPSOLQJ WHFKQLTXH IRU JOREDO RSWLPL]DWLRQ 7RZDUGV *OREDO 2SWLPLn ]DWLRQ $PVWHUGDP >@ %DNHU : / t )DUUHOO $ $Q LQWURGXFWLRQ WR FRQQHFWLRQLVW OHDUQLQJ FRQWURO V\Vn WHPV +DQGERRN RI ,QWHOOLJHQW &RQWURO 9DQ 1RVWUDQG 5HLKKROG 1HZ @ %DUQHV & $SSOLFDWLRQV RI QHXUDO QHWZRUNV WR SURFHVV FRQWURO DQG PRGHOLQJ ,QWHUn QDWLRQDO &RQIHUHQFH RQ $UWLILFLDO 1HXUDO 1HWZRUNV f (VSRR SDJHV >@ %DUWR $* 6XWWRQ 56 t $QGHUVRQ & 1HXURQ OLNH HOHPHQWV WKDW FDQ VROYH GLIILn FXOW OHDUQLQJ FRQWURO SUREOHPV ,((( 7UDQVDFWLRQV RQ 6\VWHPV 0DQ t &\EHUQHWLFV >@ %DUWR $* 6XWWRQ 56 t :DWNLQV &-&+ 6HTXHQWLDO GHFLVLRQ SUREOHPV DQG QHXUDO QHWZRUNV $GYDQFHV LQ 1HXUDO ,QIRUPDWLRQ 3URFHVVLQJ 6\VWHPV >@ %D[W : t :KLWH + %RRWVWUDSSLQJ FRQILGHQFH ,QWHUYDOV IRU &OLQLFDO ,QSXW 9DULn DEOH (IIHFWV LQ D 1HWZRUN 7UDLQHG WR ,GHQWLI\ WKH 3UHVHQFH RI $FXWH 0\RFDUGLDO ,QIn DUFWLRQ 7HFK 5HS >@ %HOOR 0* (QKDQFHG WUDLQLQJ DOJRULWKPV DQG LQWHJUDWHG WUDLQLQJDUFKLWHFWXUH VHOHFWLRQ IRU PXOWLOD\HU SHUFHSWURQ QHWZRUNV ,((( 7UDQVDFWLRQV RQ 1HXUDO 1HWn ZRUNV SDJHV >@ %HOOPDQ 5 ( '\QDPLF 3URJUDPPLQJ 3ULQFHWRQ 8QLYHUVLW\ 3UHVV 3ULQFHWRQ 1>@ %HUWVHNDV '3 t 7VLWVLNOLV -1 1HXUR'\QDPLF 3URJUDPPLQJ $WKHQD 6FLHQWLILF %HOPRQW 0$

PAGE 192

>@ %LVKRS & 1HXUDO 1HWZRUNV IRU 3DWWHUQ 5HFRJQLWLRQ &ODUHQGRQ 3UHVV 2[IRUG >@ %UHQW 5 $OJRULWKPV IRU 0LQLPL]DWLRQ ZLWKRXW 'HULYDWLYHV 3UHQWLFH +DOO (QJOH ZRRN &OLIIV 1>@ %U\VRQ $( t +R < $SSOLHG 2SWLPDO &RQWURO +HPLVKSHUH 3XEOLVKLQJ &RUSRUDn WLRQ 1HZ @ &DPSROXFFL 3 3LD]]D ) t 8QFLQL $ 2QOLQH $OJRULWKPV IRU 1HXUDO 1HWZRUNV ZLWK +5 6\QDSVHV ,((( ,QW &RQI RQ 1HXUDO 1HWZRUNV ,&11 3HUWK $XVWUDOLD >@ GH 9ULHV % t 3ULQFLSH -& $ 7KHRU\ IRU 1HXUDO 1HWZRUNV ZLWK 7LPH 'HOD\V $GYDQFHV LQ 1HXUDO ,QIRUPDWLRQ 3URFHVVLQJ 6\VWHPV 0RUJDQ .DXIPDQ SDJHV >@ )HOGNDPS /$ t 3XVNRULXV *9 7UDLQLQJ FRQWUROOHUV IRU UREXVWQHVV PXOWLVWUHDP '(.) 3URFHHGLQJV RI ,((( ,QWHUQDWLRQDO &RQIHUHQFH RQ 1HXUDO 1HWZRUNV 2UODQGR )/ SDJHV >@ )HOGNDPS /$ t 3XVNRULXV *9 'HFRXSOHG .DOPDQ 7UDLQLQJV RI QHXUDO DQG IX]]\ FRQWUROOHUV IRU DXWRPRWLYH V\VWHPV 3URFHHGLQJV RI ,((( 976 :RUNVKRS RQ )X]]\ DQG 1HXUDO 6\VWHP DQG 9HKLFOH $SSOLFDWLRQV 7RN\R >@ )XNDQDJD 6WDWLVWLFDO 3DWWHUQ 5HFRJQLWLRQ QG (GLWLRQ +DUFRXUW %UDFH -RYDQRY LFK &DPEULGJH 0DVVDFKXVHWWV >@ *H 53 $ ILOOHG IXQFWLRQ PHWKRG IRU ILQGLQJ JOREDO PLQLPL]HU RI D IXQFWLRQ RI VHYn HUDO YDULDEOHV 0DWK 3URJUDP SDJHV >@ +DONMDHU 6 t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n PDWLRQ 3URFHVVLQJ 6\VWHPV SDJHV

PAGE 193

>@ +DUGOH : $SSOLHG 1RQSDUDPHWULF 5HJUHVVLRQ &DPEULGJH 8QLYHUVLW\ 3UHVV >@ +LQWRQ *( /HDUQLQJ WUDQVODWLRQ LQYDULDQW UHFRJQLWLRQ LQ PDVVLYHO\ SDUDOOHO QHWn ZRUNV 3URFHHGLQJV 3$5/( &RQIHUHQFH RQ 3DUDOOHO $UFKLWHFWXUHV DQG /DQJXDJHV (XURSH SDJHV >@ +U\FHM 7 0RGHOEDVHG WUDLQLQJ PHWKRGV IRU QHXUDO FRQWUROOHUV 3URFHHGLQJV RI WKH ,QWHUQDWLRQDO &RQIHUHQFH RI $UWLILFLDO 1HXUDO 1HWZRUNV %ULJKWRQ >@ +U\FHM 7 1HXURFRQWURO -RKQ :LOH\ t 6RQV ,QF 1HZ @ -DQJ -65 6XQ &7 t 0L]XWDQL ( 1HXUR)X]]\ DQG 6RIW &RPSXWLQJ $ &RPn SXWDWLRQDO $SSURDFK WR /HDUQLQJ DQG 0DFKLQH ,QWHOOLJHQFH 3UHQWLFH +DOO SDJHV >@ .DLODWK 7 /LQHDU 6\VWHPV 3UHQWLFH+DOO ,QF (QJOHZRRG &OLIIV 1>@ .DOPDQ 5( 2Q WKH JHQHUDO WKHRU\ RI FRQWURO V\VWHPV 3URFHHGLQJV RI WKH VW ,QWHUn QDWLRQDO &RQJUHVV RQ $XWRPDWLF &RQWURO 0RVFRZ SDJHV >@ .DOPDQ 5( 'HVLJQ RI VHOIRSWLPL]LQJ FRQWURO V\VWHPV 7UDQV $60( SDJHV >@ .R]D -5 *HQHWLF 3URJUDPPLQJ 7KH 0,7 3UHVV &DPEULGJH 0$ >@ /HYLQH :6 7KH &RQWURO +DQGERRN &5& 3UHVV >@ /HY\ $9 t 0RQWDOYR $ 7KH WXQQHOLQJ DOJRULWKP IRU WKH JOREDO PLQLPL]DWLRQ RI IXQFWLRQV 6,$0 6FL 6WDW &RPS SDJHV >@ /LSWDN %* 2SWLPL]DWLRQ RI 8QLW 2SHUDWLRQV &KLOWRQ %RRN &RPSDQ\ 5DGQRU 3HQQV\OYDQLD >@ 0LOOHU :7 t :HUERV 31HXUDO 1HWZRUNV IRU &RQWURO 7KH 0,7 3UHVV &DPn EULGJH 0$ >@ 1JX\HQ t :LGURZ % 7KH WUXFN EDFNHUXSSHU $Q H[DPSOH RI VHOIOHDUQLQJ LQ QHXUDO QHWZRUN 1HXUDO 1HWZRUNV IRU &RQWURO SDJHV >@ 1L[ t :HLJHQG $ (VWLPDWLQJ WKH PHDQ DQG YDULDQFH RI WKH WDUJHW SUREDELOLW\ GLVWULEXWLRQ 3URFHHGLQJV RI WKH ,-&11 SDJHV >@ 2PLGYDU 2 t (OOLRWW '/ 1HXUDO 6\VWHPV IRU &RQWURO $FDGHPLF 3UHVV 6DQ 'LHJR &$ >@3DSRXOLV $ 3UREDELOLW\ 5DQGRP 9DULDEOHV DQG 6WRFKDVWLF 3URFHVVHV UG HG 0F*UDZ+LOO 1HZ
PAGE 194

>@ 3HDUOPXWWHU % *UDGLHQW &DOFXODWLRQV IRU '\QDPLF 5HFXUUHQW 1HXUDO 1HWZRUNV $ 6XUYH\ ,((( 7UDQVDFWLRQV RQ 1HXUDO 1HWZRUNV 9RO 1R SDJHV >@ 3HUURQH 03 *HQHUDO DYHUDJLQJ UHVXOWV IRU FRQYH[ RSWLPL]DWLRQ 3URFHHGLQJV &RQQHFWLRQLVW 0RGHOV 6XPPHU 6FKRRO SDJHV >@ 3HUURQH 03 t &RRSHU /1 :KHQ QHWZRUNV GLVDJUHH HQVHPEOH PHWKRGV IRU K\EULG QHXUDO QHWZRUNV $UWLILFLDO 1HXUDO 1HWZRUNV IRU 6SHHFK DQG 9LVLRQ SDJHV >@ 3RJJLR 7 t *LURVL ) 5HJXODUL]DWLRQ DOJRULWKPV IRU OHDUQLQJ WKDW DUH HTXLYDOHQW WR PXOWLOD\HU QHWZRUNV 6FLHQFH SDJHV >@ 3RODN ( &RPSXWDWLRQDO 0HWKRGV LQ 2SWLPL]DWLRQ $FDGHPLF 3UHVV 1HZ @ 3RPHUOHDX '$ 1HXUDO QHWZRUNEDVHG YLVLRQ SURFHVVLQJ IRU DXWRQRPRXV URERW JXLGDQFH $SSOLFDWLRQV RI $UWLILFLDO 1HXUDO 1HWZRUNV ,, YROXPH SDJHV 63,( %HOOLQJKDP :DVKLQJWRQ >@ 3XVNRULXV *9 t )HOGNDPS /$ 'HFRXSOHG H[WHQGHG .DOPDQ ILOWHU WUDLQLQJ RI IHHGIRUZDUG OD\HUHG QHWZRUNV 3URFHHGLQJV RI WKH -RLQW &RQIHUHQFH RQ 1HXUDO 1HWn ZRUNV f 6HDWWOH SDJHV >@ 3XVNRULXV *9 t )HOGNDPS /$ 1HXURFRQWURO RI QRQOLQHDU G\QDPLFDO V\VWHPV ZLWK .DOPDQ ILOWHU WUDLQHG UHFXUUHQW QHWZRUNV ,((( 7UDQVDFWLRQV RQ 1HXUDO 1HWn ZRUNV 6HDWWOH SDJHV >@ 5DPDQDWKDQ 5 ,QWURGXFWRU\ (FRQRPHWULFV +DUFRXUW %UDFH -RYDQRYLFK 3XEOLVKHUV >@ 5DR 66 (QJLQHHULQJ 2SWLPL]DWLRQ 7KHRU\ DQG 3UDFWLFH -RKQ :LOH\ t 6RQV ,QF 1HZ @ 5XPHOKDUW +LQWRQ t :LOOLDPV 5 /HDUQLQJ LQWHUQDO UHSUHVHQWDWLRQV E\ HUURU EDFNSURSDJDWLRQ 3DUDOOHO 'LVWULEXWHG 3URFHVVLQJ ([SORUDWLRQV LQ WKH 0LFURn VWUXFWXUH RI &RJQLWLRQ SDJHV 0,7 3UHVV 0$ >@ 6FKDUI // 6WDWLVWLFDO 6LJQDO 3URFHVVLQJ 'HWHFWLRQ (VWLPDWLRQ DQG 7LPH 6HULHV $QDO\VLV $GGLVRQ:HVOH\ 3XEOLVKLQJ &RPSDQ\ 1HZ @ 6LQJKDO 6 t :X / 7UDLQLQJ PXOWLOD\HU SHUFHSWURQV ZLWK WKH H[WHQGHG .DOPDQ DOJRULWKP $GYDQFHV LQ 1HXUDO ,QIRUPDWLRQ 3URFHVVLQJ 6\VWHPV 0RUJDQ .DXIPDQQ 6DQ 0DWHR &$

PAGE 195

>@ 6RIJH '$ t :KLWH $ 1HXUDO QHWZRUN EDVHG SURFHVV RSWLPL]DWLRQ DQG FRQn WURO 3URFHHGLQJV RI WKH ,((( &RQIHUHQFH RQ 'HFLVLRQ t &RQWURO SDJHV >@ 6RXOLH )) t *DOOLQDUL 3 ,QGXVWULDO $SSOLFDWLRQV RI 1HXUDO 1HWZRUNV :RUOG 6FLn HQWLILF 6LQJDSRUH >@ 6WDLE :( 7KH ,QWHOOLJHQW $UF )XUQDFH 1HXUDO QHWZRUNV UHYROXWLRQL]H VWHHOPDNn LQJ 3URFHHGLQJV RI WKH :RUOG &RQJUHVV RQ 1HXUDO 1HWZRUNV SDJHV >@ 7LEVKLUDQL 5 $ FRPSDULVRQ RI VRPH HUURU HVWLPDWHV IRU QHXUDO QHWZRUN PRGHOV 1HXUDO &RPSXWDWLRQ >@ 7LNKDQRY $1 t $UVHQLQ 9< 6ROXWLRQV RI ,OO3RVHG 3UREOHPV 9+ :LQVWRQ :DVKLQJWRQ '& >@ 7URXGHW 7 7RZDUGV SUDFWLFDO FRQWURO GHVLJQV XVLQJ QHXUDO FRPSXWDWLRQ 3URFHHGn LQJV RI WKH ,QWHUQDWLRQDO &RQIHUHQFH RQ 1HXUDO 1HWZRUNV YROXPH SDJHV >@ 7V\SNLQ <= )RXQGDWLRQV RI WKH 7KHRU\ RI /HDUQLQJ 6\VWHPV 1DXND 0RVNYD >@ 9DVLOLHY 29 2SWLPL]DWLRQ 0HWKRGV :RUOG )HGHUDWLRQ 3XEOLVKHUV &RPSDQ\ $WODQWD *( >@ :DOSROH 5( 3UREDELOLW\ DQG 6WDWLVWLFV IRU (QJLQHHUV DQG 6FLHQWLVWV 0DFPLOODQ 3XEOLVKLQJ &RPSDQ\ >@ :HLJHQG $6 t *HUVKHQIHOG 1$ 7LPH 6HULHV 3UHGLFWLRQ )RUHFDVWLQJ WKH )XWXUH DQG 8QGHUVWDQGLQJ WKH 3DVW $GGLVRQ :HVOH\ 3XEOLVKLQJ &RPSDQ\ >@ :HQGHPXWK $ 2SSHU 0 t .LQ]HO : 7KH HIIHFW RI FRUUHODWLRQV LQ QHXUDO QHWn ZRUNV 3KYV $ SDJHV >@ :KLWH '$ t 6RIJH '$ +DQGERRN RI ,QWHOOLJHQW &RQWURO 1HXUDO )X]]\ DQG $GDSWLYH $SSURDFKHV 9DQ 1RVWUDQG 5HLQKROG 1HZ @ :LOOLDPV 5t =LVSHU $ /HDUQLQJ $OJRULWKP IRU &RQWLQXDOO\ 5XQQLQJ )XOO\ 5HFXUUHQW 1HXUDO 1HWZRUNV 1HXUDO &RPSXWDWLRQ 9RO SDJHV >@ :LOOLDPV 5t 3HQJ $Q (IILFLHQW *UDGLHQW%DVHG $OJRULWKP IRU 2Q/LQH 7UDLQLQJ RI 5HFXUUHQW 1HWZRUN 7UDMHFWRULHV 1HXUDO &RPSXWDWLRQ 9RO SDJHV >@:HUERV 3 %H\RQG UHJUHVVLRQ QHZ WRROV IRU SUHGLFWLRQ DQG DQDO\VLV LQ WKH EHKDYn LRUDO VFLHQFHV 3K 7KHVLV +DUYDUG 8QLYHUVLW\ &DPEULGJH 0$

PAGE 196

>@ :HUERV 3$GYDQFHG IRUHFDVWLQJ PHWKRGV IRU JOREDO FULVLV ZDUQLQJ DQG PRGHOV RI LQWHOOLJHQFH *HQHUDO 6\VWHPV
PAGE 197

%,2*5$3+,&$/ 6.(7&+ :HVOH\ &XUWLV /HIHEYUH ZDV ERP LQ $WKHQV *HRUJLD $XJXVW &XUW JUDGn XDWHG IURP *UDQGYLOOH +LJK 6FKRRO LQ *UDQGYLOOH 0, DQG FRQWLQXHG KLV HGXFDWLRQ DW 0LFKLJDQ 6WDWH 8QLYHUVLW\ 'XULQJ KLV IUHVKPDQ \HDU DW 0LFKLJDQ 6WDWH &XUW DFFHSWHG D FRRSHUDWLYH HGXFDWLRQ SRVLWLRQ ZLWK ,QWHUQDWLRQDO %XVLQHVV 0DFKLQHV ,%0f LQ 2ZHJR 1< &XUW FRQWLQXHG KLV HGXFDWLRQ VSOLWWLQJ HDFK \HDU HYHQO\ EHWZHHQ 0LFKLJDQ 6WDWH DQG ,%0 $IWHU VHYHUDO \HDUV &XUW OHIW 0LFKLJDQ 6WDWH WR ILQLVK KLV XQGHUJUDGXDWH HGXFDWLRQ DW WKH 6WDWH 8QLYHUVLW\ RI 1HZ

xml version 1.0 encoding UTF-8
REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd
INGEST IEID ESPED3W9Y_5JHR3P INGEST_TIME 2015-02-10T19:51:02Z PACKAGE AA00024529_00001
AGREEMENT_INFO ACCOUNT UF PROJECT UFDC
FILES