Group Title: Department of Computer and Information Science and Engineering Technical Reports
Title: Simulation-based real-time decision making for route planning
Full Citation
Permanent Link:
 Material Information
Title: Simulation-based real-time decision making for route planning
Series Title: Department of Computer and Information Science and Engineering Technical Reports
Physical Description: Book
Language: English
Creator: Lee, Jin Joo
Fishwick, Paul A.
Publisher: Department of Computer and Information Science and Engineering, University of Florida
Place of Publication: Gainesville, Fla.
Copyright Date: 1995
 Record Information
Bibliographic ID: UF00095335
Volume ID: VID00001
Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.


This item has the following downloads:

1995180 ( PDF )

Full Text


Jin Joo Lee
Paul A. Fishwick

Department of Computer & Information Science & Engineering
University of Florida
Gainesville, Florida 32611, U.S.A.


Decision making is an active area of research in simu-
lation, systems engineering and artificial intelligence.
One subset area of decision making, automated route
planning, is covered in our paper with our approach
being based on the technique of simulation rather
than on purely heuristic or geometric techniques.
This new technique is called simulation-based plan-
ning. Simulation-based planning is useful for route
planning under various conditions including uncer-
tain locations and events with potential adversarial
activity. We claim that it is only by using simulation
that one can make the most effective plan in uncertain
and complex environments. An experimental design
is outlined along with our plans for further develop-


At a fundamental level, general AI planning, decision
making, intelligent control and route (path) planning
in robotics, all strive to solve a common problem
based on some model of a given process, determin-
ing what actions will affect the process in a desired
way. The problem appears different because each area
deals with different levels of abstraction and applica-
Planning becomes very complex for real world plan-
ning problems that take place in an environment over
which the planner has no control, such as with an-
other agent or an enemy or when there is uncer-
tainty of available information or agents' reactions.
In such cases, accurate prediction of the resulting
states of plan execution will be difficult. To over-
come this increase in complexity of reasoning, many
new approaches have been introduced (Schoppers
1 'I; Dean and Kanazawa 1 I -; Kanazawa and Dean
1989; Hammond 1989). To handle these uncertain-
ties in real time is the ultimate goal of any planner.

Simulation-Based Planning (SBP) can solve the prob-
lem of prediction of uncertain environment by allow-
ing the use of multiple simulations of simulation mod-
els to predict the behavior of individual objects us-
ing data sampled from a predetermined distribution.
Moreover, SBP can produce plans in real time since
we can allow simulation of plans at different levels
of abstraction. A high-level, simulation can be done
by simply sampling numbers from a distribution or
a complex low-level simulation can be done by simu-
lating the state change in greater detail at each time
step. Our previous work in military mission planning
with Computer Generated Forces shows preliminary
results of our SBP approach (Lee and Fishwick 1994).
The related background research areas to planning
are discussed in Section 2. In Section 3, we dis-
cuss SBP as a general methodology. Our example
problem-the rover route planning problem-is pre-
sented in Section 4 and experimental design issues are
discussed in Section 5. Finally, conclusions appear in
Section 6 and future work appears in Section 7.



To view the problem of route planning as part of a
larger picture and to gain better insight into the na-
ture of our problem, we will overview different areas
that are relevant to planning.

2.1 Artificial Intelligence

The general problem of planning in AI is commonly
identified with problems that are highly conceptual
where actions are of the form "Go To S,,I.. n iI I.. I
and bl; milk" (Russell and Norvig 1 '.). A plan
is an ordered set of these high-level actions. Here,
the concern is not how one will physically (at a de-
tailed level) get to a supermarket, but rather on the
ordered set of actions whose logical effects will satisfy
the goal. STRIPS (Fikes, Hart, and Nilsson 1972a;

Fikes, Hart, and Nilsson 1972b) is a classical exam-
ple of such an approach to planning. This approach
is reasonable if the execution of the produced plan is
not a responsibility of the planner. Difficulties arise
when execution becomes part of the planning system.

2.2 Intelligent Control

Intelligent control deals with problems that are more
physical and less conceptual. The problem of steer-
ing a cargo ship to a desired heading (Antsaklis and
Passino 1993) is a typical problem in intelligent con-
trol. Even though this whole task can be just simply
stated as -1. r the ship to heading x"; at the control
level, we are more concerned with tuning the control
input to physically steer the ship to a precise heading.

2.3 Decision Science and Game Theory

Decision science involves the creation of decisions
based on a game-theoretic foundation. Given the
current -I ,i.- of the world," one can embark upon
several courses of action (decisions) each of which
will yield a payoff or utility (Luce and Raiffa 1 i';).
Games can be naturally extended to continuous sys-
tems (Basar 1 I''.)(often found in simulation) by
equating the input (or forcing) function to a continu-
ously changing decision which alters the payoff given
the corresponding state changes.


In the simulation literature, simulation is defined as
"the discipline of designing a model of an actual or
theoretical physical system, executing the model on
a digital computer, and analyzing the execution out-
put" (Fishwick 1 ,'',). In the planning literature,
planning is defined as using models to formulate se-
quences of actions and given a sequence of actions,
models are used to simulate the future as it would oc-
cur if the actions were carried out (Dean and Wellman
1991). So simulation provides the robot with informa-
tion which can be used to suggest modifications or to
compare the proposed sequence with an alternative
sequence. Humans, who plan using a similar over-
all approach, have models built and stored in their
brains for most objects or systems. These models are
used to formulate sequences of actions which would
occur in the future if a plan was executed. There-
fore, once simulation models have been built for a
system, simulation can be used as a tool to provide
the system with information useful for evaluating an
hypothesis. It is logical that we employ simulation
within the planning process to gather information

about each candidate plan (sequence of actions) and
to compare them. Some recent work (West, Mellon,
Ramsey, Cleary, and Hofmann 1 r' i.) presented in this
conference also bears significant relation to our work.
Their work focuses on speeding up the execution of
models for military strike-planning using parallel and
distributed simulation. Our approach can potentially
use their method for further increasing the speed of
plan and route evaluation, in addition to using paral-
lel replication by simulating, in parallel, models with
separate factor values.
Once a plan is chosen for execution, the simula-
tion data that was generated during the planning
process can be used to match with the current real
world state. This can be compared to a common
technique used in adaptive control theory where a
reference model is compared with the actual perfor-
mance data in order to tune the controller to a desired
state (Antsaklis and Passino 1993).


Our focus is near-optimal route planning. Route
planning is in-between the higher level of symbolic
AI planning and the lower level of intelligent con-
trol. There are several application areas that are re-
lated to route planning. Mission planning within the
military domain almost always involves route plan-
ning. Routes can greatly affect the success of the
whole mission, whether the mission takes place on
ground or in the air. Some of our previous research
was focused on this aspect; we used simulation-based
planning to perform mission planning at the company
level of command (Lee and Fishwick 1994). Using
simulation-based planning in the military domain was
a natural extension of the already existing wargame
approach where the mission plans were tested off-
line via constructive simulations. Related work by
(Czigler, Downes-Martin, and Panagos 1994) demon-
strates the usefulness of simulation as a decision tool.
Robot route (i.e. path) planning is another appli-
cation area. If there is little uncertainty involved, as
is sometimes the case in many robot route planning
problems, the existing approaches such as potential
fields (Barraquand, Langlois, and Latombe 1992) do
quite well. But, when uncertainties exist in the envi-
ronment, these methods alone cannot produce good
results. The main algorithm of producing a graph of
traversable paths and searching the graph for a best
route is still the same, however. The part that is dif-
ferent is how we measure the goodness of a route. If
the goal is to select a route that is the shortest in
distance, we can use any of the standard algorithms

Figure 1: Mars Microrover. Provided by permission
of Jet Propulsion Laboratory (\I .I11I. Gat, Harri-
son, Wilcox, R., and Litwin 1 r'.)

that exist for finding shortest paths in a graph. But,
if the problem is in an environment that is unknown
or uncertain, we must use different ways to evaluate
each path. Our claim is that we can use simulation to
quantatively compute the outcome of a future action.
The way we do this is by using simulation models for
each entity or object in the world using data sampled
from a probability distribution, and performing mul-
tiple simulations to obtain an estimate of the state
In 1996, NASA plans to launch a spacecraft to Mars
to explore the environment of the planet (\I I II.. -
Gat, Harrison, Wilcox, R., and Litwin 1 '.'). The
spacecraft will carry an 11 kg rover, called the Mi-
crorover, that will move around the vicinity of the
landing site to explore the territory for a duration
of approximately 1 to 4 weeks. Figure 1 shows the
Microrover traveling over a rock.
Because the Martian surface is not completely
known, JPL is undergoing a process of performance
evaluation of the rover's autonomous navigation sys-
tem with varying terrain characteristics. The Micro-
rover testbed contains the Microrover vehicle and an
indoor test arena with overhead cameras for auto-
matic, real-time tracking of the true rover position
and heading. In the arena, they have created Mars
analog terrains by randomly distributing rocks ac-
cording to an exponential model of Mars rock size
and frequency created from Viking lander imagery.
Figure 2 shows a nominal Mars terrain which was
adopted from (\I Iil Gat, Harrison, Wilcox, R.,
and Litwin 1 1'.). The term nominal is explained in
more detail in Section 5. JPL has decomposed the

rover navigation task into four functions: 1) goal des-
ignation; 2) rover localization; 3) hazard detection;
and 4) path selection. Although these four functions
are integrated, for purposes of research, we will fo-
cus mainly on route selection. The first three func-
tions are largely connected with problems in sensors
(stereo camera pair, wheel encoders, solid state turn
rate sensor, light stripe ranging sensor, inclinometers)
and their uncertainties. Path selection uses the infor-
mation gathered from these three functions to nav-
igate to a goal destination. Currently the route se-
lection is achieved by a simple behavior control algo-
rithm which is reactive and does not take any excess
knowledge-such as maps-into account. Our approach
is to use simulation at this stage of route planning to
select the route to the goal. Once a more detailed
map is constructed of the vicinity, the planning will
involve multiple simulations of each virtual route to
the goal.

4.1 Simulation-Based Route Planning Sys-

Figure 4 illustrates the basic components of our
Simulation-Based Route Planning system. Initially,
the planner takes the goal location of the route as in-
put and selects a route plan for output. This selected
plan is the input to the Control Subsystem which per-
forms a supervisory control of the process. The out-
put of the rover process is the actual sensory output
of the rover. The sensory output will include camera
images, hazard detections and position information.
Along with the plan, a simulation log (the simulation
data that was produced previously during the plan
evaluation process) of the chosen plan is provided as
input to the Control Subsystem. This can be used to
serve as a reference model to track the state of the
execution in order to monitor its progress towards the
goal. The monitoring information can be used further
to tune the system towards the goal (i.e. correct its
route or position estimation) or to generate a failure
signal to the planner as soon as it decides that the
current route is unlikely to succeed.

4.1.1 Planner

The planner has three major modules:

1. The Route Generator uses a map of the vicin-
ity leading to the goal to extract rocks that are
large enough to be considered as obstacles. The
threshold size is determined by the size and abil-
ities of the rover. In the case of the Mars Micro-
rover, it has been determined to be rocks that
have diameters larger than 23 cm. In Figures 2


2 --



3 4 5 6 7 8 9

Figure 2: Map of Nominal Mars Terrain with Visible Routes


Figure 3: Map of Nominal Mars Terrain with Reduced Visible Routes

and 3, the rocks are represented by hexagons
drawn with dotted lines. There are three rocks
which are big enough to be considered as obsta-
cles. A bounding box is drawn around each of the
rocks. A visibility graph is produced connecting
the start location and the goal location via these
bounding boxes (Figure 2). The dotted lines in
the figure represent routes that will be eliminated
via the reduced visibility graph method. The visi-
bility graph is a common approach used to create
paths between obstacles (Latombe 1991) by con-
necting vertices of each obstacle to every other
obstacle, including the start and goal location.
The reduced visibility graph is a graph having
the minimal Euclidean route length between the
start and goal location and the obstacles. This
is what appears in Figure 3. The reason for
generating routes between obstacles is because
we use the obstacles as landmarks during navi-
gation. If a direct (as opposed to indirect which
means the route touches an obstacle) route exists
to the goal location, the reduced visibility graph
will produce it since this is the shortest route to
the goal.

2. The Simulator simulates each route of the pro-
duced graph and records the result. The sim-
ulation is based on the physical and empiri-
cal models of the rover and the terrain. The
physical model includes specific characteristics
such as translation (max 0.67 cm/sec) and rota-
tion speed, step climbing ability (max 19.5 cm),
and specification for the hazard detection sen-
sors (range of view is about 120 deg. with 30
cm max distance). We also incorporate empiri-
cal models for dead-reckoning error and hazard
detection error (failure rate of 1 in 1000 given
that hazard frequency is 1/100). For the terrain,
we use the Moore's model of rock size-frequency
distribution to create terrains of different rock
sizes and frequencies. More detail about this is
discussed in Section 5. Using these models of the
rover and the Martian surface, we perform mul-
tiple simulations of each route. There are several
ways to proceed in the simulation. To perform
in real-time, it's preferable that we proceed in a
breath-first manner; for example, in Figure 2, we
simulate routes to the first set of obstacles and
then to the second set and so on. This approach
is similar to the approach we have taken in our
previous work in mission planning (Lee and Fish-
wick 1994). To reduce the amount of compu-
tation, we use the A* search method if we can
build a heuristic function which can estimate the

cost of the remaining route. Another possibil-
ity is the branch and bound method used in the
area of Operations Research. Due to uncertain-
ties which exist in the models, simulations must
be performed multiple times using the available
stochastic information to reduce the variance of
the outcome variables.

3. The Plan Evaluator/Selector evaluates the
results of the simulations and selects a route for
execution. Currently, three elements are consid-
ered: 1) the outcome of plan failure or success;
2) the total time elapsed; and 3) final position of
the rover. These elements can be combined into
a single score but, because the objectives may be
different in different situations, we evaluate them
individually and then select a plan according to
a predetermined criterion.


In simulation, experimental design is a method of
choosing which configurations (parameter values) to
simulate so that the desired information can be ac-
quired with the least amount of simulating (Law
and Kelton 1991). In experimental design terminol-
ogy, the input parameters and structural assumptions
composing a model are called factors and the output
performance measures are called responses. Our ex-
perimental design approach to the SBP method is to
choose different rock distributions (both in terms of
size and frequency) and different routes as our factors,
and the simulation results according to the evalua-
tion function as our responses. Thus, we vary two
elements of simulation within the planning process:
1) the route; and 2) the terrain. If the planning ex-
periment is performed using the full factorial design
approach, the number of combinations to simulate
will be prohibitive. However, optimization techniques
such as response surfaces and metamodels can be used
to alleviate the problem. Currently, JPL is experi-
menting their Microrover by performing test runs on
Mars analog terrains which has been created artifi-
cially by randomly placing rocks according to a rock
size-frequency model developed by Moore (\1....... and
Jakosky 1989). Moore's model is based on data ob-
tained from images taken by Viking Lander 2. Since
a similar rock density is expected for the Microrover
experiment, the same model can be used. The orig-
inal Moore's model for rocks down to a diameter of
0.14 m is represented by N = 0.013D-2.66, where N
is the cumulative frequency of rocks per square meter
with diameters of D and larger. This model predicts
that about 18.8% of the landing site area is covered

Figure 4: Simulation-Based Route Planning System

by rocks. However, the model used by JPL so far in
creating the actual test terrains is based on the modal
value of the surface rock cover over the whole planet,
which is estimated to be at 6%. The terrain created
from this model is called nominal. A computer sim-
ulation is under development at JPL in order to test
terrains with rock frequencies ranging up to 19%.
We base our simulated terrains on this model in
creating simulated Mars analog terrains. There are
three major factors to consider in creating the terrain:
1) rock sizes; 2) rock frequencies; and 3) rock place-
ments. For each simulation, the ranges of rock sizes
and frequency are sampled from the Moore's model.
Once the rover is on Mars, the map that is generated
from photographs will be used to place rocks that
are large enough to be obstacles in their absolute lo-
cations. The remaining terrain in between the large
rocks can be estimated through the Moore's model
during simulation. We can use any additional infor-
mation that is available (distribution of rock sizes and
frequency) of the landing site to design a non-uniform
sampling distribution. Initially, the sampling distri-
butions are uniform. By using visual information of
the landing site, we can build rock distributions that
are similar to the actual terrain characteristics. The
placement of these smaller (less than 23cm in diam-
eter) rocks will be random. If time permits, multiple
simulations with different small rock placements will
be performed. On a higher level, the percentage of
rock cover can be varied to be between 6% to 19%.
Figure 5 illustrates the variable factors in the simu-
For the simulation, we use discrete time step sim-
ulation and the following algorithm:

While (Goal is not reached) do
Sample sensor data
Execute control action on rover

Update rover state variables
Update current clock time by AT
End While

Until the rover's state variables indicate it has
reached the goal location, the planner continues above
loop. Sampling sensor data involves sensing the rocks
placed and sized on the map according to the pre-
viously discussed method. Depending on the size
and location of a rock, the planner may also invoke
the hazard detection sensors-indicating that the rover
will have to maneuver around the obstacle. Smaller
rocks (with diameter less than 23 cm) are ignored
by the hazard detection sensor but is likely to cause
the dead-reckoning error to increase. This effect can
be captured by incorporating a dead-reckoning er-
ror model into the simulation. This is planned for
later experiments. The control algorithm of the rover
is very simple: if an obstacle is detected then the
rover turns in place until the hazard is no longer de-
tected. If there is no obstacle then it continues to
move forward while turning towards the goal loca-
tion. The control action is calculated to take place
for AT seconds and the state variables (location x,y)
are updated accordingly using control dynamics of
the rover.


We have shown how we are able to build a real-time
route planning system, the Simulation-Based Route
Planning System, by integrating ideas from different
but related fields of research.
There are several advantages to the method of
simulation-based planning:

1. By employing models to simulate and predict

route n


6% -- 19%

0.05m 0.23m

rock placement (random)

Figure 5: Sampling Range and Distribution of Simulation Factors

the outcome of the process, we are able to cap-
ture the effects more accurately and completely
(given that the models are built appropriately).

2. The use of simulation models allows us to use
standard simulation analysis methods in tuning
the simulation models to closely reflect actual

3. The ability to use distributed object-oriented
concept in planning without having to reason
about the combined effects of agent's actions or
changes in the environment within one central
node enables the evaluation of a plan as a nat-
ural result of simulating different models in the

4. The ability to I I I:" progress and performance
during execution allows finer tuning of the exe-
cution process.

5. We can easily extend the set of models to include
additional properties (especially properties that
may be difficult to create physically such as cre-
ating the atmosphere of Mars) in testing plans.

Some potential difficulties exist in using simulation
in the planning process. Defining appropriate models
may be difficult and time consuming. The simulation
process itself can be computationally intensive. How-
ever, we plan to overcome this problem by varying
the level of aggregation of our simulation models.


Once the implementation of the rover problem
domain in our Simulation-Based Route Planning

(SBRP) system is finished, we will experiment with
our method by building a response surface of the
problem and then using various ways to search for
the near optimal solutions. In the long term, we plan
to extend our SBRP system to the domain of Air
Force mission planning problems.


We would like to thank Science Applications Inter-
national Corporation (SAIC, Orlando, Florida) un-
der contract #4515164 for partial support of this re-
search. Continued partial support of this work will be
possible through contract #4514169-12 from Rome
Laboratory, Griffiss Air Force Base, New York. We
would also like to thank the JPL Mars Microrover
group for their permission to use Figure 1 and various
data appearing in (\1I ,1li -, Gat, Harrison, Wilcox,
R., and Litwin 1 1'.). Disclaimer: this work is not
funded by JPL. The authors have used Figure 1 and
other rover related data as a sample application of
the simulation-based planning procedure.


Antsaklis, P. J. and K. M. Passino. 1993. An Intro-
duction to Intelligent and Autonomous Control.
Kluwer Academic Publishers.
Barraquand, J., B. Langlois, and J. Latombe. 1992.
Numerical Potential Field Techniques for Robot
Path Planning. IEEE Transactions on Systems,
Man, and Cybernetics 22(2), 224-241.
Basar, T. 1995. Dynamic Noncooperative Game
Theory Academic Press.

Czigler, M., S. Downes-Martin, and D. Pana-
gos. 1994. Fast Futures Contingency Simula-
tion: A "What If' Tool for Exploring Alter-
native Plans. In Proceedings of the 1994 SCS
Simulation MultiConference, San Diego, CA.
Dean, T. L. and K. Kanazawa. 1 1 -. Persistence
and probablistic inference. Technical Report
CS-87-23, Department of Computer Science,
Brown University.
Dean, T. L. and M. P. Wellman. 1991. Planning
and Control. Morgan Kaufmann.
Fikes, R. E., P. E. Hart, and N. J. Nilsson. 1972a.
Learning and Executing Generalized Robot
Plans. Artificial Intelligence 3.
Fikes, R. E., P. E. Hart, and N. J. Nilsson. 1972b.
Some New Directions in Robot Problem Solv-
ing. In Machine Intelligence 7. Edinburgh Uni-
versity Press.
Fishwick, P. A. 1995. Simulation Model Design and
Execution: Building Digital Worlds. Prentice-
Hammond, K. 1989. Cased-based planning. In Per-
spectives in Aritificial Intelligence, Volume 1.
Academic Press.
Kanazawa, K. and T. L. Dean. 1989. A Model
for Projection and Action. In Proceedings of
IJCAI-89, Detroit, MI, pp. -. 999.
Latombe, J. 1991. Robot Motion Planning. Kluwer
Academic Publishers.
Law, A. M. and W. D. Kelton. 1991. Simulation
Modeling and Analysis. McGraw-Hill.
Lee, J. J. and P. A. Fishwick. 1994. Real-Time
Simulation-Based Planning for Computer Gen-
erated Force Simulation. Simulation, 299-315.
Luce, R. D. and H. Raiffa. 1 I.';. Games and De-
cisions. John Wiley and Sons. Later re-printed
by Dover, 1 I-
Matthies, L., E. Gat, R. Harrison, B. Wilcox,
V. R., and T. Litwin. 1995. Mars microrover
navigation: Performance evaluation and en-
hancement. Technical report, Jet Propulsion
Laboratory, Pasadena, CA. To appear in Au-
tonomous Robots Journal.
Moore, H. J. and B. M. Jakosky. 1989. Viking
Landing Sites, Remote-Sensing Observations,
and Physical Properties of Martian Surface Ma-
terials. International Journal of Solar System
Studies, 164-184.
Russell, S. J. and P. Norvig. 1995. Artificial Intel-
ligence A Modern Approach. Prentice-Hall.
Schoppers, M. 1 1 7. Universal Plans for Reactive
Robots in Unpredictable Domains. In Int. Joint
Conference on Artificial Intelligence.

West, D., L. Mellon, J. Ramsey, J. Cleary, and
J. Hofmann. 1995. Infrastructure for Rapid Ex-
ecution of Strike-Planning Systems. In Proceed-
ings of the 1995 Winter Simulation Conference,
Washington, D.C.


JIN JOO LEE received the B.S. degree in Com-
puter Science from Ewha Womans University, Ko-
rea in 1988 and the M.S. degree in Computer Sci-
ence from Brown University in 1991. After receiving
the M.S. degree, she was a research engineer at Hu-
man Computers Inc., Korea until 1992. She is cur-
rently a PhD student in the Computer and Informa-
tion Science and Engineering department at the Uni-
versity of Florida. Her research interests are in AI
planning, simulation and control. Lee's homepage is

PAUL A. FISHWICK is an Associate Professor
in the Department of Computer and Information Sci-
ence and Engineering at the University of Florida.
He received the BS in Mathematics from the Penn-
sylvania State University, MS in Applied Science from
the College of William and Mary, and PhD in Com-
puter and Information Science from the University of
Pennsylvania in 1986. He also has six years of indus-
trial/government production and research experience
working at Newport News Shipbuilding and Dry Dock
Co. (doing CAD/CAM parts definition research)
and at NASA Langley Research Center (studying en-
gineering data base models for structural engineer-
ing). His research interests are in computer simula-
tion modeling and analysis methods for complex sys-
tems. He is a senior member of the IEEE and the
Society for Computer Simulation. He is also a mem-
ber of the IEEE Society for Systems, Man and Cy-
bernetics, AC\I and AAAI. Dr. Fishwick was chair-
man of the IEEE Computer Society technical com-
mittee on simulation (TCSIM) for two years (1988-
1990) and he is on the editorial boards of several
journals including the AC\I Transactions on Model-
ing and Computer Simulation, IEEE Transactions on
Systems, Man and Cybernetics, The Transactions of
the Society for Computer Simulation, International
Journal of Computer Simulation, and the Journal
of Systems Engineering. Fishwick's home page is

University of Florida Home Page
© 2004 - 2010 University of Florida George A. Smathers Libraries.
All rights reserved.

Acceptable Use, Copyright, and Disclaimer Statement
Last updated October 10, 2010 - - mvs