PAGE 1
stationary Analysis and Optimality Ccffiditions for (a,S) Policies in MultiCanTodity Inventory Control Prcblams By NABIL S. FADUR A DISSERTATION PRESENTED TO THE GRADUATE COUNCIL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFZLEJyiENT OF THE REQUIREME^7^S PDR THE DEGREE OF DOCTOR OF PHILDSOPHY UNIVERSITY OF FLORIDA 1972
PAGE 2
UNIVERSITY OF FLORIDA 3 1262 08552 8643
PAGE 3
ACKNOWLEIXIEMM'S I would like to ejq^ress my thanks to my advisor. Professor B. D. Sivazlian, for his advice and guidance. I am also grateful to Professors J. F. Bums, R. G. Blake, and Z. R. PopStojanovic for their assistance, encouragement and support. Thanks goes also to Mr. Philip L. Raphals for his fast and excellent typing. Finally, to my wife, goes special thanks for her unending patience and encouragarent. This research was partially st^jported by Project THEMES, ARDD Contract nAHCX)468 CX)002. 11
PAGE 4
TABi:Â£ OF CXKTENTS Page ACKNOWLEDGEMENTS ii ABSTRACT V Chapter I. INTRCÂ»UCnCN .' 1 1 . 1 Introduction 1 1 . 2 Literature Revisff 6 1.3 Definition of the Problan and an Outline of the C3iapters 20 II. MAIHEMATICAL PRELIMINARIES 26 2 . 1 Introduction 26 2 . 2 The Generalized Convolution 26 2.3 On the Solution of an Integral Equation 36 2 . 4 Other Mathematical Concepts 46 III . MATHEMATICAL FORMULATION 48 3.1 IntroductiCTi 48 3. 2 Mathematical Formulation 49 IV. THE OPTIMIZATIOJ PRDBLEiyi 71 4 . 1 Introduction 71 4 . 2 Characterization of T 72 4.3 Necessary and Sufficient Conditions 76 4.4 Geanetric Reforniolation of the Problem 98 4 . 5 The Functicxi L (x) 100 V. OCMPUTATIOSIAL ASPECTS OF THE PROBIÂ£M AN EXAMPLE 103 5.1 Introduction 103 5.2 Exaitple 104 5 . 3 Ccnputational Aspects 105 5.4 Numerical Exarrple 116 111
PAGE 5
Chapter Page VI. CONCLUSIONS AND REOOMMEMDATIOSIS 128 6 . 1 Conclusions 128 6 . 2 Extension and Recxxmendations 131 GLOSSARY OF NOTATION 134 APPENDIX 5A COMPUTED OUTPUT RESULTS AND PROGRAM LIST 136 BIBLIOGRAPHY 146 BIOGRAPHICAL SKETCH 149 IV
PAGE 6
AbstracÂ± of Dissertation Submitted to the Graduate Council of the University of Florida in Partial Fulfillment of the Requirotents for the Degree of Doctor of Philosophy STATIONARY ANALYSIS AND OPTIMALTTY OONDITICNS FOR (o,S) POLICIES IN MULTIHXM^DDITY INVHWORY OCWTROL PRDBLEJ^ By Nabil S. Faour March, 1972 Chairman: Professor B. D. Sivazlian Major Department: Indxastrial and Systems Engineering An moonmodity (m >^ 1) inventory control model with periodic review is formulated and studied when a dyadic stationary ordering (a,S) policy in e"' is used. It is assumed that: (1) the ordering decision at the beginning of each of a sequence of periods of time is affected by a single set v^ cost, a linear variable ordering cost and holding and shortage costs; (2) the demands for the items in each period are identically and independently distributed continuous random variables; (3) the delivery of orders is immediate; and (4) catplete badklogging of unfilled demands is allowed. Next the model is analyzed by minimizing the e^qjression for the stationary total expected cost per period. The set of simultaneous equations used to determine the optimal policy parameters are restated in terms of a real valued function in E . Finally, the optimization analysis is restricted to the special case of a twoocmtodity inventory oontrol problem v^ere the danands for the items obey the expcxaential distribution, and the holding and shortage costs are linear. The integral equation vdiich is used in solving for the optimal
PAGE 7
policy parameters is cxinveirted into a linear hyperbolic partial differential equation of the second order with boundary conditions. This boundary value prcblan is solved analytically. The solution is then used to determine the optimal policy parameters. In a numerical exanple a corputer program is developed to determine the cptimal policy parameters. VI
PAGE 8
CHAPTER I INTRODUCTION 1.1 Introduction Ihe replenishment and control of inventories is a caution practice in the fields of business, eocnanics and management in general. Sinply, an inventory is a stock of physical goods v^ch are held or stored for future use, sale or production. Inventory problems are integrated in nature. Ihey may involve production, scheduling, distribution of catmodities, or oarbination of all. In the past twenty years, much work has been done to formulate and apply the proper mathanatical tools for determining economic decision rules in managing the varioxas inventory systems . Viewed as decision problems , the analysis of such systems essentially proceeds in three steps: (1) the formulation of a model expressing a set of relations between a set of decision variables whose values are to be determined; (2) the establishment of an objective function of the decision variables so as to minimize or maximize an eoononic objective; and (3) a numerical method to determine the values of the decision variables that optimize the objective function. In this introduction we shall restrict ourselves to models in v^ch decisions are made at the beginning of each of a nunber of eqiaal periods of time, and in vArLch the demands are indqpendently and identically distributed random variables from period to period. The decision at the beginning of each period is affected by a fixed set up cost, a
PAGE 9
variable ordering cxDst and a holding and shoirtage costs. Backlogging and irtmediate delivery of orders are allcv/ed. IWo principal distinct approaches have Been used in analyzing such inventory models both in theory and practice. In the first aj^roach, the syston is viewed as a multistage decision process and the technique of dynamic progranming is enployed in finding the optimal policy that minimizes the total expected cost over the duration of the process. To ensiire that the optimal policy is of the "sinple type" certain restrictions are inposed on the cost factors that affect the decision rules. More precisely, when the ordering cost is linear and v^en the sum of holding and shortage costs L(x) is convex, the optimal policy is of the "sinple type." However, vrtien the duration of the process is infinite, a second approach is often used: A "sinple type" ordering policy is chosen to be used in each period. Under the chosen sinple policy the sequence of inventory levels at the beginning of each period forms a Markov process, whose stationary behavior can be analyzed. Here, instead of optimizing the total cost over the horizon period, \4iich is meaningless, we minimize the stationary total expected cost per period . The optimal values of the decision variables are called the steadystate or stationairy solutions to the inventory problem. Sane of the advantages of the stationary approach over the dynamic progranttiing ^proach are: 1. The stationary ^proach provides us with information about the dependency of the optimal policies on the many parameters involved in the model and about sensitivity of costs as function of the policies, while this information is not provided by following the dynamic progranming approach.
PAGE 10
3 2. As the notiber of ocnttodities increases, the nuirber of the state variables increases drastically, and the use of dynamic progranining techniques in obtaining cptimal policies for the inventory problem will be timsconsutning even vdien using the fastest ccnputers. 3. Frequently, for lewcost high turnover items, optimal policies are not really required. It is sufficient to obtain sinple analytical approximations to optimal policies, and these are <Â±)tained from the stationary ^jproach. The main ooncem of this research will be to find the stationary solutions to the mocmnodity (m >^ 1) inventory prcÂ±)lem. For the one ocmtodity, the theory of inventory has studied the statiaiary behavior of the system and provided cptimal decision rules for inventory replenishment under a linear ordering cost and a convex holding and shortage costs function. The results of the oneocnmodity inventory system are valid for multiocmnodity inventory system where each comnnodity is treated independently. However, in practice this independency assarption is unocrrtDn and it is more realistic to incorporate the interdependency in the set up order cost of the oortmodities \4iich arises in the nature of the following types of problems. 1. Ordering different ccimodities at the same time from a oamcai vendor may reduce the procurement ordering costs, thus incurring one ordering cost. 2. Several warehouses used as storage and distribution points are si;53plied by a single factory. Whenever an order to replenish a given commodity stored by the warehouses is
PAGE 11
made to the factory, a fixed set up oost is incurred whether the warehouses order individually or simultaneously. The configuration of this systan is shown in Fig. (1.1) . The two problans are similar in structure and it is practical to take advantage of the reduced procurement cost in controlling the inventory of the cottnodities . Furthermore, if one can solve the multiocmnodity prcijlem, then the multiwarehouse problem, as given previously, is solved. The procurement cost of ordering Â£ >^ 0^ units is assumed to be equal to a set ip cost plus a linear purchase cost T C Z = (Cj, Co,..., c ) where c^ >_ 0, i = 1,.. . , m, is the unit purchasing cost of item i. The fixed set up oost is incurred only if an order is made. Thus if z > units are ordered, the fixed cost will be a function of z, say K(z) In general K(z) will take as many different values as the nunber of alternative different ordering decisions. For the two conmodity prcblen K(z) = Ki K2 K if Zi > and Z2 = if zi = and Z2 > if Zj > and Z2 > if Zi = and Z2 = v*iere Ki, K2 and K are all nonnegative, and the inequality max (Kj , K2) 1 K < Ki + K2 is satisfied. Thus , the procurement cost is given by K{z) + c'^z
PAGE 12
Â•^ Pm
PAGE 13
1.2 Literature Review Vfe shall first survey the literature on the steactystate or stationary solutions to the onecamodity inventory system operating \3nder a (s,S) policy; next we shall elaborate on the work done on the stationary solutions to the mcomnodity inventory system operating under a (o,S) policy. For the onecoimodity problem we shall consider a c^amic inventory nvodel. At the beginning of each period, a decision is made to order up to S if the stock falls below s, otherwise, nothing is ordered. The denands, {D.}, in successive periods (i = 1, 2, . . .) are assured to be described by a sequence of nonnegative random variables, independently and identically distributed with a joint density function '^'ri ^ * ^ Â• Delivery of orders is inmediate and total backlogging of unfilled dotiands is assumed. The ordering decision in each period is affected by a set vp cost K, a linear purchase cost cz where z is the quantity ordered and c the unit cost, a holding cost for carrying inventory, and a shortage cost for not meeting demands. L(x) represents the expected inventory holding and shortage oosts for being in stock X at the beginning of the period. Pollowing the pioneering work of Anrcw, Harris, Marchak [1], Karlin [2] studies the steady state for this model operating under a (s,S) policy. With this ordering policy, the sequence of inventory levels Xi , X2 , . . . , X , . . . , v^ere X. is the inventory level at the end of the ith period, forms a Markov process. Karlin shews that the distribution of X^ coverages in the sense of distributions to a stationary distribution with the density f {Â•) given by
PAGE 14
f(x) = f 'P (S X) 1 + 4'(S s) Ss (S s) + I i;(t) ^(S X t) dt for s < X < S (1.1) 1 + 4'(s s) for <Â» < X < s vAiere i;{x) = f (>^"Nx) n=l and 'i'(x) = j; D^^Nx) n=l 4) ( * ) and $ ( Â• ) are the wellknown syntxDlic notations for the n^ fold convolution of (()(Â•) and $(Â•), respectively. Itie stationary distribution is clearly independent of any econoTdc consideration. By inposing the assumed cost structure on the process the total expected cost per period denoted by g(s, Ss) is given as K + L(S) + L(S x) i/)(x) dx g(s, s s) = TTim "" "^ r J n (1.2) vdiere Q = S s and \i = E(D) / E(*) is the expectation operator.
PAGE 15
Greenberg [5] obtains f (Â•) , the transient soluticn for the (s,S) inventory system. Using the generating function approach, the stationary distributic^ f ( Â• ) is given by the sta n da r d Abelian theorem lim " f (x) = (1 z) I z"f^(x) z^l n=l v^ch results in (1.1) . Iglehart [7] shows that the optinal values of s* and Q* minimizing g(s, S s) are given as the solution to the following equations L'(Q + s*) + L ' (Q + s x) ij; (x) dx = K + L(Q* + s*) + L(s*) = L(Q* + s x) \l){x) dx 1 + "PCQ*) where L ' ( Â• ) is the derivative of L ( Â• ) Â• Another approach in studying the (s,S) inventory model is follcwed by Sivazlian [11] . He defines f (x) to be the total expected cost of operating the system for n periods when no order is placed initially, i.e. , s <_ x < S, h (x) to be the total expected cost of operating the system for n periods v^en an order is placed initially, i.e., Â°Â° < X < s. Then f (x) and h (x) satisfy the following functional equations
PAGE 16
'X~S f (x) = L(x) + 1 f .(x t) 4)(t) dt h (x t) 4) (t) dt s < X < S (n > 1) xs h (X) = K + c(S x) + L(S) + n rSS f T (S t) ^ (t) dt ni + 1 h , (S t) (J> (t) dt Â» < X < s (n >^ 1) J Ss "~ Starting with these basic functicaial equations and using Howard's ^proadi [6] he deduces the stationary total ei^ected cost per period as given by (1.2) . The optimal values for s and Q = S* s* that minimize the total ejqaected cost per period are given as the soluticais to the set of equations ,* /.* M(s*,Q ) = K and 9M(s*,x) = x=<3" vihere ,rÂ„/ * VI _ L{L(s*) L(s* + x)} L{M(s ,x)} 1 lU(x)} the operator I. { } is the Laplace transform operator.
PAGE 17
10 Several corputational itethods for cÂ±)tairujig values of s and Q* are given and elaborated later by Sivazlian.^ For the case of gartma distributed demand and linear holding and shortage costs, a dimensional analysis is carried to reduce the number of variables. A numerical inversion technique of the Laplace transform e:^ressians using Gaussian quadrature is used to solve for s* and Q* . For large values of Q, Sivazlian [11] and Roberts [3] shew that s* is the solution to the equation ^^ [1 *(v)] dv = hTT^ I ^' ^ ^ ^ ^^^^1 and Q* /Â„ y , where y and a^ are the mean and the variance of the randan variable D, h is the unit holding cost, and p the unit shortage cost. Sivazlian [11] also shows that for small Q,s* and Q* are the solutions for the following approximate system of equations L'(S*)=P L"(s*) =^+p4,(0) It is inportant to note that an integral equaticn of the renewal type in one dimension is solved by each of the authors in carrying their analysis. In Scarf's work [2] , the density function f (Â•) as given by (1.1) is a solution of a renewal type integral ^See additional references.
PAGE 18
11 equation. Sivazlian [11] in finding the total expecÂ±ed cost per period solves a renewal type integreil equation. For the case v*ien the desriand is discretely distrib\ited, Veinott and Wagner [16] develop the nvethod of stationary or renewal analysis for the ccrputation of an optimal (s,S) policy. The total es^jected cost per period is given by Ss K + L(S) + I MS j) i)(j) g(s, S s) = Â— + c^ ^ 1 + H'(S s) which is a discrete version of (1.2) . Ihe algorithm provided for searching the (s,S) policies to find one that minimizes g(s,S s) consists of two steps. Step 1. Determine lower bounds (s,S) and upper bounds (s,S) on s and S . St^ 2. Find the collection of all (s,S) policies that minimizes g(s,S s) over the class of (s,S) policies falling within the bounds defined in Step 1. Any policy in the oollection is optimal. An interval bisection or a Fibonacci search technique is then eitployed if A ju JL to find the optimal values of s and Q = S s . Note that this procedure does not guarantee the optimality of g{s,S s) since, in general, g(s,S s) is not a unimodal function in s and Q.
PAGE 19
12 RDr other work done in the area of oneocninodity inventory, the survey articles of Scarf [9] and Veinott [18] should be noted. Let us now t\xcn to the mocrmodity (m >^ 1) prctolem. As itost practical problems in inventory involve more than one product, there has been substantial interest recently in studying the theory of multicamodity inventory problons. The prcblem we shall consider is the familiar dynamic inventory system with periodic review [17] . Danands, {D.}, for the items over a sequence of successive periods (i = 1,2,...) are assumed to be described by a sequence of nonnegative random variables, independently and identically distributed. No specific restrictions will be placed \;?x3n the statistical properties of the danands for corfttxiLties within a given period; thus, in general, the donand distribution is given by a joint density function j^it) (t > 0) . Ocnplete backlogging and iimiediate delivery of orders are assured, and as previously discussed the ordering decision in each period is affected by a single set up cost, a linear ordering cost and holding and shortage costs. Following Sivazlian [13], we define the following: In if", let n = {x3Â« Â£S} v*iere x = (x^ ,X2,. . . ,xj is the inventory levels of all itars prior to a decision and S = (Si ,82 , . . . ,S^) is a nonnegative vector. Let the hypersurface that subdivides the set Q into the subsets a and a^ be defined inplicitly by an equation of m variables, Z( S xÂ° ) = 0, xÂ°efi, v^ere the realvalued function Z( S x) has the following properties:
PAGE 20
13 1. Z( S x) is defined and ocaitinuoxos for V xefl 2. Z( S x) < V xeoÂ° 3. Z( S x) > V xea Define r = {xjxe^, Z( S x ) = 0} and u)(x) = {tt Â£ x; t, x eo^]; The stationary policy considered is to order all itans for xea . For the case of two ccmnodities , Fig. (1.2) illustrates gecmetrically the decision regions. It is inportant to note that for x^n, the (a,S) policy is not feasible. However, once a stock level reaches a level xefi, the policy is and remains feasible. With this ordering policy, the sequence of inventory levels Xi ,X2 , . . . ,X , . . . , vdiere X. is the inventory level of all items at the beginning of the i^h period, forms a Markov process since the demands are independently distributed frcxn one period of time to the next and since the (a,S) policy depends only on the last state of the process. Sivazlian [10] and [13] studies the stationary behavior of the systan for the special case v^en T is admissible, i.e., rnA(x)= (})(null set) where A(x) = {tx < t < S , xca^} . If W(^) denotes the distribution function of the stationary distribution of of the stock levels imttediately after a decision, then
PAGE 21
14 6M(y) = M if Y = S; \l) {y) *^ if ^ edc v^ere 8<^ = a^ S, M is the probability of placing an order, and \1){Y) satisfies the integral equation ii (y) = M4> ( S y ) + 4) (x) (j) ( x y ) dx
PAGE 22
15 o^ = aj(S) Fig. (1.2). (a,S) ordering policy for a to^cxjimodity inventory system.
PAGE 23
16 Sivazlian [10] shows that the solutican of the integral equation is given by (n) 'J^{y) =M I <})^"' ( S y) n=l v*iere (x) with itself and M = 1 + li^(t) dt Steu (S) The optimal decision variables Si, S2 , . . . ,S and the oonfigioration of the admissible set r are derived by minimizing the total ej^sected cost per period given as m g(a,S) = E(C) = KM + I [c.E(D.) +h.E(Y.) + (h. +p.) E(B.)] (1.3) where C is the total cost per period; and for items i = l,2,...,m c. = variable cost/unit h. = holding cost/ijnit at the end of a period p. = shortage oost/unit at the end of a period B. = unfilled demand at the end of a period Y^ = inventory level at the end of a period
PAGE 24
17 E(*) denotes the expectation operator. For the twooonmodity problem v^en the demands for the items are independently and exponentially distributed with (f* j^ (t . ) = X . e~ """ "^ <_ t. < Â», X^ > (i = 1,2) , the optimal policy is investigated for the twD given classes of adnissible T Case 1. Here r is the set of points for v^di Xj (Sj xi) + X2(S2 X2) = a vy^iere a is an unknown positive parameter to be determined. Case 2. Here T is the set of points for whioh itax [Xj (Sj Xj) bj , X2 {S2 X2) b2] = v^ere bj and b2 are unknown positive parameters to be determined. For both cases the optimal decision variables are ootputed. Moreover Sivazlian shews by numerical exaitples that the joint (a,S) policy yields a lower operating cost than the case when eaoh item is ordered separately. In another paper Sivazlian [12] considers the multiooninodity control problem operating vinder a stationary (a,S) policy. The joint density function of demand for all items in a given period is given by *(t) = n 4). (t.) < t. < 0Â°, (i = l,2,...m) i=l 1 1 1
PAGE 25
18 viiere 4) (t.) = X.e ^ ^ (i = 1,2, . . . ,in) . T is admissible and here it i 1 1 is the set of points for vghich Xi(Si Xi) + X2(S2 X2) + ... + X^(S^ X J = a a > In this case, let A = lij(t) dt 'stÂ£a)(S) By generalizing the concept of Dirichlet's multiple integrals, Sivazlian shews that A can be ejq>ressed as a single integral . Next/ similar ejqsressions for E(Y.) and E(B.) are obtained. For the case v^en the variable a takes on large values, explicit analytic solutions for the optimal values of S and a are obtained by minimizing (1.3) . It is also shewn that when m, the nvjÂ±>er of itans, beccres very large and a is finite, the prcÂ±iability of placing an order in any given period tends to 1, and that it is almost optimal to use the "order 14) to" policy. For the twDocniTodity problan, Wheeler [20] follows Iglehart's [7] approach for the onecannodity (s,S) model in showing that , . C (x) g(a ,S ) = n>oÂ° n v^ere C (x) is the total expected cost of operating the system for n periods, given that x is the initial inventory level and an optimal policy is used in each period, and g(a*,S ) is the optimal stationary total e3q)ected cost per period. Pollcwing Greenberg's [5] notations
PAGE 26
19 and procedure, Wheeler also atta:npts to find the transient solutions for the probability distribution of the stock levels. Wacker^ formulates and stiidies the multicomiodity inventory problem with periodic review v^en a stationary mdimensional (s,S) policy is followed. Following the work of Roberts [3] he gives analytical formulas for determining the asyrptotic reordering region v^en the set up and shortage costs have large values. For the case when the demand D in each period is discretely distributed with probability mass function <\>^P^it) , vrfiere x is the stock level at the beginning of the period, and (J'^^^i/O) < 1/ Johnson [8] uses the policy iteration method to choose the optimal (a,S) reorder policy which minimizes the stationary total expected cost per period. Further he provides a ccnputaticnal algorithm for finding bounds and characterizing the optimal policy. The algorithm is a search procedure and only attenpts to find the ocaif iguraticn of the optimal policy pointwise. Finally, the work of Soland [15] is briefly mentioned. In solving some renewal processes in twodimensions, Soland solves a oontinxious review inventory problem involving two products for the case v^en the demand is discretely distributed. In minimizing the total expected cost per cycle (defined as the time between ordering) he concludes that the optimal policy is to order both products at the end of a cycle. Here again, as in the oneocmnodity prcblan, a renewal type integral equation in mdimensions is solved by each of the authors in carrying their analysis. ^See acMitional references.
PAGE 27
20 A more ejqslicit ctefinition of the problem is given next. 1.3 Definition of the Problem and an Outline of the Ch^ters A multioamiodil^ inventory system with periodic review operating lander a stationary (a,S) policy is considered. The ordering decision in each period is affected by a single set up cost K, a linear variable ordering cost c = (cj , Cj , . . . ,c ) , and the expected holding and shortage cost function, L(x) , conditional on being in stock level x = (xi , X2,...,xJ at the beginning of a period; L{x) is assuned to be twice dif f erentiable . Danand, {D.}, for the items over a sequence of periods (i = 1,2,...) , is assumed to be independently and identically distributed oontinuous nonnegative randan variables with continuous joint density function 4'p,(t) . Delivery of orders is irmediate and total backlogging of unfilled donands is assumed. Under a stationary (a,S) policy, as discussed before, either all items will be ordered to bring the inventory level to S = (Si, Sz, . . . ,S^) if X, the inventory level at the beginning of a period prior to making a decision, is in a (ordering region) or nothing is ordered c if xeo (not ordering region) . The principal concern of this study is to find the optiniality condition for (a,S) policies. This is done by minimizing the ej^ression for the stationary total expected cost per period with respect to the decision variables that characterize the policy being used. The study begins in Chapter II by introducing seme mathematical concepts that are needed in subsequent chapters. First the concept and properties of a generalized convolution operator defined on locally
PAGE 28
21 integrable realval\aed functions of in (m > 1) variables are introduced; next the solution of an integral equaticsi of the renewal type in m dimensions is studied. In Ch^ter III the analysis for deriving an analytical ejqjressicn for the stationary total ejqaected cost per period, proceeds in three steps: (1) the basic functional equations which relate f (x) , the total es^jected cost of operating the system for n periods v^en starting with xea , and h (x) , the total expected oost of operating the system for n periods vrfien starting with xea, are derived; (2) for large n it is shown that f (x) = ng + u(x) and h (x) = ng + v(x) ; \diere g is the nÂ— Â— nÂ— "^ stationary total ejqaected cost per period and u(x) , v(x) are f\jnctions of the initial stock levels; and (3) an ancilytical ej^ression for g, which is a function of (a,S) , is derived by solving a renewal type integrail equation that results frcm the use of the asyrnptotic ejqaressions for f (x) and h (x) in the fvmctional equations. In Chapter IV, the optimality conditions for (a,S) policies are studied. The minimization of the ejqnression for the stationary total expected cost per period with respect to the decision variables proceeds as follows : First, it is shewn that at optimality r = {xxeJ^; L(x) C* = 0}, viiere C* is the minimum value of g(a,S) excluding the variable ordering oost. Then the necessary and sufficient conditions for the existence of prcper relative minima of g(a,S) at the pair (S*,C ) are established. In Chapter V, the stucfy considers the twooomnodity inventory control problem v*iere the demands for each ocmnodity, in a given period.
PAGE 29
22 are independent and exponentially distributed, and vAiere the holding and shortage costs for each coimodity are linear. The equations used to solve for the optimal policy parameters are analytically determined. In a numerical exarrple a ootputer program is developed to determine Sj*, S2* and C*. Finally, in Chapter VI recamendations for future research are presented. As carpared to other works, in the mocmnodity (m > 1) inventory control problem c^jerating under a stationary (a,S) policy, this research is more general, in the sense that no assurtptions are made about the configuration of the ordering region or the admissibility of r. The caaly restriction on the set r = {xxefi, Z ( S x) = 0} that subdivides the set fi into the ordering region, a, and the not ordering region, a , is that the real valued function Z( S x) is continuous and defined for all xef2. jybreover the ordering regions must be sinplyconnected. The ordering regions considered by other researchers, in the moonmodity problem for oontinuoios demand, are all special cases of the general case considered here. In E^, Fig. (1.3) shows geonetrically the configuration of sane of the regions that are considered in this research. Fig. (1.4) shows some ordering regions that will not be considered. The mdimensional ordinary convolution operation used by other researchers in studying the system ixnder consideration is a special case of the generalized convolution operation introduced in Ch^ter II. For the case when the set r is admissible, the generalized cxjnvolution of any fxmction reduces to the ordinary convolution and
PAGE 30
23 SzJfS Sa^ Si (Fig. (1.3)
PAGE 31
24 Fig. (1.4).
PAGE 32
25 all the generalized oonvolxition operation properties with the ccmnutative law hold. The analysis \ased in deriving the e3q)ression for the stationary total expected cxDst per period is similar to Sivazlian's [11] approach for the oneocxtinodity problem when operating \jnder a (s,S) policy. The asynptotic ej^ressions for the costs of operating the system are deduced by ^:pealing to the results of Yosida and Kakutan [22], The set of equations that will be used to determine the c^timal parameters of the adopted policy are similar in form to the set of equations used by Sivazlian [11] in detentdning the optimal values of s* and Q* = S* s* for the one ocrmodity problem.
PAGE 33
CHAPTER II MATHEMATICAL PRELIMINARIES 2.1 Introduction In this chapter a number of mathematical concepts that will be needed in subsequent chapters are discvissed. First, the concept and properties of a generalized convolution operator defined cai locally integrable realvalued functions in m(m >^ 1) variables are studied and then the solution of a renewal type integral equation is given. 2.2 The Generalized Convolution Let S be a positive vector in a space of m (m >^ 1) dimensions ; i.e., S > Â£ <=> Sj^ > 0, i = l,2,...,m. In e'", define the set U = {x I X j^ S} / and let the hypersurf ace that subdivides the set fi into the sxÂ±)sets a and a^ be defined inplicitly by an equation of m variables ; Z ( SxÂ° ) = , x" efi , where the realvalued function Z ( Sx) have the following properties: 1. Z ( Sx) is defined and continuous for Vxefi; 2. Z (Sx) < Vxea^ 3. Z (Sx) > Vxea Next, define the sets: (a) (ij(S) = {x  xefi; Z ( Sx) < 0}; (b) u (x) = { t I t, xeo) (s_) ; t Â£ x} ; (c) r = {x I xe^; Z(S2X) = 0}; (d) 0) ' (t ,x) = {v I t
PAGE 34
27 (e) r"(x) = {t I t <_x; J^wCS)}; (f) r'{x) = r"(x) a)(x) I^t us make the transformation u = xt , tew (x) , then the set R(x) is the image of the set w(x) , r^ is the image of T , and Rj ( xt) is the image of w' (t,x) . In the particular case ra = 2, Fig. (2.1) illustrates geometrically the effects of the transforTtBtion on the above defined sets. For v^R(x) , define the set Rj (u) = (t  1 t <^ u; teR(x)}. Consider the set of functions D, f (t)eD is a real valued function of ^(2.1) m variables (t, ,t, , . . . ,t ) for which we have * ^ m 1. f (t) = f or t < 2. f (t) is locally integrable on Rj (u) On this set of functions, addition and gconvolution operations are defined as follows: rvaf 1. Addition (f 4g) (t) = f (t) +g(t); Def f 2. gcJonvolution (f*g) (u) = I f ( uv) g(Y_) dy verTT^) then the following relations are satisfied 1. f(t), g(t)eD => f(t) + g(t)eD 2. f (t) + g(t) = g(t) + f (t) (Connutative Law) 3. {f(t) + g(t)} + h(t) = f(t) + {g{t) + h(t)} (Associative Law) where the equality sign is used in the ordinary sense. And the set of functions D is an abelian group vdth respect to addition in v^ch the zero element is the functicm f (t) =0 and the inverse of the functiCTi f (t) is f (t) .
PAGE 35
28 y
PAGE 36
29 Properties of gConvolution 1. Closure Law If f (t), g(t)eD, then h(u), given by h(u) = f f(u2t) g(t) dt ^teRi (u) belongs to D. Proof ; h(u) is locally integrable on RjCu) since both functions f and g are locally integrable on R , (u) . Moreover h(u) is identically zero for u < . 2. Distributive Law If f , g, h, e D then [f * (g+h)] (u) = (f * g) (u) + (f * h) (u) Proof : By definition [f * (g + h)] (u) = f f(u^ [qiY) + h{^)] 6^ = f f(u2^) g(jr) d^ J^^eRi (u) +f f(ur^) h{Y) ^ ' ^gRi{u) = (f * g) (u) + (f * h) (u) 3. Associative Law If f , g, e D then [f * (g * h)] (u) = [ (f * g) * h] (u) Proof ; By definition [f * (g * h)] (u) = [ f (uv) I g(vy) hiY) djr_ dv Â•'veRj (u) ^Y.^Ri (v) = f [f(uv) g(vy) h(Y) d^. dv (2.2)
PAGE 37
30 v^ere W = ( {v_,y)  veR (u) , ^eR (v) } vMch is equivalent to the set W. defined as 2 Wj = { iy.'Y) ! Y. <. Y. 1 y.' Z' Ye^i ^H.^ ) Also by definition we can write [{f * g) * h] (u) = I [ f f( uyt) g(t) dt] TMy) ^ (2.3) Let t = V Y^, then (2.3) becones [(f * g) * h] (u) = I [ I f(u2v) g(v2V) dv] h(y) d^ VeRj (u) vyeR^ ( uy) Since g ( vy ) = for vy < Â£, vye R^ ( uy) implies that yt^^ (Z'H.^ ' where r' (jlÂ»u) ={vl Y.l_vl_UrJ!ERi (u)) Â• Hence we have [(f * g) * h] (u) = j r I f (u^v) g(v2r^) dv] h(j^) djr ^Rj (u) \cR' (y,u) = 11 f^Hli:) g
PAGE 38
31
PAGE 39
32 4. CormTutative Law ; We observe that in general, if f , geD, then (f*g) (u) ^ (g*f ) (u) By definition (f*g)(u) = f f (u t) g(t) dt teRi (u) Let u t = v; then (f*g) (u) = I f (V) g(u:^) dv ?^ {g*f) (u) veRj (u) viiere Ri (u) = {v uV Â£Ri (u) }. Note : the goonvolution of feD with itself ccmnutes. Moreover , if we define for n = 1,2, f(l)(u) = f(u) f (n) (E) = (f (nl)*f) = j f (n1) ^Hlt) f (t) dt (n ^ 2) teRi (u) then f ran the goonvolution associativity law we can write f(n)(ll) = j f(nl)(y=^) f(t) dt = j f(u^) f(^_i)(t) dt teRj (u) teRi (u) (n >_ 2) If we let ut = v; then Ri (u) is mapped into R^ (u) arxJ we can write f(n)(ii) = j f(nl)
PAGE 40
33 (Z.X) X2 X2y2 R{x) >A iâ€¢ xiyi Ri ( xy) H Ri (u) Ri ( xy) 5 Ri (u) X2 T m Ri (xy) ^^toCx^) Xi (Fig. (2.3).
PAGE 41
34 Henceforth we conclude that the set of functions D defined by (2.1) forms a ncmccmnatative ring with respect to addition and convoluticn operations . Remark 2.1 ; It is important to note that, in the case v^en for all ueR(S) , Rj (u) = {t I < t < u} , the ordinary convolution in m (m > 1) dimensions is defined and all the previous properties with the Ccranutative law hold. Ihus, in a sense, the ordinary convoluticai is a special case of the gcoivolutiai. It nay also be noted that for the case when r^ is admissible the gCCTivolutiai of any function reduces to the ordinary ccnvolution. Remark 2.2 ; lÂ£t (p{x) be a locally integrable density function of the randan variable D . For every xew (S) , define ^[l)M = I 4)(t) dt xte nj (x) X f^^^X) = I (j)(t) dt and for n >^ 2 *(n)^2i) = J "^(nD^^^lt) (tCt) dt XtE h) (x) X ff'^^x) = I f(nl)(xt) (t)(t) dt I
PAGE 42
35 Fran the above definitions and the definition of the set a)(x) r we have by induction *(n)(x) 1 ^^^hx) (n > 1) Thxjs, ?Â«I>(n)(x) < It^'^^x) n=l ^ ' ~n=l By definition *(x) = r<)(t) dt < I ^ L.UCti ti,...,tn,) dti... dt.... dtn, Jo JO Â•'o ^0 Xi
PAGE 43
36 = f ^ $.^^^x.t.) ()).(t.) dt. =$^^Ux ) Jn^ 1111 1 i i cind ^^^Ux) < i.^")(x^) (n > 1) Henceforth, n=l ^^> n=l ~ ~ n=l ^ ^ Fran renewal theory in one dimension [2] , 00 I i.^^' (x.) converges uniformly to il). (x.) , v^ich is tJie expected n=l ^ "Â• number of renewals that have occurred in the close interval [0,Xj] . Therefore, the series being dominated by a tmiformly convergent series is itself uniformly convergent for every xew (S) . 2.3 On the Solution of an Integral Equation Consider the set of functions F, f (t)eF is a real valued function of m variables (t ,t , . . . ,t ) for which we have J 2 m 1. f (t) H f or t < 2. f (t) is locally integrable on a)(S)
PAGE 44
37 Lonna 2.1 ; Let f , geF and h be a locally integrable function cxi u(S) For xea)(S) , if we denote by a(x) = f fCx^t) [ JtecoCx) Y,^oi (t) g(ty) hi^) dj^J dt b(x) = [ f(xy^) g(v) dv] h(^) dy^ '^eco(x) ^VÂ£a)( xy ) C(x) , = [ [ ^e 10 (x) ' veRi ( xy) then, a(x) = b(x) = C(x) Proof ; In (2.5) g ( ty ) = for YL ^ t. Hence we can write (2.5) (2.6) f ( xy^) g(v) dv] h(Y) ^ (2.7) a(x) = f(x2t) [ 'tea)(x) '^ew(x) Let t = V + y , then (2.6) beoones g(ty) h(Y) <^] dt (2.8) b(x) = [ Â•'^ea)(x) f (xt) g(ty) dt] h (y^) d^; tye o) ( xy ) Now since tye o) ( xy) inplies that teu)(x) , we can write b(x) = [ [[ f(x2t) gitr^) dt] h(^) d^ Â•'^e(i)(x) Â•'tea)(x) (2.9)
PAGE 45
38 The integrals given by (2.8) and (2.9) exist . Itierefore, by Fubini's theorem the iterated integrals are equal. Usance for Xje:u)(S) we have a(x) = b(x) Fran (2.6) since g(v) =0 for v^R, ( xy) eo) ( xy) , we have ^M = I f f _ f( xyv) g(v) dv ] hi^) d;i_= C{x) "^yeto(x) ^veR, (xv) And this cxanpletes the proof of the lerma. Note : Let xyv = t, then fron (2.7) C(x) = f [ f f(t) g( xyt) dt ] h(Y.) d^;; ^ileu(x) ^teRi(xy) Renark 2.3 : Let (b(t) be a locally integrable density fxinction of the randcra variable D , and let ^. v (t ) , n >^ 1 , be the nf old gCOTivolution of ({)(t) with itself. From the gcOTivolution properties we have f _ (t, ( xyv) 4, (V) dv= f 4, ( xyv) (t)(v) dv veRi(x;^) ^^'^1 htR^{J^ ^^'^f Hence fron Larma 2.1 f
PAGE 46
39 Similarly frcm the gcx>nvolution properties and Lentia 2.1 we have ' tea) (x) *(nl)(^) iew(t) ' ( ty) h(i^) &^ dt Â•'^eu)(x) In E^, Fig. (2.4) illustrates georetrically the sets tuCx) , o) ( xy) , and Ri ( xy) . Theorem 2.1 : Let (J)(x) be a locally integrable density function of the randan variable D. Ihen, for every xe(D(S) and a(x)  < Â«> there exists a unique solution to U(x) = A(x) + I ,. (xHt) , (f), . (xt) is the nfold gconvoluticxi of Â£n^(n) n=l (n) (f)(xt) with itself, and \^{x)\\ = Si;^ a(x) . xeo) (S) Proof : To prove solution existence and uniqueness, successive approximation method is used. Consider the sequence of functions defined inductively as follows: Uo(x) = A(x) Ui(x) = A(x) + f Uo(t) (j)(xt) dt tea) (x)
PAGE 47
40 h4
PAGE 48
41 and U ^^^ (x) = A(x) + f Uj^(t) ^(xt) dt (n >_ 0) ^tea)(x) where the subscripts refer to the inductive process. Let A,denotes the collection of all such functions. Aj farms a Beinach space in vtiich the norm of an element is defined by U(x) = Sup U(x) xew (S) To proceed in the proof, first v;e wish to show that the sequence of fimctions {Un(x)} converges to a function U(x) for all xeu)(S). Next, we shall shew that this function U(x) satisfies (2.10), and finally that U(x) is a unique solution to (2.10). From the recurrence relations we can write " ^1 ^^) U^^'^) = f 'l'(^t) I " (t) U , (t) ] dt (n > 1) n+1 n J^^^(j^) n n1 Â— Ta3cing absolute value of both sides we obtain l"n+l(?i) "n^^H ^ f ^(x:!) Un(t) V^iit) I dt (n > 1) 'tew(x) Let m = A(x) I Upon taking norms of both sides, of the above relations, we get for n = 0, 1, 2, . . . U (x) U (x) I < mf (j)(t) dt = m*, . (x)
PAGE 49
42 Uj (x) "i (x) I < mf $ (xt) t (t) dt = m* (x) and K^lM U^(x) l"ij__^^^i(n)(?ilt) Mt) dt=m^^^^j(x) xte o) (x) (n > 0) With this inecfuality at our disposal we now show that the sequence of functions (Uj^{x)} is a Cauchy sequence with respect to the norm. The proof is as fellows : Taking norms of both sides, we obtain 00 Since, from Ratiark 2.2, the series J] 1'( j (x) is uniformly convergent n=l " for every xew(S) then the sequence of functions {Uj^{x)} converges to a function U(x) for every xew(S) . Moreover, U(x) is an element of Aj since Aj is a cotplete Banach space. r^t us find an analytical expression for U(x) . Let U(x) be an initial solution for (2.10). Using the results from Latma 2.1, Remark 2.3,
PAGE 50
43 and the results of the goaivolution properties, we get by induction for N = 1, 2, . . . and Uj(x) = A(x) + U^W = A(x) + = A{x) + = A(x) + teu) (x) ({Â•(xt) U(t) dt tew (x) <>>(xt) [A(t) + I ^t,e(o(t) 4>(tt,) U(t,) dt,] dt ()>(xt) A(t) dt teu) (x) Â«t> (xt) [ I ^'(tit,) U(t,) dt, ] dt teu (x) tj eu) (x) (tÂ» (xt) A(t) dt + f p, . (xt) U(t) dt tew(x) ^tea)(x) ^^' N1 U..(x) = A(x) + j; f <>in)^^~^^ ^(t) dt ^ n=l^teu)(x) ^"^ I *{N)(2Slt) U(t) dt (N > 2) tea) (x) where ^ , . (xt) , n > 1 , is the nf old gconvolutiCTi of ( xt) with (n) Â— itself. Therefore, U(x) is given by CO U(X) =A(x) + I f , .(>^) A(t) dt ~ ~ n=l ^teu)(x) ^^' and (2.11) follows.
PAGE 51
44 To show that (2.11) satisfies (2.10) we proceed as follows : Using (2.11) we can write , U(x) f (t) (xt) U(t) dt = A(x) + f \i) (xt) A(t) dt Â•'tew(x) ^tew(x) f {xt) [A(t) + f \l){tt^) A(tj) dt,] dt ^tea)(x) Jt,ew(t) Applying the goonvolution properties, Lemia 2.1, Remark 3.3, and the definition of tj; (t) , we get U(x) f ((Â»(xt) U(t) dt = A(x) + y f *, ^ (xt) A(t) dt ~ ^tea)(x) ~" n=l ^tew(x) ^^' f () (x^t ) A (t ) dt + f ((I ( xt) Â• 'tea)(x) Â•'tea)(x) n=l %ea)(t) ^""^ i Â—i _ = A(x) + y f <\>, .(xt) A(t) dt f (j)(xt) A(t) dt n=l Jtea)(x) ^"' ^teu)(x) n^l itea)(x) <"+!) = A(x) Thus (2.11) satisfies (2.10). Conversely, we can show that if U(x) satisfies (2.10), then U(x) satisfies (2.11). Using (2.10) we can write
PAGE 52
45 A(x) + f ^ixt) A(t) dt = U{x) f (j)(x2t) U(t) dt ''tEw(x) ^te(jj(x) + f ^(xt) [ U(t) f <.(tt,) U(t,) dt,] dt ^teu)(x) ^t eu)(t) ' ~ ~ ~ ^1^^^^ Using, again, Leaina 2.1, Remark 2.3, and the definition of ^{t} , we get A(x) + I (xt) A(t) dt = U(x) f Â» (xt) U(t) dt tea)(x) ^tEu(x) 00 + n f 't'fn)^'^^) ^^(^) '^t ] n=l ^tEw(x) ^"^ ~ I J ''>(n+i)(?Elt) U(t) dt n=l 'tea)(x) ^""^^^ = U(x) Thus the solution for (2.10) is given by (2.11). Now if the set T, as defined in Section 2.1, is admissible, then (t> (t) = (t){n) (t) where (Â»^") (t) is the nfold, n > 1, ordinary (n) Â— Â— convolution of i>(t) with itself. Thus the solution for (2.10) will be given by 00 . U{x) = A(x) + I (t> ^"^ (xt) A(t) dt n=l ^tEa)(x) Q. E. D.
PAGE 53
46 2.4 Other Mathgratical Concepts As we shall see later the following theorem fron calculus, Oliivstedl will be needed in Chapters IV and V. Theorem 2.2 ; Mean Value Theorem for multiple integrals. If f (t) is a OOTitinuous function on a ccrpact set T for which A(T) = I dt ^ 0, then there exists a point v in the interior ^T of T such that .(T) f f(t) dt = f (V) A( Remark 2.4; Frcm the theory of Linear Operators in Banach Space the follcvdng results will be needed in Chapter III. Define the operation (x) = f a (xt) (})(t) dt +f a(x) (^(t) dt ^xte(o(x) ^ xte r'(x) = f a(x,t) ())(t) dt 'xter"(x) = (j) ( xv) a(x,v) dv xea)(S) 'vet" (X) With t" = 1, t" is the n*^ iterate of T, for n = 1, 2, . . . T is defined for all functions a{v) which are bounded for all ver" (x) , xew (S) . The collection of all such functions forms a Banach (M) Space in which the norm of an elonent is defined to be iSee additional references.
PAGE 54
47 Sup Ia{v) ver" (S) Foo: such integral operation T with bounded density kernel, Yosida and Kakutan [22] showed that tP = Z T. + P" ^^ 1 i=l = B + P" n = 1, 2, . where B is the ergodic part of the operation t" and P^ is the nonergodic part of '^, k is the number of proper values of T, T^ (i = 1, 2, . . ., k) is a ccmpletely continuous linear operation and p'^ (n ^ 1) is a corpletely oontimxxis linear operaticai (which might vanish) with Ml < x_^ (l+e)" where r and e are positive constants independent of n.
PAGE 55
CHAPTE3^ III MATHEMATICAL FOTMULATION 3 . 1 Introduction An mcomiDdity (m >^ 1) inventory control system operating under a (a,S) policy is considered. At the beginning of each period a decision to order or not to order is made depoxling on the stock level X = (Xj, Xj, . . ., x ) . The ordering decision in each period is affected, as discussed previously in Chapter I, by a single fixed setup cost K, a linear variable ordering cost c = (Cj, Cj, . . ., cr ) , and the expected holding and shortage cost L(x) conditional on being in stock level x at the beginning of a period. Demand, (Dj^), for the itenis over a sequence of periods (i = 1, 2, ... .) , is assumed to be independently and identically distributed continuous random variables with joint density fimction (i)^{t) , t > 0, t = (t,, t,, . . ., t ). uÂ— Â— * m Iinnediate delivery of orders and ocrplete backlogging of unfilled demands are assumed for all ccfmodities. As discussed in Chapter I, under the adopted stationary (ci,S) policy, the sequence of stock levels at the beginning of each period forms a discretetime Markov process. Ttiis chapter is devoted to finding an expression for the stationary total expected cost per period based on the assumed ordering policy. 48
PAGE 56
49 3,2 Mathenatical Formulation For an n period prcfclem (n >^ 1) , let X = initial stock level for the items prior to making a decision, f (x) = total expected cost of operating the system for n periods when no order is placed initially, i.e.. h (x) = total expected cost of operating the system for n periods when an order is placed initially, i.e., xea. If no order is placed initially, i.e., xca^, then f (x) satisfies the functional equation f , (X) = L(x) (n = 1) f^(x) = L(x) + I Vl^^^^ ^^^^ ^ xtew(x) + f hÂ„_i(xt) {t) dt (n > 2) ^xter> (X) ^ ^ ' (3.1) However, if an order is placed initially, i.e., xea, then hu.(x) satisfies the fimctiOTial equation
PAGE 57
hi (x) = K + c (S^x) + L (S) h^(x) = K + c (S^Xs + L(S) (n 1) 'L Stew(S) Vl^Â— ^ *^^^ ^ + I , h (S^t) 4. (t) dt (n > 2) I ^StÂ£r^(S)5o ^ ^ ~ Â— J Fran (3.1) and (3.2) 50 (3.2) h (x) = K + c (Sx) + f (S) n Â— n Â— xea(n > 1) (3.3) To find, ej^licitly, an analytic expression for the stationary total expected cost per period, the followir^ theorem, based on Howard [6] and White^ work, is needed: Theorem 3.1 For large n (a) h^(x) = ng + v(x) xea (b) f^(x) = ng + u(x) xea^ where g is the stationary total expected cost of the process per period and u(x) and v(x) are functions of the initial state of the process . Proof : Fran (3.1) and (3.3) for xea^ fi(x) = L(x) fj^(x) = L(x) + xtecj (x) + {K+ c [S(xt)]} <),(t) dt ''xterMx) ~ Â— + Vl^^) J^^^^^i ^^j ^(t) dt (n >. 2)^ (3.4) ^See additional references.
PAGE 58
51 Let us denote by (X) = f a(x) = {K + c (S(xt))} (f(t) dt 'xterMx) (3.5) b ,(S,x) = f AS) ni ni Â— xter^ (x) (t)(t) dt (n > 2) (3.6) Upon using (3.5) and (3.6) in (3.4) we get fi(x) = L(x) f (x) = L(x) + a(x) + b ,(S,x) + I 4. (xt) f , (t) dt n Â— Â— Â— ni . , . ni Â— Â— tew (x) (n > 2) Fran the above relation, on using the goonvolution prc^jerties and Remark 2.3, we obtain by induction for n = 2,3,... f2(x) = L(x) + a(x) + bi (S,x) + f3(x) = L(x) + a(x) + b2(S,x) + (^(xt) L(t) dt tea) (x) teu) (x) * (xt) X [L(t) + a(t) + bi(S,t) + (tti) L(ti) dti] dt Â•'tjlÂ£U)(t) = L(x)+ a(x) + b2(S,x) + 4> (xt) [L(t) + a(t) ' tew (x) + bi(S,t)] dt + I (j)(xt) ' tea) (x) I.. eo)(t) <}.(tti) L(ti) dti] dt
PAGE 59
52 L(x) + a(x) + b, (S,x) + f 4) ( xt) [L(t) + a(t) ~ ^tea)(x) + b, (S,t)] dt + f 4) , . (xt) L{t) dt and n2 f (x) = L(x) + a(x) +b (S,x) + y ((,,.. (xt). "" ~ ^1iÂ£l Jt^Â„(x) <^' [L(t) + a(t) + b ^ . .(S,t)] dt f ^. ,. (xt) L(t; (t) dt (n > 2) tea) (x) v*iere (j) ( xt) , n >_ 1 , is the nf old gconvolution of <(> ( xt) with itself. On regrouping terms, we get n1 f_(x) = tL(x) + y tEO) (X) ((Â». (xt) L(t) dt 1 n2 , + [a(x) + y ~ i=l tew (x) (t, (xt) a(t) dt] n2 , + [b (S,x) + I 2) (3.7)
PAGE 60
53 To proceed in the proof, first, we shall siinplify (3.7) by substituting for a(x) as given by (3.5). Mext, we shall find an analytical expression for f (S) and use it in finding the expression for f (x) , jKa^, for large n. Finally, we shall determine the expression for hyj(x)f xea, for large values of n. Let y = (mi ,y2 f Â• Â• '/M ^ r where y is the expected value of 'Â— ni 1 the randan variable D., i = 1, 2, . . . , m. Fron (3.5) and the definition of the sets w (x) and r ' (x) , we have a(x) = f [K + c'^(S(x2t))] (xt) dt tew (x) = K + 0*^(3(xy)) I [K + c'^(St)] (xt) dt '^tew(x) Upon using the gconvolution properties and Lema 2.1 we obtain for i = 1, 2, . . . f (), (xt) a(t) dt = f i>, , (xt) {K + 0*^(5(ty)) ^tea)(x) t^) 'tew(x) ^^' [ [K + c'r(5ti)3 6(tt,) dti) dt 'tie(Â»)(t)
PAGE 61
54 I 4>{xHb) [K + 0*^(3(ty))] dt tea)(x) f r,)(xt) [K + c'^(S2t)) dt 'tew(x) and, f (J) ( xt) a(t) dt= f () ( xt) [K + c'^(S(tij))] dt ^tFMfxl (i) ^trM(x) (i) f r,xi\(xt) [K + c'^{St)] dt (i > 1) Upon using the above relations, we obtain n2 x) + y f (}),.. (xt) a(t) dt = K + c'^(Sx) + c^v i=l ^tÂ£a)(x) ^^' J {K + c'^(St)} (i)(xt) dt tea)(x) n2 Ml j (t)(j^)(xt) [K tc'^( St+y) ] dt tea){x) f 'J'rxix (?^) fK + c'^(St)] dt On siitplifying cind regrouping terms we get n2 a(x) + Y f (J),., (x^t) a(t) dt i=l ^tewCx) ^^^ T ^~^ f T = K + c ^ [1 + y * , . V (xt) dt ] + c (Sx) i=l ^tEai(x) ^^>
PAGE 62
55 " I *(nl) (?^) f^ + c'^CS^t)] dt (3.8) .' Now, we shall proceed to find an analytical expression for fj^CS), n > 2. Fran (3.4), for x=S, we have f,(S) = L(S) fCS) = L(S) + f f fSt) (j)(t) dt ^Steu)(S) ""^ + f [K + c\] 4,(t) dt ^SterÂ»(S) 'SterMS) f(t) dt + f^ (S) f 4,(t) dt (n > 2) Let us denote by c(S) = f [K + c\] 2) (3.11). 'Ster'(S) Note T as defined is a bounded linear operation. Substituting (3.10) and (3.11) in (3.9) we get fj(S) = L(S) fn(S) = L(S) + T(f^_]^(t)) (S) + c(S) (n > 2) '(3.12).
PAGE 63
56 Fran (3.12), by induction for n = 2, 3, . . . , we have f2(S) = L{S) + c(S) + T(L(t)){S) Â£3(8) = L(S) + c(S) + T{L(t) + c(t) + T(L(t)))(S) and, n2 . fÂ„(S) = L(S) + c(S) + I T^{L(t) + c(t)) (S) i=l ~ ~ + T"l(L(t))(S) (n>2) (3.13) Pron Yosida and Kakutan [22] we have k T^ = I T. + p^ j=l 3 = B +P^ (i = 1, 2, . . .) (3.14) v^ere B is independent of i. The linear operations T^ (j = 1, 2, . . . , k) and P are explicitly defined in Renark 2.4. On substituting (3.14) in (3.13) , we get f^(S) = L(S) + c(S) + [.(B+PÂ») + (B+P2) + . . . + (B+P""^)] L(t)4c:(t))(S) + [B+P""^] (L(t))(S) (n>_2) Upon regrouping terms, we obtain
PAGE 64
57 fjj(S) = L(S) + c(S) + (nl) B(L(t)) (S) n1 + (n2) B(c(t))(S) + [ I Pi ] (L(t))(S) i=l n2 . + t I P^ ] c(t))(S) i=l (n > 2) (3.15) Fran (3.6) and the definition of the sets w(S) and r* (S) , we can write b ,(S,x) = fÂ„ .(S) f (t.(t) dt nlnlJ^_terÂ»(x) = ^nl^S) [f ()(t) dt f ()(xt) dt ] tew(x) = fnl(S) [1 f *(xt) dt ] (n > 2) ^te(D(x) and for n > 2 n2 i=l ., (xt) b , ^ (S,t) dt (i) nli Â— tea) (x) = fn_i(S) [1 I (t'Cxit) dt] + tea)(x) n2 i=l nli f *H\(?Slt) [1 f
PAGE 65
58 n2 = f fS) [1 <}.(xt) dt] + I f , i(S) . [f Â„.iv(xit) dt ] (n >2) (3.16) Denote by F^ = f (f)>.> (xt) dt (i = 1, 2, . . .) (3.17) Jtea)(x) ^^' with F " = 1. Using (3.17) in (3.16) and sinplifying, we get n2 'tew(x) n1 ^nl^^'H) + .1 f , ,*(i)(^) Vli(S,t) dt 1=1 tew(x) n1 Â• 1 Â• = y f . (S) [F^"^ F^] i=l ^^= y f ,(S) [f""^ F^] +f.(S) [f"'^ f"'2] i=l "1 + f,(S) [F^"^ F^'^1 (3.18) On substiuting (3.15) in (3.18) and expanding v;e can write for n >^ 2 n2 'tÂ£w(x) Vl(S,x) + y I ^ (x^) b^_i_i(S,t) dt 1=1 tÂ£w(X) n3 = y [L(S) +c(S) + (nli) B(L(t))(S) 1=1 ~ ~ " nli . + (n2i) B(c(t))(S) + [ I P^l (L(t))(S) + [ I P"] (c(t))(S)] [F^" f"] j=l
PAGE 66
59 + L(S) + C(S) + B{L(t)) (S) + P* (L(t)) (S)] [f"~3 f""^] 12 ji1. L(S) [f"" f"'] = [L(S) + c(S)] [F" F^ + F^ F^ + . . . + f""^ F^~^ 13 n2 + f""^ f" ] + L(S) [f""2 F^1] + B(L(t)) (S) [ (n2) [F" PÂ»] + (n3) [FÂ» F*] + . . . + 2[F^~^ f""^] +[f""^ F"2] 1 + B(c(t)) (S) [(nS) [F" F^]+ (n4) [F' F^] + . . . + n2 [f""* f""^] 1 + [ E P3] (L(t))(S) [fO FÂ»] " j=l n3 . 2 ^ + [ Z P3] (L(t))(S) [FÂ» F^] + . . . + [ I P"] j=l j=l (L(t)) (S) [f"""* F^"^] + P' (L(t)) (S) [Fr^3 f""2] n3 n4 . + [ I P^] (c(t))(S) [f" fM + [ I P"] {c(t))(S) j=l j=l [FÂ» F^] + . . . + PMc(t)){S) [f""^ f""^]
PAGE 67
60 On sinplifying, we obtain n2 ^nl'^''\ilL(x,*<"'''"^^'' dt = L(S) + c(S) c(S) F^"2 L(S)F^"^ + (n2) B(L{t))(S) n2 . B(L(t))(S) I F^ + (n3) B{c(t))(S) i=l n3 . n2 . B(c(t))(S) j; F^ + [ 5; PJ] (L(t))(S) i=l j=l + [ p""^(L(t))(S)FÂ» p""^(L(t))F2. . . P>(L(t)){S)F"'^] + [ P"^(c{t))(S)F' P^'^(c(t)(S)F' . . . PMc(t))(S)F"'^] + [ I I>h (c(t))(S) = L(S) + c(S) c(S)f""^ L(S)f""'" + (n2) B(L(t)) (S) n2 . B(L(t))(S) I F^ + (n3) B(c(t))(S) i=l n3 . n2 . B(c(t))(S) 5; F^ + [ [ P^] (L(t))(S) i=l i=l + [ I V^] (c(t))(S) I P"" ^ '(L(t))(S)F^ i=l i=l n3 . J _ Jpn2i (c{t))(S)F (3.19) i=l
PAGE 68
61 Let n2 nl D^ (S,x) = [ ^ P^] (L(t))(S) + t I P"] (c(t))(S) i=l i=l I p""^"^(L(t))(S)Fi i=l n3 J . . I p""^'^(c(t))(S)F^ Vn>2) (3.20) i=l 'flien (3.19) becxmes (after substituting for F^, i1 D n2 I i(x) n2 / b (S,x) + I 0/.x(xt) bÂ„ , .(S,t) dt "1 i=l ^teu)(x) ^^' ""^ ^ = L(S) + c(S) c(S) f *, ^^ (xt) dt ~ " ~ Â•'tea3(x) ^'^2) L(S) f , ,.(xt) dt + (n2) B(L(t))(S) ^tea3(x) ^""^^ n2 f + (n3) B(c(t))(S) B(L(t))(S) I (>,.v(xt) dt ~ ~ ~ ~ i=l ^teco(x) ^^' ~" n3 , B(c(t))(S) I 2) (3.21) ' Using (3.8) and (3.21) in (3.7) we get n1 fÂ„(x) = [L(x) + I f *,..(xt) L(t) dt] ^~ ~ i=l ^tÂ£a)(x) ^^'
PAGE 69
62 + [K + c n2 , i=l JtEa)(x) ^"^ + c'^{Sx) f , ,,{xt) [K + c'''(St)l dtl + L(S) + C(S) C(S) te(jj(x) ,Â„.2,'!^)& tew (x) n2 + (n3) B(c(t)){S) B(L(t))(S) I ~ ~ i=l n3 tea) (x) L(S) f (J), . . (xt) dt+ (n2) B(L(t))(S) * (i\ ( xt) dt I B(c(t)) (S) y f <),.x (x^) dt + D^ ,(S,x) i=l ^tea)(x) ^^^ n2 (n>_2) (3.22) Next, we shall study the asynptotic behavior of fj^(x) for large n. For 00 n large we have, because of the cx)nvergence of 7 (^ , . ( xt) to ^ ( xt) , i=l ^i' (n) (xt) Â» I (),.,(xt) ^ il>(xt) i=l ^^' Moreover, from Rerark 2.4, P^, i >^ 1, is a cx3tpletely continuous linear operation (which might vanish) with P" l< < 1 (l+e) i = 1, 2,
PAGE 70
63 where r and e are positive caistants independent of i. Thus, for large n, D _2(S,x) as given by (3.20) converges to a function D(S,x) vAuLch decreases rapidly as n beoomes large. Using these last remarks in (3.22) we get, for large n. f^(x) = L(x) + I iHxt) L(t) dt + K tew (x) + c\[l + f ij(xt) dt] + c'^(5x) + L(S) ^tetj(x) + c(S) + nB(L(t)+c(t)) (S) 2B(L(t)) (S) 3B(c(t)) (S) B(L(t))(S) f it>(xt) dt ^tea)(x) B(c(t))(S) f i;(xt) dt + D(S,x) ^tÂ£w(x) Or, f (x) = ng + u(x) xca (3.23) n Â— Â— "~ where g = B(L(t)+c(t)) (S) is independent of n, and u(x) is a fianction of X . Now from (3.3) for large n h (x) = K + c'''(Sx) + fÂ„(S) n Â— " ~ = K + c'^(Sx) + ng + u(S) = ng + v(x) xeo (3.24)
PAGE 71
64 where, v(x) = K + c'^(Sx) + u(S) xea (3.25) Q. E. D. Theoran 3.2 ; The stationary total eiqsected cxjst per period, g, is K + L{S) + I L(St) i/;(t) dt JCÂ•t<^/../C^ 'StÂ£a)(S) g = + c\ (3.26) 1 + ij^a) dt 'steaj(S) where ipit) = I
PAGE 72
T v(x) = v(0) ex xfea 65 (3.29) Using (3.29) in (3.27), we cÂ±)tain ng + u(x) = L(x) + f (nl)g 4, (t) dt ^xtcr"(x) + [ u(xt) ,})(t) dt Â•' Xtc o) (x) + f [v(0) c'^(xt)] (,(t) dt Jxtcr'(x) = L(x) + (nl)q f4)(t) dt + f u(x^) <,(t) dt 9. ^xtea)(x) i + [[v(0) c'^(x2t)] 4,(t) dt Jo [ [v(0) c'''(x2t)] 4,(t) dt Jxteco(x) Siiiplifying, we get g + u(x) = L(x) + v(0) [1 f <^(t) dt] xtea)(x) 'xtea)(x) Let xt = y, then (3.30) beocnes [c'^(xt) 4,(t) dt + f c'^(xt) (j)(t) dt h ''xtEa)(x) + f u(xt) (t)(t) dt (3.30)
PAGE 73
66 g + u(x) = L(x) + v(0) [1 [ (xy) d^] ''^G(d(x) I c\ (i)(xy) dy. + c Y. 4> (xy) d^ ^0 ^ilew(x) + [ u(^) (t) (xy) d^ xea^ (3.31) ^yea)(x) Let fT r T c H(x) = L(x) c Y. (^ (xy) dv + c Y. (t> (xy) dy xea ^0 Jiiea,(x) (3 32) i(x) = [ (})(X2^) d^. (3.33) ^Y.ew(x) On using (3.32) and (3.33) in (3.31), we obtain aftex transposing g to the R.H.S. u(x) = ^ + H{x) + v(0) [1 $(x)] + [ (l,(xy) u(v) dy (3.34) Zew(x) Let A(x) = g + H(x) + v(0) [1 *(x)] (3.35) and (3.34) beocmes u(x) = A(x) + I (&(xy) u(Y.) ^ xeo*^ (3.36) ''^Â£a)(x)
PAGE 74
67 Each cxitiponent of A(x) as given by (3.32) and (3.35) is bounded over the set (d(S). Thus, A(x) is a bounded function over the oorpact set u,(S). Fran Theorem 2.1 the solution for (3.36) is given by u(x) = A(x) + \l) (xy) A(^) d^^ xeo^ (3.37) 'yetÂ»Â»(x) where i()(t) = Y , Jt) . n=l (") ~ Using (3.35) in (3.37) we get u(x) = g + H(x) + v(0) [1 (x)] + f yji (xy) [g + E{y) + v(0)[l i(j^)]] djr This last relation is true for all xea . Set x = S and we get u(S) = g + H(S) + v(0) [1
PAGE 75
(n) Note $ , (S) = $ (S) (i) 68 (b, ,(Sy) dy (n^^l) (3.40) From (3.39) the coefficient of v(0) transposed to the R.H.S. is 1 + [1 i(S)] + f iKSiZ) [1 *(Z)] d^;: Â•'yeuj(S) ' ilew (S) Fran the definition of J , (S) (n >_ 1) , the gconvolution properties, ^(n) ~ and Remark 2.3, we have for n > 1 f *,Â„, (Â§11) *(Z) = f (n) (Sil) [ I ^ *(Zlt) dt] ^ = f [ f d.(n)( Sty)
PAGE 76
69 Or, g = K + c^ + H(S) + I it)(Sy) H(y) djr ^vea)(S) 1 + f ii;(Sy) d^ (3.43) Let M = [1 + f t)(Sy) d^]"'. Substituting in (3.43) for H(t) as ^yew(S) defined by (3.32), aives g = M {K + c^S + L(S) + I c\ (KS^x) dy; ^ileo)(S) F c'y (t) (Sy) dy. + f Â»HSy) (L(j^) ^0 ^Ycui(S) f c'^t (t) (y^) dt f c\ d) (^) dt] dy ^trM(v) ^n Â— 'tew(v) dy} Upcsi sinplifying and arranging terms, we get g = M {K + 0*^5 + L(S) + f \l) (Sy) L(v) dv 0*^5 ^Sye(D(S) + c u + I c\ (i> (Sy) dy + f H;(Sy) Kr^..(c,^ ^iiÂ£w(S) 'yea)(S) T [ [ c t (t)(yt) dt ] d^ + c\ iJ;(Sy) d^ Â•'tew(v) Â•'j^ea)(S) f c\r\){S^ dv } ^VEa)(S) v*iere ]i^ is the expectation vecÂ±or. Siirplifying further we get
PAGE 77
70 g = M {K + L{S) + f i>{S:X^ UZ) ^}+ c\ ' Syeu) (S) + f i;(Sy) [ f <{)(yt) c\dt] dy 'Yew(S) ^teo)(^) I c y ij (Sy) d^ } = M {K + L(S) + ji (Sy) l(y.) djr } + c y Â•'Syeb)(S) + M { I c'''^ ^ ( Sy) ^ Yea)(S) n=l ^^eai(S) I f c:T^*(n)<^^ ^ ^ n=l Veu(S) ^"' 'yew! Note the last result follows frcm the definition of ^ (t) , the convolution properties, and Larrra 2.1. Upon simplifying the last relation and substituting for M, (3.26) follows. Q. E. D. As given by (3.26) the stationary expected cost per period, g, is a function of (a,S) , and will be written as g(cr,S) .
PAGE 78
CHAPIER IV THE OPTIMIZATICN PROBLEM 4.1 Introduction In this ch^ter we shall provide necessary and sufficient conditions for the existence of an optimal (a,S) policy that miminizes the expression for g(a,S) as given by (3.26) . The unknowns to be determined are the decision variables S. (i = 1, 2, . . . , m) and the set r. In Section 4.2 we shall first determine the configuration of the set r and then reduce the minimizaticai of g(a,S) to finding optimal values for S and an additional variable C, vtiere C will be identified. The purpose of Section 4.3 will be to determine the necessary and sufficient conditions for the existence of the pair (C* ,S) that minimizes g (a ,S) . The necessary conditions will be given for the general mooitniodity inventory prcÂ±)leni, wiiile the sufficient conditions will be given only for the twocommodity problem with separable L(x) . Ihe set of equations satisfied by (C*,S*) will be restated in terms of a real valued function M(C*,x) , where M(C*,x) will be defined. Section 4.4 will give a geotvetrical formulation of the cptimizaticn problem in terms of the function M(C*,x) and its inherent prc^jerties . Finally in Section 4.5 the linear form for LCjc) , the conditional expected holding and shortage cost function, will be considered. 71
PAGE 79
72 4t2 Characterization of r In this section we shall first characterize the set r, r = {xjxen; Z (Sx) = 0}, by deterndning explicitly the relation Z (Sx) = up to a constant. This then will be used to reduce the miniirdzation problan into a problem in differential and integral calculus. Theorem 4.1 ; Under an optimal (a*,S*) policy, g(a*,S*), the minimum stationary total expected cost per period, satisfies the following relation: g{o*,s*) = LCx") + c\ x'er* (4.1) where jj^ is the expectation vector of the random variable D with joint density function A (t) . Proof: The proof is based on the definitions of h^(x) , xea^ and f^(x) , Xeo*^, as given in Chapter III. At optiirality we must have for x'eF* h (x") = f (x") (n > 1) n Â— n Â— Â— We note that this last relation if explicitly established will determine the configuration of r*. For large n, f (x) as given by (3.1) can be written f (x) = L(x) + f [(nl)g + u(xt)] ^{t) dt " ~ ^xteaj(x) + f[(nl)g + v{xt)] {t) dt
PAGE 80
73 J [(nl)g tv(xt)i(t>(t) dt xea^ (4.2) 'xHtew(x) V*iere g is the stationary total expected cost per period and u(.) and v(.) are well defined in Chapter III. For x"er*, and the definition of w(x), the set a)(xÂ°) is an enpty set. Hence, from (4.2) we can write fj^Cx") = L(x'') + F [(nl)g + vCx^t)] <^(t) dt, x'el* (4.3) using (3.25) and the asynptotic ejqpressiOTi for f (x') in (4.3) we obtain T ng + K + c (S*x'') + u(S*) = L(x'') + r [(nl)q + K + c ( S*x +t) + u(S*)] <(.(t) dt ^0 ~ ~ ~ = L(x'') + (nl)g + K + c'^( S*xÂ° ) + c'^^vi + u(S*) siitplifying we obtain (4.1). 0. E. D. Ihe result of Theorem 4.1, as we shall see later, will be of great inportance in finding the necessary and .sufficient conditions for the ecistence of an optimal policy that minimizes the expression for g(a,S) as given by (3.26). An explicit expression for L(x'), x^eP*, can be iirmediately derived by ccnparing (3.26) and (4.1) to yield
PAGE 81
74 K + L(S*) + f L( S*t) if)(t) dt ^S*tca)(S*) L(xÂ») = j^ (4.4) , 1 + \ij(t) dt ^S*tea)(S*) which is sijnilar to Iglehart's result [7] for the single ccmodity case operating under the ( 3;S) policy. For notational purposes v;e shall sonetimes represent L{x'') by the symbol C*, C* refers to the value of the total expected cost per period excluding the variable ordering cost. Corollary 4.1 : r* = {xxGn; L(x) C* = 0} (4.5) Proof ; Fran Section 2.2 r* = {xxen; Z ( S*x) = 0} Fran Theorem 4.1 L(x) c* = xer* Hence, Z ( S*x) = L(x) C* = xer* and (4.5) follows. Q. E. D. Let us make the transformation x = S*t ; then (Chapter II) the image of the set a)(S*) is R(S*) and the image of r* is To* = {tt > 0; L( S*t) C* = } (4.6) The sets R(S*) and a)(S*) are functions of C* and often will be referred to as R(S*Â»C*) and w(S*,C*).
PAGE 82
75 Wfe may rewrite the expression of L(x ) , x"er*, as K + L(S*) + f L( S*t) \l)(t) dt ^R(S*,C*) L(x^) = . (4.7) I + I ^^(t) dt R(S*,C*) Â•nie relation given by (4.1) , then, may be written T. K + L(S*) + f L( S*t) ij^(t) dt ^R(S*,C*) g{a*,s*) = . + c"y 1 + \l,{t) dt ^R{S*,C*) Fran this last relation it nay be noted that to minimize g(o,S) relative to the set a and S is equivalent to K + L(S) + I L{Frt) i!Â»(t) dt Rf^ C) Min { g, (S,C) = 7 } (4.8) 1 + ijj(t) dt ^R(S,C) which is a problem in integral and differential calculus. Let M = [1 + f ipit) dt ]~\ < M < 1, ^R(S,C) then the minimization problan as given by (4.8) is equivalent to: min {g,(M,C,S) = M [K + L(S) + f L(St) (^(t) dt] } * ^r> /o r>\ 'R(S,C) S.T. {g3(M,C,S) = M [1 + f \\){t) dt] 1 = } ^R(S,C) (4.9)
PAGE 83
76 4.3 ^tec:essary and Sufficient Conditions The Lagrange funcÂ±ion, G{M,C,S,X), for (4.9) is given by G(M,C,S,X) = g2(M,C,S) + Xgj(M,C,S) = M{K + L(S) + f L(St) i;(t) dt} ^R(S,C) + X{1 M Mf ,J;(t) dt } (4.10) 'R(S,C) By definition a point (M*,C*,S*) is a proper relative minimum for g2(M,C,S) if g2(M*+^, C*+AC, S*+AS) g2(M*,C*,S*) > or equivalently AP = G(M*+AM, C*+AC, S*+AS ,X*fAA) G(M*,C*,S*,X*) > (4.11) To determine the necessary and sufficient conditions for the existence of a relative minimum for (4.9) , we shall first obtain an analytical expression for AG as defined in (4.11). On using (4.10) in (4.11) the expression for AG can be written as AG = (M*+AM) {K^+ L( S*+AS) + f L( S*+ASt) t(;(t) dt } ^R( S*+AS ,C*+AC) + (X*+AX) (1 (M*+AM) (M*+AM)f r\,{t) dt } ''r(S^+AS,C*+AC) M* {K + L(S*) + f L( S*t) \l){t) dt } ^R(S*,C*)
PAGE 84
77 X* {1 M* M* iÂ»(t) dt > ^R(S*,C*) Upon recirranging terms we obtain AG = M* {L( ^4A5) L(S*) + f [L( S*Â»Ast) X*] i^ (t) dt ^R(SMAS,C*4AC) f [L( S*t) \*] ^:t) dt + AM (k + L( S*fAS) A* ^R(S*,C*) + f [L( S*4ASt ) X*] i)(t) dt } ^R( S*fAS >C*fAC) + AXd (M*+AM) (M*fAM) ilfCt) dt } (4.12) ^R( S*fV\S ,C*fAC) If we denote by AR(S*,C*) the incremental set between R( S*fAs ,C*>AC) and R{S*,C*) , then we can write f [LfS^+AS^t) X*] \l)(t) dt 'r(s*+as,c*+ac) ^R(S*, ^R(S*,C* [L( S*t ) X*] ii;(t) dt C*) [L( S*tASt) L( S*t) ] ^{t) dt ) + f [L( S*+ASt) X*] rl)(t) dt ^AR(S*,C*) (4.13) and
PAGE 85
78 J [L( S*+ASt ) A*] i){t) dt = f [L( S*HASt) A*] ii;{t) dt ^R(S*,C*) ~ Â— + [ [L( S*+ASt) A*] ti;(t) dt (4 14) Using (4.13) and (4.14) in (4.12) we obtain /G = M* {L(S^N^) L(S*) + f [L( S*tASt) L(S*t)] ij;(t) dt ^R(S*,C*) Â— + f [L(S^+^t) A*] Mt) dt } ^tf^(S*,C*) ~ + ^ {K + L(S^+AS) A*[l + [ ,,(t) dt] ^R(S*,C*) ~~ + I L(S*+ASt) i;(t) dt ^R(S*,C*) ~ ~ + f [L( S*+ASt ) A*] nt) dt} Â•'ar(s*,c*) Â— + AA {1 {M*+m) (M*+/!M) f \p{t) dt } (4.15) ^R( S*+AS ,C*+ac) Let t''eAR(s*,C*) and let A{m) denotes the area of the increniental set Z1R(S*,C*). Then from TheorOT 2.2 J [L( S*+ASt) A*] ij(t) dt m(s*,c*) ~ Â— = [L( 5*+^t Â°) A*] Mt") Aim using this in (4.15) v;e obtain
PAGE 86
79 AG = M* {L(S*+AS) L(S*) + f [L( S*^ASt) L( S*t) ] ii;(t) dt ^R(S*,C*) + [L( S*^ASt ) X*) \l)it^) A(AR)} + AM {K + L( S*tAS ) X*[l + f t;(t) dt] + f L( S*tASt ) \l)(t) dt ^R(S*,C*) Â•'R(S*,C*) + [L(S*+ASt'') X*] 'j/(t'') A(AR)} + AX {1 (M*+AM) (M*+AM) 'Mt) dt } (4.16) Â•'r(s*+as,c*+ac) Let t^eV^*. Expandinq !.(Â•) in a Taylor series, with a rerrainder, rM'), about (S*,C*), the expression for AG as given in (4.16) becones m 3L(S) I , m m 3*L(S) AG = M* { J: I ASi + 5" I I AS . AS . i=i 8s, s^s* ^ ^ 1=1 j=i as^aSj ' s=s* ^ => , m 3L(St) [ I :=Â— I AS 'r(SÂ«,C*) 1=1 3S. &=S* . m m 3^L(St) 1=1 j=l 3Sj^9Sj &=S*
PAGE 87
80 m 3L(Sto) I + {L( S*t ) \* + I I AS. i=l 3S. S=S* 1 m 8L( S*t ) + I At^} i;(to) A(AR)}+ AM (K + L(S*) i=l 3tjL t=to m 3L(S) I + I I AS. \* [1 + ^(t) dtl i=l 9S^ &=S* ^R(S*,C*) m 9L(St) I + [L( S*t ) + I I AS.] ,J;(t) dt ^R(S*,C*) i=l 9Si S=S* m 3L(Sto) I + [L( S*tJ X* + ^ I AS. i=i as^ &=S* m 9L( S*t) + I At.] i;(t^) A(AR)} + AX {1 (M*+AM) i=l at. t=to (M*+^) [ ^it) dt } ^R(SM^,C*+AC) + RMS*,C*,M*,X*; S*+AS >C*fAC,M*+AM.X*+AX) (4.17) The above expression for AG will be used to determine the necessary and sxifficient conditions for gj (Â£,C) to have a proper relative minimum at (S*,C*) .
PAGE 88
81 Necessary Conditions We shall first find the set of simultaneous conditions that (S*,C*) will satisfy and then restate these conditions in terms of the function M(C*,x) which will be defined later. The following Lama is needed. Lgma 4.1 ; At cptimality K + L(S*) + f L( S*t) ^(t) dt ^R(S*,C*) X* = = C* > n (4.18) 1 + f ti;{t) dt ^R(S*,C*) Proof ; The partial derivative of G(M,C,S,X), as given by (4.10), with respect to M, evaluated at (M*,C*,S*,X*) , yields = K + L(S*) + f L( S*t) ij;(t) dt and 3M (M*,C*,S*,X*) 'R(S*,C*) X* [1 + f t>(t) dt] = Jr(s*,c*) K + L(S*) + f L( S*t) ^^(t) dt Â•'r(s*,c*) * 1 + f i,{t) dt Jr(s*,c*) Fran this last result and (4.7), (4.18) follc*^. Q. E. D.
PAGE 89
82 Theorgn 4.2 ; nie optimal values of C > and S > , C* and S* , vtiich ndnimize g^ {S,C) , satisfy the follodng set of equations: 9L(S) ^ptQ* 9L(St) as. S=S* 'R(S*,C*) 3S^ ^S* C* L(S*) + 'R(S*,C*) ii;(t) dt = (i = l,2,...,m) (4.19) [C* L( S*t) ] i)(t) dt = K (4.20) Proof: The partials of G(M,C,S,X) with respect to S^ (i = 1, 2, . . . ,m) are detained frcm (4.17) 9G 3S^ (M*,C*,S*,X*) = M*[ 3L(S) as . s=s* 1 aL(st) 'R(s*,c*) as^ ^=s* ^{t) dt aA(AR) + [L( S*t o) X*] i;(to) ] = ^\ (i = 1, 2, . . .,m) where the point toePo*. Smce by (4.6) and (418) X* =C* = L(S*t,). t,er/, (4.19) follows. To prove the second part of the theorem we have frcm (4.18)
PAGE 90
C* [1 + f ,;(t) dt] = K + L{S*) + f L( S*t) r\){t) dt ^R(S*,C*) ^R(S*,C*) And this last relation can be written as C* L(S*) + f [C* L( S*t) ] iKt) dt = K R(S*,C*) 0. E. D. Theorem 4.3! Let the function M(C*,x) satisfy the integral equation M(C*,x) = C* L(x) + f M(C*,X;;t) (^ (t) dt (4.21) Â•'r(x,C*) = C* L(x) + f (t) (xt) M(C*,t) dt (4.22) ~ ''a)(x,C*) Suppose that C* aixl S* exist, which minimize g, (C,S) subject to C > and S > 0. Then, if S* > and C* > 0, (C*,S*) is a solution to the set of equations M(C*,S*) = K 3M(C*,x) I 3x . JC=S* (4.23) = (i = 1, 2, . . ., m) (4.24) Proof: Define the functions 3L(x) . 3L(Kt) M, (C*,x) = + rl){t) dt (i = 1, 2, ..., m) 3x^ 'r{x,C*) dxi (4.25) 1
PAGE 91
84 M(C*,x) = C* L(x) + f [C* L(xt)] ii)(t) dt (4,26) ''r(x,C*) = C* L(x) + f [C* L(t)] i) (xt) dt (4.27) Â•'u(C*,x) Let X = S* in (4.25) and (4.26) then ccnparing the results with (4.19) and (4.20) we obtain M(C*,S*) = K (4.28) M(C*,S*) = (i = 1, 2, . . ., m) (4.29) From Theoran 2.1 the solution of (4.22) is given by (4.27). Thus M(C*,x) = M(C*,x) (4.30) Hence from (4.28), (4.23) results which proves the first part of the theoran. To prove the second part of the theoran we can write from (4.26) m{C*,x) = M(C*,x+Ax) M(C*,x) = C* L( x+Ax) + f [C* L( x+Axt) l i>{t^ dt ^R(x+Ax,C*) [C* L(x) + f [C* L(x2:t)] iKt) dt] ^R(x,C*) If we denote by AR(x) the incratental set between R( xfAx ,C*) and R(x,C*) , then the expression for AM(C*,x) can be written as
PAGE 92
85 Z*1(C*,x) = [L( x+ X) L(x)l + f tC* LC x+Axt) ] ,j;(t) dt ^R(x,C*) + f [C* L( xÂ»/yt )] \i,{t) dt 'AR(x) ^R(x, (C* L(xt)] Mt) dt C*) Let t^eARCx) arri let A(AR(x)) denotes the area of AR(x). Using the Mean Value Theoran for multiple integrals, TTieorejn 2.2, we can write /^(C*,x) = [L(x+^) L(x)] f [L( xfAXt) L(X2t)li;(t) dt JR{X,C*) + [C* L( x+Axt P)] ,^(tÂ») A(AR(x)) (4. 31) Expanding L{x+Ax), L( x+Ayt Â°) in a Taylor series, with a ranainder, R' ( xÂ»AX ,x;C*) , about x, the expression for ^(C*,x) becomes m 3L(x) f m aL(xt) mc*,x) = I AX. I ^Ax. V;(t) dt i=l 3x^ ^ Jr{x,C*) i=l 3x^ m 9L(xtÂ°) + [C* LCxt") I AX^l ij;(tÂ») A(AR(x)) i=l 3x^ + R* (x+Ax,x;C*)
PAGE 93
86 As Ax , t " tjeFo (x) . Hence lijn (Lfxt") C*> = 0. Dividing Ax^n AM(C*,x) by Ax^ and lettijig Ax^ ^ we get 3M(C*,x) 3L(x) 3L(xt) "(j;(t) dt ] (4.32) 9Xj^ 8x^ jR(x,C*) 3x^ (i = 1, 2, . . ., m) Copnparing (4.25) and (4.32) we cistain 3M(C*,x) = M. (C*,x) (i = 1, 2, . . ., m) 3x. Hence frcm (4.29), (4.24) follows. 0. E. D. Sufficient Conditions So far the necessary ocnditions for the determination of the optimal values of C and S v/ere discussed. To study the sufficient conditions that have to be satisfied, vre shall assume that the function L(x) is separable, that is, m L(x) = y L. (x.) where L. (x.) for all i is twice differentiafale. Now, since gs (M*+AM,C*+AC, S*+AS) = 0, and since the necessary conditions, derived above, require that each derivative of G(M,C,S,X) vanish (4.17) reduces to
PAGE 94
AG = G(M*+AM,C*+AC,SM^,x*+AX) G(M*,C*,S*,X*) 87 T m m aL(St) u u , AS. AS. j=l i=l 3S.8S. S=S* ^ J 1 3 J m m 9L(St)  * J ill AS. AS. ^{X) dt ^R(S*,C*) j=l i=l 9S.8S. S=?* ^ 3 1 : m aL(StJ I + {L(S*to) A* + j; =^ I AS. i=l 3S^ ^S* ^ m 3L(S*t) 1=1 9t^ ^=tc At^ ]'<''t^) A(AR)] + RÂ»(S*,C*,M*,X*;S;Â»+AS,C*+AC,M*+AM,X*+AX) (4.33) **iere to = (tjo, tj^, . . ., t_^) eP/, rMÂ«) is the remainder terms mo of the Taylor series expansion, A(AR) is the area of AR{S*,C*) , and L\(t^), L'\(t^), (i = 1, 2, . . ., m), are the first and second derivatives of L. (t.). 1 1 Since L( 5*t p) = A* = C*, we have fron (4.33) on taking the second partial derivatives, for i, j=l, 2, ...,m a^G =M*[L "(S*) + f L."(S*.t.) ^{t) dt *.r*.^*.\*\ ^ 1 ^R(S*,C*) ^ ^ ^ ~ "~ 3S^2 (M*,C*,S*,A*) + L.'(S.*t.,) K,(to) aA(AR) 3S^ (C*,S*) (4.34)
PAGE 95
88 3S.3S. (M*,C*,S*,X*) = M* ^ ^i'(V^io) ^(^)^(c*,s*) {ifij) (4.35) 5^G ' 3S.aC (M*,C*,S*,X*) = M* L '(S *t ) i^(t ) ^^^^^ \ ^^i ^iO^ '*^^io' 9C {C*,S*) (4.36) 9C^ (M*,C*,S*,X*) = M* 'l'(to) 8A(AR) 'Zl 9C (C*,S*) (4.37) Note the value of tg is not the same in (4.34) , (4.35) , (4.36) , and (4.37) . A sufficient condition for (4.9) to have a proper minimum a (M*,C*,S*) is that the matrix A given belov, is positive definite.
PAGE 96
89 9C^
PAGE 97
90 v*iere the point t*, t* = (t/,tj*)cr5*, has different values for (i=l,2) Proof : For the tvro c cjii ii o dity problem we must have (a) > 0, and this iirplies frcm (4.37) that 3C2 9a (Ar) I 9C (C*,S*) 3'g 9'g (b) 3e 3Si = > 9V, (4.39) 3c3Si > and fron (a) this inplies 3^0 3Si (M*,C*,S*,X*) > (4.40) Hence fron (4.34), since M* > 0, (4.38) f or i = 1 follows. 3^0 a^G a^G (c) 3^G (3S2' 3e 9Si' 3c3Si )'} 3^G { 3^G 3'g 3'g 3'g 3Si3S2 3^ 3S,3S2 3c3Si 3c3Sj 3'g 3'g 3^G d^G 32g + { . 3C3S2 3c3Si 3Si3s2 3c3S2 3Sj^ Upon regrouping terms we obtain } >
PAGE 98
91 d^G a^G 3^0 d^G { ( ) 2 } a^G 32G a^G a^G d^G + 2 . ( ) 9Sj3S2 9C3Si acaSj 3C2 9Si9S2 a^G gJG J Â•( ) > as, 2 acaSj Upon using (a) and (b) we get a^G a^G a^G a^G ^ a^c a^G a^G { Â— Â• ( ) } + 2 . . > aSj* 3c2 3s,2 acas, as,as2 acas, acasj (4.41) New a sufficient condition for aA(AR) > 3C (C*,S*) is that, for any toero*Â» L.'(S.*t. ) < (i=l,2). The proof proceeds as follows: r * = {t IL (S *t ) + L (S *t ) = C* } 'Â— ' 1 1 10 2 2 20 ' In particular if A = (t ,t )er *, Fig^ (4.1), then 10 2 C* = L (S *t ) + L (S *t ) 1110 2 2 20 If we increment C by AC, then for B(t ,t +At ) and D(t +At ,t ) 10 20 20 10 10 20
PAGE 99
92 and Thus C*+AC = L, (Si*tio) + L2(S2* (t2o+At2o)) =L, (S,* (t,o+At,o)) +L2(S2*t2o) AC = L2'(S2*t2o)At2o = L, Â•(S,*tio)At,o IP tc = L2'(S2*t2o) ; ^ L,'(S,*t,o) A*^2 Atjo ^ tc > if L2'(S2*t2o) < n; > if L,'(Si*t,o) < ^^20 At,o And henoe 3A(AR) I >0 9C (C*,S*) if Li'(Sitio) < n (i=l,2) for all toePo*, which proves the first part of the theorem. Using this condition and the fact that 9A(AR) I I > (i=l,2) 9SjL (C*,S*) vre get from (4.35) and (4.36)
PAGE 100
93 g^G a^G < 0; < 0; asjasj (M*,c*,s*,x*) acas, (m*,c*,s*,x*) 9'G acasj (M*,c*,s*,x*) < (4.42) tioTt Fig. (4.1)
PAGE 101
94 Hence, from (4.42) 3*G a^G a^G 2 . . < and (4.41) inplies that a^G a2G a^G a^G 2 { Â— Â• ( ) } > aSj^ ac2 aSj2 acas, Frcm (b) we have a^G a^G a^G ^ . { ) > 8c2 as, 2 acas, Â•Rierefore fron (4.43) it follows that a^G aSj^ (M*,C*,S*,A*) and (4.38) for i=2 is proved. Theorem 4.5: > n (4.43) 0. E. D. If (S,C*) is a solution to the set of equations (4.19) and (4.20), then (S*,C*) is a proper relative minima of gj (S,C) if (i) L.'(S.*t.o) < S Vt = (tio, t,,)ET,* (i=l,2)
PAGE 102
95 (ii) M{C*,x) is a strictly concave fxinction of x, where M(C*,x) is given by (4.21) Proof ; Expanding L( x+AX) t L( xÂ»/yct Â») in a Taylor series with a ranainder R^ (x+^,x;C*) , fibout x, the expression for AM, in E*, as given by (4.31) beocmes 2 2 AM(C*,x) ={ };l. '(xJAx. + ^ L."(x.)Ax.*} i=l 1 1 1 i=l { 2 [ I L.'(x.t.)Ax. R(x,C*) i=l 2 + I L."(x.t.)Ax.2] i^(t) dt} + {C* L(xt'') i=l 2 y L. Â•(x.t.'')Ax.} il;(t'') A(AR(x)) 1=1 + p2(x,xj7^;C*) (4.44) v*iere AR(x) is the incremental set between R(x+^,C*) and R(x,C*), A(AR(x)) is the area of AR(x), and t"eAR(x) . Let to = (tio,t2o)Gro(x). Then from (4.44) we can write 32m{C*,x) , Z_= {L."(x.) + L."(x.t.) li^(t) dt 3x.* ^ ^ iR(x,C*) 1 ^ ^ 1 Â—
PAGE 103
96 3A(AR(x)) + L.'(x.t.J ^(t,) > (i=l,2) 9x. (x,C*) (4.45) 3^M(C*,x) 9x.9x^ 1 J = L.'(x.t.^)^(t,) 9A(AR(x)) 9Xj {x,C*) iT^j (i,j=l,2) (4.46) Now, M(C*,x) is a strictly concave function of x if and only if the Hessian is negative definite. Hence we irust have a^M(c*,x) (a) < 0; (b) ^^ 3'm(C*,x) 3^M(C*,x) W 8x^9x^ a^M(C*,x) 9^M(C*,x) 9x, 9x. 1 2 9x. or, frcm (4.45) for i=l, (a) L,"(x.) + f L,"(x,ti) iMt) dt * ^R(x,C*) > + Li'(x,tio) 'I' (to) 9A(AR(x)) 9xi (x,C*) > (4,47) and 3*M(C*,x) 9^M(C*,x) 92m(C*,x) 2 (b) Â• { } > 9x, 9X2 9x9x2 from vMch
PAGE 104
97 d^mc*,x) 9^M(C*,x) =. =^ > 9xi^ 3x2^ Hence, frcn (a) and (4.45) for i = 2, we have L2"(Xi) + f L2"(X2t2) ^Pit) dt Â•'r(x,C*) + L2'(X2t2d l'(tp) ^^^^ (^^c*) ^ Â° ^^^^^ Set X = S* in (4.47) or (4.48) , then for (i = 1,2) L."(S.*) + 1 1 J L."e*t.) li'(t) dt R(S*,C*) ^ 1 ^ ^ T . /c * 4^ , ^4^ 9A(AR) ' + L.'(S.*t.o) ^(tp) ^3Â— (5^^^*j > 0, t;, = (tio, t2o)ero* merefore, condition (4.38) will hold if M(C*,x) is a strictly concave fvnction of x. Q. E. D. Ifrif ortunately , as stated, the sufficient conditions are not usable in practice.
PAGE 105
98 4.4 Gectnetric Reformulation of the PrcÂ±>lem At this point it is of interest to use the inherent characteristics of the function M(C*,x) in giving gecnetric reformulaticai of the problan for determining the local minima at the point (S*,C*) in E^. If S* and C* satisfy the necessary and sufficient conditions for a proper relative minima, then as we have seen in Section 4.3 determining S* and C* is equivalent to constructing the function M(C*,x) v^ose equation is given by (4.22) with the following properties : (a) for x^er* M(C*,xÂ°) = 9M(C*,x) ' ' = L.'(x ") (i = 1,2) 9x. x=xÂ° 1 1 1 (b) M{C*,x) achieves a maximum value of K at x = S*, i.e., M(C*,S*) = K Graphically, this reformulation is illustrated in Fig. (4.2). M(C*,x) is plotted for various values of the parameter C*. The function
PAGE 106
99 M(C*,x) f a = a(K) Li(Xi) + L2{X2) = C* / / ^/ /^ ' I I I r K 1'/ S* Fig. (4.2). The Function M(C*,x)
PAGE 107
100 a = a(K) is the locus of all points S* at which the function M(C*,x) attains a maximum value of K for various values of C*. 4.5 The Function L(x) Fran new on we shall consider the case when L(x) has the form m L(x) = I L. (X.) i=l ^ ^ where each L^(x^) is continuous and twice dif f erentiable . The function L(x) is the cost charged over a given period of time excluding the ordering cost and in general it is the holding and shortage costs. Consider the case when the holding and shortage costs for each item are linear. Let for item (i = 1, 2, . . . , m) h^ = Unit holding cost at the end of the period p^ = Unit shortage cost at the end of the period Then L(x) is the total ej^jected holding and shortage costs Measured at the end of the period and is given by m 1*0 /Xj^ foo L(x) = ^h ... ... (X t ) (j)(ti,...,t. ,...,t Jdti...dt....dt i=l^'o'o '0^^ ^ ^ ^ " m Â«) ,oo oo (4.49)
PAGE 108
101 If 4). (Â•) and $. (Â•) denote respectively the marginal density fi3ncÂ±ion and marginal distribution function of the darand for itan i (i = 1, 2, . . . , m) in a given period, then from (4.49) we <Â±)tain LW = I i=l hfl^ (^i^i) h^h^ ^^i ^ Pif (v^^i) ^i^^i) *^i (4.50) Setting L. (x.) = h. XX.' X ^ (x.t.) *. (t,) dt. + p, J (t.x.) 4>i(t.) dt. Â•'x. = (h. + p.) fi *. (t.) dt. + p. (Mix. ) (i = l,2,...,m) (4.51) we get from (4.50) L(x) = I L. (X.) i=l ^ ^ L. (x.) denotes the marginal expected holding and shortage costs for item i measxxced at the end of a period. Frcm (4.51) it is clear that the function L. (x.) is twice different iahle at all points for v^di 1). (x.) is ccntinuous. In fact for (i = 1, 2, ,m)
PAGE 109
102 Li'(xi) = (hi+Pi)'I>i(xi) Pi L^"(x^) (h.+p.)(t.^(x.) and L. (x.) is strictly convex for all points x. for which <{> (x.) y 0. Hence , L (x) , being the sum of strictly convex functions , is itself a strictly convex function.
PAGE 110
CHAPTER V OOMPOTATIOqAL ASPECTS OF THE PRCBIiM AN EXAMPLE 5.1 Introduction In this chapter the analysis shall be restricted to the ccnputational aspects of the optimization problem for the special case of a twooomTodity problem v^ere the demand for the items obeys the exponential distribution, and the holding and shortage costs are linear. In Section 5.2 we will present the twoccmnodity problem under consideration and obtain explicit expression for the function L. (x.) , (i = 1,2) , as given by (4.51) , and its first and second derivatives. Ihe purpose of Section 5.3 will be to determine analyticcil e^ressicffis for the set of equations used to determine the optimal policy parameters. The integral equation (4.22) will be converted into a partial differentieil equation of the second order. Riemann method will be used to solve this boundary value problan at the point S. Finally in Section 5.4 a numerical exanple will be considered. Nvmerical methods will be used to solve for Sj*, S2*/ and C*. 103
PAGE 111
104 5 . 2 Exaitple Consider a twocxmrodity inventory control problem v*iere the demand, (Dj), for the items over a sequence of periods (j = 1,2,...), is assured to be independently and identically distributed continvious randcm variables with joint density function ^(t) = Mt) = x^x^e~^i^^~^2t2 The marginal probability distributiai function is given by ^j^i\) = I ^ At) dt = 1 e'^i^^i (i = 1,2) Fran (4.51) we can write L. (X.) = (h. + Pi) I ^ 4>. (t.) dt. + p. (u. X.) = (hi^p.) I" (le^iti)dt.^p.(l,x.) = (h,.p,)(x^.^ eiXi_^),p.^^_j 1 1 On taking the first and second derivatives of (5.1) , we get V (x.) = (h. + p.) e'Vi + h. (i = 1,2) (5.2) r2) (5.1)
PAGE 112
105 I^"(x ) = X^Chi+Pi) e > (i=l,2) (5.3) Using (5.1) in (4.5) we get hj+p, XjXj hj+Pz X2X2 r* = {xxen; h,x, + h^Xj + e + e Xi Xa h, hz = C*+ Â— + Â— } (5.4) X, X2 5.3 Ccciputational Aspects From Theorem 4.2, Si*, S2*, and C* are the solutions for the set of equations given by (4. 19) and (4.20). Using (5.1) and (5.2) in (4.19) and (4.20) we get h. (h.4p.) e ^ ^ + f [hi (hi+Pi) Â•'r(s*,c*) Xi(Si*ti) e ] 4,(t) dt = (i=l,2) (5.5) C* 2 (h.+p.) XiS^* hi y [ Â±Â±e " ^ + h.S.* ^ 1 + [C* i4i Xi Xi Jr(s*,c*) 2 (h.+p.) Xi(Si*ti) I [ ^_i_e +hi(Si*ti) i=l X^ hi _] ,,(t) dt = K (5.6) X^
PAGE 113
106 where '<'(t) = I it) n=l ^^' Fran (5.5) we can write ^iti Or f (hi+Pi)e ^ ^ 4;(t) dt Xj^S^* Â•'r{s*,c*) [h. (hi+Pi) e ] = J (i=l,2) 1 + ip(t) dt ^R(S*,C*) e ii;(t) dt 1 ^R(S*,C*) hi Si* = Â— In [ Â—7 ] (i=l,2) Xi ^(t) dt R{S*,C*) hi+Pi The cdtplexity of findinq an analytical expression for ti^(t) dictates abandonix>g the direct solution of (5.5) and (5.6) for Si*,S2*, and C*. Now frcm Theorem 4.3 the values of S* and C* are the real positive solutions to the set of siimiltaneous equations given by (4.23) and (4.24), i.e. 8M{C*,x) M(C*,S*) = K; I = (i=l,2) ,. ^. 3X^ 3F=S* v*iere M(C*,x) is given by (4.21).
PAGE 114
107 To solve the set of equations given by (5.7) , we shall first find an expression for M(C*,x) and then use this excression to determine Si*, Sj*, and C*. One way to solve for M(C*,x) is to convert the integral equation (4.22) into a partial differential equation of the second order. Theorgn 5.1 ; In E^ if ())(t) = XjXj e (A^ > 0, t^ >_0; i=l,2) and L(t) = Lj(tj) + ^(tj), where L^(t^), i=l,2, is given by (5.1), then M(C*,x) satisfying the integral equation (4.22), i.e. M(C*,X ) = C* L(x) + 1 Â»(xt) M(C*,t) dt (5.B) ''a)(x,C*) is the solution for the boundary value problem 9*M(C*,x) 3M(C*,x) 3M(C*,x) + Xj + Xj = XjXj [C* hjXj hjXj] 9Xj5Xj 3Xj 9Xj and for x^eF* M(C*,x'') = 3M(C*,x) X.x," = L.'(x.Â°) = (h.+p.)e ^^ h. (i=l,2) 9X; X=x'' "1 _ _ (5.9)
PAGE 115
108 Proof ; Before deriving the partial differential equation, we will note that for i,j=l,2 a^ft) = \^{t) (5.10) 3ti 3'<(>(t) atj^atj = XiXj(t.(t) (5.11) The partial differential equation will be derived directly from the integral equation using again the method of varicible increment on AM as in Chapter IV. AM = M(C*,x+Ax) M(r*,x) = C* L(x+Ax) + J (xt) M(t) dt] ''(i)(x,C*) adding and subtracting ({> ( x+Axt) M(t) dt to the right hand ^a)(x,C*) side we get eifter sirrplif ication AM = {L( xfAx) L(x)} + I I*(xÂ±^Slt) (fi (xt) ] M(t) dt ^a)(x,C*)
PAGE 116
109 + ] Â» ( x+Axt) M(t) dt w(x+Ax,C*) f (j) ( xtAxt ) M(t) dt (5 12) ^a)(x,C*) ~ ~ Let Aa)(x,C*) denote the increiTental set between a)(x+Ax,C*) and u)(x,C*), A(Aw(x)) denote the area of Aw(x), and t^tdjuiM. Using the Mean Value Thearem for irultiple integrals, fTheorem 2.2), we can write J ({> ( xHAxt ) M(t) dt 'aj(x+Ax,C*) (j)(x+AxHb) M(t) dt (x)(x,C*) = 4'( x^Axt Â°) MCt") A(Aw(x)) (5.13) Using (5.13) in (5.12) we obtain AM = {L( x+Ax ) L(x)} + f [(t)( xtAxt ) (t'(xt)] M(t) dt ^a){x,C*) + (((( x^Axt ") MCt") A(Aa)(x)) Expanding L( xfAx ) and 4) ( xtAxt Â° ) in a Taylor Series about the point x, the expression for AM can be written as 2 2 AM = { ^ Li'(Xi)Axi + I Li"(x^)Ax.M i=l i=l 2 34)(xt) ] [ I Ajc ^w{x,C*) i=l 9x^ ^
PAGE 117
110 , 2 2 8^(l)(xt) + j I I Ax.AX^ ] M(t) dt + {^(xt") j=l i=l ax^9x^ ' 2 a(f,(xtÂ°) + y AX.} MCt") A(Aa)(x))i+ RMx+Ax,x;C*) i=l 3x. " ~ 1 (5.14) where R ( x+ax ,X;C*) denotes the remainder term of the expansiai. On using (5.2), (5.3), (5.10), and (5.11) in (5.14) we get XiX, XjXj AM={(hj+Pi)e hi}Ax,{(h2+P2)e h2> Ax2 Xj(h,+p,)e Ax,* X2(h2+P2)e Ax^^ + I [ Xj(xtÂ°)Ax. X,<)(xtÂ°)Ax,} . M(t') A(Aw(x)) + R*(x+^,x;C*) (5.15) Figure (5.1) shows gecitietrically the incremental areas due to positive increments in x, and x^. Now fron (5.15) we have for i=l,2 9M(C*,x) X.x. A.xr = (h^+Pi)e ^ ^ hi ^i J 't'(xt; t) M{t) dt ax^ a)(x,C*)
PAGE 118
Ill 9A(Aa)(x)) + <(>(xt ) M(tj,) toeAwfAx^) or e3q>licitly 3M(C*,x) XjX, , = (h,+Pj)e hj X, <}'(xt) M(t) dt 3x. ^u)(x,C*) 8A(Aw(x)) + ((.(Cx^tjg) M{x,,tj,) (5.16) 3x, where, from Fig. (5.1), t^^e [Xj,X2]. 3M(C*,x) XjXj , = (hj+P2)e ^2 ^2 Â» (xt) M(t) dt 3x, u)(x,C*) 9A(Aa)(x)) + (t.(x,t,5,0) M(t,o,Xj) (5.17) Ax, 2 Â•vtjere, from Fig. (5.1), t,je[xj,x,]. 3*M(C*,x) = XjXj (f> (xt) M(t) dt 3Xj3Xj w(x,C*) 3A(Aa)(x)) Xj(}>(0,Xjt2 5) M(Xi,t2j) 3x J 3A(^)(x)) X\()(Xit,5,0) M(tjo,X2)
PAGE 119
112 Â•tÂ»^J Â•P X Cm
PAGE 120
113 + ,^(0,0) M(x,,x^) ^5.^3j v^ere f ran Fig. (5.1), t^^t[xÂ„x^] and t^e[x^,x^]. Using (5.16) and (5.17) and the explicit expression for M(x) = l(0,X2t2 5) M(x,,tjo) 3x, 3A(Aa)(x)) Xi
PAGE 121
U4 A2 "2 3M(C*,x) 9M(C*,x) = ^2 Â—^^ ^1 Â—^ + A2X2[C* hiXj h2X2] Hence the partial differential equation given in (5.9) follows. Since for xÂ°er* the set w(xÂ°) is an enpty set, we obtain from (5.8) M(C*,xO) = C* L(xO) = and 3M(C ,x) dx. 1 = L.' (x.O) (i = 1,2) On substitxiting for L.' (x.Â° ) as given by (5.2) the boundary conditions follav. Q. E. D. Theorem 5.2 ; The solution of the boundary value problem (5.9) at the point S is given by M(C*,S) = f Q^l(Six^)X^{S2X2) Iq[2AiX2(Si Xj) (S2 X2) ] [(hi + pi) e"^i^i hi] dxi
PAGE 122
115 + ^^2 J [C* h,x, hjXj] . a)(C*,S) X,(SiXj) XjfSaXj) e l^(2/X,X2(Six,) (S2X2)] dxidxj (5.19) v^ere 1^[ ] is the rxriified Bessel function of the first kind of zero arder. Proof ; In solving the boundary prc^lem, which is a linear hyperbolic particil differential equation of the second order, we will use a method due to RiÂ€iTannÂ» Miller^. The value of the function at the point S is given by { f "(S,x) .,x, ^r* 3M(C*,x) M(C*,S) Â«= ( I U(S,x) dxi XjXj ^r* 3x, + I X,Xj[C* h,x, hjXjl U{S,x) dx ) (5. 20) ^w(C*,S) where U(S,x) is the Green function for the problem satisfying the following boundary problem 9x,8x2 9xj 3x2 9U = X U v/hen x^ = S^ ax, 'See additional references. (5.21)
PAGE 123
9U_ 9X2 = X2U v*ien Xj = S] 116 (5.21) U = A1A2 when Xj = Si, X2 = S2 A solution for (5.21) is given by (see [25]) U(x,S) = A1A2 er^i(SiXi)A2(S2X2) ip[2 AiA2(Si Xi)(S2 X2) ] (5.22) using in (5.20) expression (5.22) and the value of M (C*,x) 1 xer' i = 1,2 as given in (5.9) , we obtain (5.19) Q. E. D. 5.4 Numerical Exanple In this nxjmerical exanple, we shall make use of the parameter values considered by Sivazlian [13] . Let us caisider the case when Xi = X2 = 1 Pi = P2 = 20 hi = h2 = 1
PAGE 124
117 K = 5.0 Using these values in (5.1), (5.2), (5.4), and (5.19) we get ^i L (x ) = +21e + Xi 1 (i=l,2) i i ~^i Li'(Xi) = 21e + 1 (i=l,2) X, ^j r* = {xXj + 21e + Xj + 21e = C* + 2 } and at the point S = (Sj,S2) (SjXj) (SjXj) M(C*,S) = e I,[2/(S,k,)(Sj3
PAGE 125
118 Lj (Xj) + ^^(Xj) = C* for the case v^ien C* = 8 w(C*,S) Xj X2 r* = {xjxi + Xj + 21e + 21e = 10} 2 3 S^ x^ Fig. (5.2)
PAGE 126
119 Numerical Analysis : In carrying out the numerical analysis for this particular exairple, the following points were taken into consideration: (i) at optimality Sj* = S2* (ii) for a given value of C , T is synmetric about the line Xj = X2 (iii) the resiiLts obtained by Sivazlian [13], table 1, for this particular exaiiple under different (a,S) policies (iv) Sivazlian [13j shows that ^ p e"^"l"'"2) io[2,^I^] dui du2 = a{l [e~^^ Io(2a) + e~^^,(2a)]} (a > 0) (v) Sivazlian [13] further shows that f f u.e"^^i'*"^2)ij2/;iru7] dui du2 = alLi^al ^ _ j^2a ^^^^a) + e'^^ Ii(2a)]} (i = 1,2) (vi) for this particular exanple, M(C*,x), as given in (5.8) is a nondecreasing function of C*.
PAGE 127
120 TaJdng the above points into consideration, we initiated the search procedure by choosing C* = 8.2 and Sj = Sj = 3.8. In ccmputing the double integral in (5.23) for a particular value of C*, Sj and S2 , the region ti
PAGE 128
121 x,/
PAGE 129
122 which is defined over the closed subr^ions A2, A3 and Ai^ is formulated by cutting each subregion into very anall squares each of area AA. , (i = 2,3,4) , where any two of these squares have at nost one side in acrman. Henoe for (i = 2,3,4) ) (C* xi X2) e<^i^i)^S2X2)i^^2 As,xi)(S,x,) 3 dx Ai = lim f (c* xi X2 ) e"^^i^i.^(S2X2.) AA^>0 j=l j j ^3 Io[2.^(Sl Xi ) (S2 X2 )] AA. 3 j ^ v^iere (xi .,X2.) is any point in A. and n. is the maxiimm nxjtiber of squares of area AA. in A. . By syninetry, the integral over Pi^ is equal to the integral over A2. (3) The line integral of the function ^(S,Xi)(S2X2)j^^^^ (g^ x,)(S2X2) ] . {21e^l} viiich is defined over the smooth arc T* is formulated by dividing the arc T* into N arcs by inserting (N 1) points Q^(xii,x2i) . Thus, J (Six,)(S2X2),^^,^ ^^^ xi) {S2X2) ]{21e^ll]} dxi
PAGE 130
123 N^ i=l AxiiM) Ax2iK) Numerical Results : To obtain nunerical resiiLts, a caTpatational algorithm in Fortran IV (Appendix 5. A) was written to solve for the cptimal policy variables. The algorithm proceeds in the following steps: Step Oxose initial values for C and the increment in C*,AC ; Choose initial values for S, S_= Sq, and an increment in S, AS . The range of vali>es of S , Sq f. S <^ SÂ° , for a given value of C*, are chosen to assure that M(C*,S) achieves a maxinon in that range; Choose M as an x:?3per liirat on K M(C*,S) . Step 1 Carpute M{C*,S) . Step 2 Check the difference between M(C*,S) and K, then a. If the difference is greater than AK, set AC* = ^ , C* = C* AC*, S = S and go to Step 1.
PAGE 131
124 b. If the difference in absolute value is greater than AK, go to Step 3. c. If this is the first tirns the difference in absolute value is less than AK; set K* = M(C*,S) , S* = S, Cj = C* and go to Step 3. Otherwise, go to Step 3. Step 3 Set S = S + AS. If the new value of S is in the range of values of S, go back to Step 1. Otherwise, if K M(C*,S) > AK over all the range of values of S, set AP* AC* = 1, C* = C* + AC* , S = S snd go back to Step 1 ; else the algorithm ccaiverges on the cptiitial values of C* and S*. For the particular exanple we are considering, v^en K = 5.0, we initiated the iterative search procedure as follcws: The values of C* were chosen over the range of values [8.22, 8.27] in increment of 0.01; Sj = S2 values ranged over the closed interval [3.8, 4.08] in increment of 0.02. The number of "steps" used in the numerical integration N^ were selected to be 60 and 100. For the case N^ = 100 we have C* = 8.2625 and S* = S* = 4.02 M(C*, S*) = 5.00694 and 8M(C*,S) = for (i = 1,2) S = S* 9S. 1 which corpare favorably with the results obtained by Sivazlian [13] when follcvdng different optimal policies.
PAGE 132
125 Grafiiically, the cjoiputed values (J^pendix 5. A) of the function M(C*,S) are plotted in Figs. (5.4) and (5.5) for the cases when Nj = 60 and 100 respectively. As an e3q)eriinental observation, it may be noted that as the number of "steps" used in carrying out the nunerical integration increases the curves beoane siroother. The accuracy of the ocnpubed results is subject to two sources of error, truncation and roundoff. In coiputing the modified Bessel function Iq (x) / only the first ten terms of the convergent series were sunned. Hcwever it was found experimentally that the rouncJoff error is the factor of significance and it is not easy to examine it analytically, since it is a function of the oorputer used and the size of the "steps" used in carrying out the nunerical integration.
PAGE 133
126 Nj= 60 5.0M(C*,S) 4.94.BC*=8.?5 C*=8.24 C*=8.?.3 C*=^.22 3.76 3.7^3 Fig. (5.4)
PAGE 134
127 Ni= 100 / 5.0
PAGE 135
CHAPTER VI CXX^CLUSICWS AND RECC*ME1Â«1DATI0NS 6.1 Conclusions A stationary multiconnodity inventory problan with periodic review has been formulated and studied vs^en a dyadic (o,S) policy is used. At the beginning of each of a sequence of periods of time the stock level of each item is reviewed and a decision to order or not to order is made. Ihe cost elanents that affected the ordering deciin each period are: the ordering oosts, the holding cost and the shortage cost. The ordering oosts are assured to be ccttposed of a purchasing cost which is proportional to the quantity of each item ordered and a single set Lp cost which is independent of the quantity ordered. The holding and the shortage costs are assured to be charged on the basis of the stock levels at the end of the period. Demand, for the items, in each period of time is described by a continuous randan vector, with a joint density function, independently distributed from period to period. Immediate delivery of orders and corplete backlogging of all unfilled demands are assumed. The inventory syston just described was treated by many researchers under different optimal (ct,S) policies. This research, as catpared with what has been done in this area, is more general in the sense no assunptions were made about the configuration of the ordering region and no specific joint density functions were considered. 128
PAGE 136
129 The mdimensional convolution cperation v^ch was introduced in Chapter II and used in the study is a generalized concept of the ordinary convolution operation. For the case v^^en T , the hypersurface that subdivides the ordering region and the notordering region, is admissible, the goonvolution of any function red\x:es to the ordinary convolution and all the goonvolution operation properties with the carmutative law hold. The gconvolution prcperties, in particular, were used in the solution of an mdimensional integral equation of the renewal type, which is similar in form to the equation solved by other researchers in this area but with no restrictions on the configuration of r. Ttie analysis used in deriving the analytical expression for the stationary level expected oost per period is similar to Sivazlian's [11] approach for the one oormodity problan when operating under a stationary policy of the (s,S) type. The asynptotic results for the problem were deduced by appealing to the results of Yosida and Kakutan [22] in the theory of linear operations. At optimality it was shown that the set T* that subdivides the policy space, Q, is given by r* = {xx e Q; L(x) C* = 0} v*iere C* is the minimum value of the stationciry total ejqsected cost per period excluding the variable cost, and L(x) is the conditional
PAGE 137
130 expected holding and shortage oost function. By using this result, the minimization problesn was reduced to finding the decision variables Sj* (i = 1,2, ,m) and C* that minimize the stationary total expected oost per period. S* and C* were given as the real positive solutions to the set of simultaneous equations M(C*,S) = K (6.1) 3M(C ,x) 9x. 1 ^ = (i = 1,2,. ..,m) (6.2) x=S vMch are siitdlar in fonn to the set of equations used by Sivazlian [11] in determining the optimal values s* and Q = S* s* for the one cciitnodity problem. M(C*,x) satisfies the integral equation M(C*,x) = C* L(x) + M(C*,X2t) (J)(t) dt (6.3) R(x,C*) in which (p is the joint density function of the demand. For the case of a twooorrmodity problem viiere the demand for the items obeyed the exponential distribution and the holding and shortage costs were linear, it was feasible to convert the integral equation (6.3) into a hyperbolic partial differential equation of the second order with boundary conditions. Analytical solution of the boundary value problan was then determined and used in (6.1) and (6.2) to determine Sj*, S2* and C .
PAGE 138
131 6.2 Extension ar>d Recxjimendations Several interesting directions for future research and develc^xnent are suggested as a result of this research. The ccnputational aspects of the prcÂ±)lan need further investigation. The necessary and sufficient conditions, established in this study, for the existence of a relative minima are valid for any continuous density function and any conditional expected holding and shortage cost function v\iiich is twice differentiable. In the case of a twooomiodity problem, v^ere the daiiand for the itans obeys the ej{ponential distribution and the holding and shortage costs are linear, it has been feasible to convert the integral equation into a partial differential equation of the second order with boundary conditions. However, what about the case when the dariands for the two items are given by, say, a gaitma distribution? Is it then feasible to convert the integral equation into a partial differential equation? And, if so, vAiat ccnputational procedure must be followed to solve for the policy parameters? Another qiiesticn also arises, how to solve for the optimal policy parameters for the twooontnodity prcfclon, when the demand for the items are exponentially distributed and the holding and shortage costs are not linear but, say, quadratic. As we have seen the determination of the cptimum policy parameters is a ccrplex task even in the case of a sinple demand density function.
PAGE 139
132 As an alternative to approximating these parameters by nutverical means, the approach used by Roberts [3] could be follaved to give analytical formulas for determining the asynptotic reordering region for the case when the holding and shortage costs are linear. In order to derive the asymptotic expressions, for the set of equations that is used to determine the policy parameters, an approximation for the renewal function i){t) dt and similar extensions of the type discussed Ir(s,c) Â— in Smith [14] are required. The work of Yahov and Bickel [21] and FarreH [4] in investigating the asymptotic behavior of the renaval function in two and higher dimensions must be noted. However, the challenging prcÂ±)lem is to find an analogous theorem to Smith's theorem involving the renewal density function in two and higher dimensions. As noted in Chapter I, under the adopted (a,S) policy, the sequence of stock levels at the beginning of each period forms a discrete Markov process. Greenberg's [5] approach could be followed to determine the transient distributions of the stock levels prior to ordering. For the stationary distributions Karlin [2] approach oould be followed. The results of this research, especially the properties of the gconvolution operation, are hoped to be of great use in determining the transient and stationary distributions of the stock levels prior to making ordering decision. T3iis study treated "the Patient Cus toner Case" v^iere all unfilled
PAGE 140
133 demands were hacklogged. For the lost case, the analysis can proceed along the same procedures vised in Chapters III and IV.
PAGE 141
GLOSSAFOr CF NOTATIGN m the nunber of cxrmodities cxnsidered E mdimensicxial (m >_ 1) Euclidianspace D a random vector representing the demand D. a randan variable representing the deinand of oonmodity i * the pnÂ±)abilitY distribution of D *. the probability distribution of D. (ji a density fxmction of D ()>. a density function of D. y the mean vector of D $ the nfold ordinary convolution of $ $ . the nfold ordinary convoluticxi of $ . .(n) '(n) the nfold ordinary convolution of ^ with itself the nfold generalized convoluticn of 4) with itself \\) i{;(t) = J 4), V (t) , the renewal density function j^i W a the ordering region a the not ordering region r the set of points that sepcirates a and a S_ the point up to which we order S. the i^ conponent of the vector S_ K the set up cost C. the unit purchase cost of product i h. the holding cost per unit of product i on hand at the end of a period 134
PAGE 142
135 p. the shortage cx>st per unit of ooriTodity i on hand at the end ^ of a period L(x) the conditional e^^jected shortage and holding cost measured at the end of a period g(a,S) the stationary total ej^jected cost per period I (x) the modified Bessel function of order n n * starred syrobols denote conditions at optimality C* the value of the total expected C30st per period excluding the variable ordering cost
PAGE 143
APPENDIX 5A OCMPUTED OLTTPUT RESULTS AND PIOGRAM LIST
PAGE 144
Algorithm Flowchart 137 Choose initial values for C* and Ac* Choose a range of values for S, SefSofS"] , so that for a given value of C*, M(C*,S) achieves a inaxiinum in that range. Choose an initial value for AS ChooseAK as an upper limit on KM(C*,S) Conpute M(C*,S) Compute AM=M(C*,S)K Set AC*=AC*/2 C*=C*AC* S=So YES Set S=S+AS YFS Â— KAM>AK am>ak" NO S>S NO YES YES Set K*=M{C*,S) S*=S Ci=C* ^J0 STOP The algorithn _v converges to "^ S* and C*=Ci
PAGE 145
138 > tO SD (M LU " I/) tÂ— O t/) Â•Â• 3 LU LU .< Q. >Â— Â•Â• I/' 5! 13 fi: Â• o a ui cj Lj 2: i>^ Â— a n OO (_) Â»Â— .. U "Â• o Q UJ O X Â•LU a: Q. O u. q: "Â— Â• a: LU Â•' yu m cc a:. Q. UJ o oo > Oj ij f\j Â• ^ O LU t/1 > nf < LL C O n UL ;' o I/) I/) o LU UJ rD z) > o o Â• o cc lU C 5: UJ ^ o UJ ^ X a. z tÂ»i IL) O Â•' > O uLL q: a O yUJ LU o o < ~ z :^ o < < oo ro ot (a: Â•' Â— r r^ >O t(^ q: a: UI LU (z z o < < Â— Â•fM Â— 2' 3: Cll oJ uo o O (/) LU LU ro >Â— rr > > ^ Â— LO < Â— < q: o 1/1 lu Â•Â• o o ^ C< O^ vO O LO 2 jlululuooluqc^ _j X ^ a: O X 00 (Â— ii^(Niia (NJ f< O II Â• *Â—Â» o ^ II Â— Â•4cc X hfM 00 * ca Â•& * 00 og Â— . fvj rÂ«Â» IÂ— cr (\i Â•o c ~. ^ r>o: r^ 00 100 II o ^~ r. Â— o a: tQ IÂ— Â•Â— a. VQC w o o hÂ»hOD Â» oooooooooouo
PAGE 146
139 o o CM X o >
PAGE 147
140 o o X o X
PAGE 148
141
PAGE 149
142
PAGE 150
143 NUMBER OF SUBDIVISIONS Nl = 100 OPTIMU*^ VALUES GF C AND S CO = 0.826399993896480 01 ITERATION NO. = 4 M(S,C) S1=S2 C 0.500069A9910199D 01 .40 19996643G664D 01 0. 82624999401 1 60D 01
PAGE 151
144 ITERATION NO. = 1 0.^99298689329910 01 . 39 199991226I96E 01 0.826399993896^80 01 ITERATION NO. = I 0.50025237931040D 01 . 393999B6267090E 01 0. 826399993896A80 01 ITERATION NO. = 2 0.^98617315300050 01 . 39 19999 1226 196E 01 . 826l9999A0
PAGE 152
145 ITERATION NO. = 3 0.30018918310^950 01 D . 39999971389771E 01 0. 82629999397323D 01 ITERATION NO. = 3 0.5001^8781866140 01 . 40 199966A30664E 01 0.8262 99993973230 01 ITERATION NO. = 4 0.498800582182030 01 . 39 199991226196E 01 0. 82624999401 1600 01 ITERATION NO. = 4 0^.499763824692570 31 . 39 399906267090E 01 0. 82624999401 1600 01 ITERATION NO. = 4 0.499769408400690 31 3 . 39 59993 I307983E 01 0.826249994011600 01 ITERATION NO. = 4 0.500003905046990 01 . 39 799976348877E 01 0.826249094011600 01 ITERATION NO. = 4 0.499904571616230 31 . 3999997138 9771 E 01 0.826249994011600 01 ITERATION NO. = 4 0.5aCV069499101990 31 .40 199966430664E 01 0.826249994011600 01 ITERATION NC. = 4 0.499933712128080 01 .40 39996 147 1 558E 01 0.826249994011600 01 ITERATION NO. = 4 0.499334954327350 31 .40 59995651 245 IE 01 0. 82624999401 1600 01 ITERATION NO. = 4 0.499084532973770 01 .4079995 1553345E 01 0. 82624999401I60D 01 ITERATION NO. = 4 0.49828455150016D 01 .40999946594238E 01 0.826249994011600 01
PAGE 153
BIBLIOGRAPHY 1. Arrow, K. J., T. Harris and Y. Maichak, "Optimal Inventory Policy," EconoT^trica , IPSl, 19, 250272. 2. Arrow, K. J., S. Karlin and H. F. Scarf, Studies in the r^thematical Theory of Inventory ctnd ^oduction , Stanford University Press, Stanford, California, 195R. 3. ArrcM, K. J., S. Karlin and H. Scarf (eds.). Studies in Applied Probability and Â•lanaagnent Science , Stanford University Press, Stanford, California, 1062. 4. Farrell, R. H. , "Limit Theorems for Stoooed Random Walk III," Ann. Math. Statist ., 1966, 15101527. 5. Greenberg, H. , "Stock Level Distribution for (s,S) Inventory Problems," Naval Research Logistics Quarterly , 1964, 11. 6. Howard, R. A. , Dynamic PrograiminQ and Markov Process , The MIT Press, Cambridge, Massachusetts, 196n. 7. Iglehart, d., "Dynamic Programning and Stationary Analysis of Inventory Problems , " Chapter 1 in Multistage Inventory ^todels and Techniques , H. Scarf, D. Gilford, and M. Shelly (eds.), Stanford University Press, Stanford, California, 1963. 8. Johnson, K. J., "Optimality and Computation of (a,S) Policies in the MultiItem Infinite Horizon Inventory Problem," Management Science , 1967, JJ, No. 7, 475491. 9. Scarf, H. , "A Surve^' of ;\nalytic TeKrhniques in Inventory Theory," Chaoter 7 in Multistage Inventory 'Videls and Techniques , H. Scarf, D. Gilford, and M. Shelly (eds.), Stanford University Press, Stanford, California, 1963. 10. Sivazlian, P. D., "Inventory Control of a MultiProduct Systan \^tb Interacting Procurenent Cost," vrorking paper No. 58, Operations Research GrouD, Case Institute of Technology, Cleveland, Ohio, 1966. 11. Sivazlian, B. D., "Optimality and Ccmputation of the Stationary (s,S) Inventory Control Problem," Technical Report No. 7, Project Themis, Dept. of Ind. & Sys. Enor., the Univ. of Fla. , May, 1968. 146
PAGE 154
147 12. Sivazlian, B. D. , "Asyrrptotic J^roximations to Ordering Policies in a Stationary (a,S) Inventory Problem," Siam J. P^ppl. Math. , 1970, 19, No. 1, 155166. 13. Sivazlian, B. D. , "Staticaiary Analysis of a Multicarmodity Inventory System with Inter acÂ±ing SetUp Cost," Siam J. i^pl. Math. , 1971, 20, No. 2, 264278. 14. Smith, W. , "Renewal Theory and Its Ranifications , " Royal Statistics Society , Ser. B. , 20, 243302. 15. Soland, R. M. , "On Sane Renewal Processes in One and Two Dimensions," Ph.D. Thesis, MIT, 1964. 16. Veinott, A. F. and H. M. Wagner, "Ccnputing Optimal (s,S) Inventory Policies," Management Science , 1965, 11, No. 5, 525552. 17. Veinott, A. F. , "Optijnal Policy for a MultiProduct, Dynamic Nonstationary Inventory Problesn," Management Science , 1965, 12, No. 3, 206222. 18. Veinott, A. F. , "The Status of Mathematical Inventory Theory," Management Science, 1966, 12, No. 11, 745777. 19. Watson, G. N., A Treatise of the Bessel Functions , Cambridge, England, University Press, 1952. 20. Wieeler, A., "MultiProduct Inventory Models with SetUp," Ph.D. Thesis, Stanford University, 1968. 21. Yahav, J., and Bickel, P., "Renewal Theory in the Plan," Ann. Math. Statist. , 1965, 36, 946955. 22. Yosida, K. , and Kakutan, S. , "Operator Theoretical Treatment of Markoff Processes and Mean Ergodic Theorem," Ann. Math. Statist., 1941, 42, 188226. Additional Pteferences Miller, K. S. , Partial Differential Equations in Engineering Problgts , PrentaoeHall, Englewood Cliffs, N. J., 1953. Olmsted, J. M. , Advanced Calculus , ;^letcinCentuxyCrof ts , New York, 1959.
PAGE 155
148 Sivazlian, B. D. , "Dimensional and Ccrputational Analysis in Stationary (s,S) Inventory Problanns with Gairma Distributed Demand," Management Scienoe , 1971, 1]_, No. 6, 307311. Sivazlian, B. D. , and P. E. Valisalo, "The Utilization of a Hybrid Cdtputer Systan in Solving for Optimal Policy Variables in Inventory Problans," Proceedings Ccnferenoe on H^rid Oatputation , 1970, 6266. Wacker, W. D. , "Approxinations to Optimal Policies in a Stationary Multioonmodity Inventory Model with SetUp Cost," Ph.D. Thesis, Washington University, 1971. Vhite, D. J., Dynamic Progranming , Oliver & Boyd, Edinbur^ & London, 1967.
PAGE 156
BIOGRAPHICAL SKETCH Nabil S. Faour was bom in Palestine on December 25, 1940. In Septennber, 1947, he became a refugee. He attended schools in Beirut, Lebanon and graduated from high school in June, 1958. Mr. Faour 's undergraduate education was taken at the Ainerican University of Beirut, where in June, 1962, he received a Bachelor of Mathannatics degree. Between Septerijer, 1962 and July, 1965, Mr. Faour worked with the Arabian Anerican Oil Conpany, Dhahran, Saudi Arabia, in a senior position as an Electronic Systenvs Analyst. The author entered the Industrial and Systems Engineering Department of the University of Florida in S^tertoer, 1967. In Augxist, 1968, he received a ^faster of Operation Research degree. In July, 1970, he joined the American Can Conpany, Greenwidi, Connecticut, and he is presently an Associate Industrial Engineer. Mr. Faour is married cind hcis a son. 149
PAGE 157
I certify that I have read this study and that in my c^inion it conforne to acceptable standards of scholarly presentation and is fully adequate, in scope and qxiality, as a dissertation for the degree of Doctor of Philosophy. / / /,/ / Boghgs D. Sivazlian, Cttainnan Assoc. Prof, of Ind. and Sys. Eng. I certify that I have read this study and that in my opinion It conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree Of Doctor of Philosophy. F. Bums 3SOC. Prof, of Ind. and Sys. Eng. I certify that I have read this study and that in my opinion It conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosophy. Robert G. Blake Assoc. Prof, of Math. I certify that I have read this study and that in my opinion It conforms to acceptable standards of scholarly presentation and is fully adequate, ui scope and quality, as a dissertation for the degree of Doctor of Philosophy. , Â— Pop stoja^ovic/ . Prof, of Math. Zoran Assoc This dissertation was submitted to the Dean of the College of Engineering and to the Graduate Council, and was accepted as partial fulfillment of the requirements for the degree of Doctor of Philosophy March, 1972 ian. College of Engineering Dean, Graduate School
PAGE 158
Â«'! ,ii^^
xml version 1.0 encoding UTF8
REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchemainstance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd
INGEST IEID E6V4I3MGI_OCXERC INGEST_TIME 20170714T21:49:31Z PACKAGE UF00097610_00001
AGREEMENT_INFO ACCOUNT UF PROJECT UFDC
FILES
