You are on page 1of 9

KNOWLEDGE-BASED OPTIMIZATION OF A STOCHASTIC INVENTORY PROBLEM USING DYNAMIC PROGRAMMING Sayed M. Irfan1 and Professor Dr.

Nazim Qureshi2 1. Department of Industrial & Manufacturing Engineering, NED University of Engineering & Technology, Karachi 75270, Pakistan. 2. Department of Industrial & Manufacturing Engineering, NED University of Engineering & Technology, Karachi 75270, Pakistan. ABSTRACT This paper presents the development of a computer-based decision-support system for optimal lot-sizing of finished goods over a multi-stage planning horizon. The product demand across the planning horizon is variable and symbolic in nature. Methods commonly employed for the lot-sizing problem (deterministic demand) are heuristics that produce near-optimal solutions. However, an optimal solution is guaranteed by the Wagner-Whitin method that can lend itself to be solved by dynamic programming. A key feature of this decision-support system is to permit incorporation of feedback information into the decision-making process by exploiting the principle of optimality of dynamic programming, which identifies optimal policies for each and every state in the problem, regardless of how that state was achieved. The demand situation considered here is a company contemplating a price reduction decision and supporting it with a promotion action. A set of rules of the type If X is low, then Y is low, for example, involving several interrelated variables that influence demand have been used. To incorporate this reasoning, a fuzzy logic approach has been employed. KEYWORDS: (Managerial decision support, dynamic programming, fuzzy logic). 1. INTRODUCTION

Inventory costs may represent a large proportion of capital costs and inventory decisions can have significant implications on a companys financial health and market credibility. In this context managers have to deal with several decision-making situations of different nature and scope and there may be the emergent need for organized, calibrated and interactive approaches to identify optimal or near-optimal strategies. One such decision problem is the lot-sizing problem where a production plan has to be devised across a planning horizon comprising distinct time-periods. There is a discrete demand for each period which is variable and a decision is required on the number of units to manufacture in each period. The underlying objective is to minimize aggregate inventory costs where the two principal cost elements, viz., holding costs and setup costs are conflicting in nature. These costs may also vary from one period to another. An element of complexity may further creep into the problem if the demand happens to be symbolic in nature. Traditionally, for a strictly deterministic demand, there are essentially three approaches to dealing with the case of a time-varying demand pattern: i. Use of the basic economic order quantity with appropriate adjustments. ii. Use of an approximate or heuristic method.

iii. Use of exact best solution to a particular mathematical model of the situation; Under a specific set of assumptions, this approach, known as the Wagner-Whitin algorithm, minimizes the total of certain costs. The latter two approaches have been the focus of much research previously. The WagnerWhitin method, in particular, has been considerably debated since its publication in 1958 [1] and, frequently, arguments of computational burden and complexity have been raised against it [3]. The development of computer technology has tended to mute the high computational burden argument but the complexity argument persists. Conclusive arguments in its favor have recently been advocated and the method is likely to be adopted more widely in the future [1]. There are important aspects of the lot-sizing problem, however, that the Wagner-Whitin method does not address, if the problem is seen in a broader context. This necessitates a more intensive analysis. The lot sizing problem is a sequential decision problem, and there may always be the occasion for marginal changes down the planning horizon due to the typically dynamic nature of the decision environment. Therefore, optimal policies may have to be revised if such changes are to be incorporated. Past or current practices do not address this potentially critical aspect of the decision-making process. The Wagner-Whitin method operates strictly on the basis of a deterministic demand. The practical context of the lot-sizing problem, however, may include a situation where artificial seasonality can be induced by pricing or promotion actions [6], and this is a situation where a judgmental input will clearly be required either due to the absence of reliable historical data or because of the urgency to gain a competitive advantage. If the demand were to be perceived in terms of symbolic information, for example, in terms of subjective categories like low, pretty low, medium, very high, high, etc., then the problem transcends into the domain of artificial intelligence. An AI language like LISP will allow direct incorporation of symbolic information and logical rule bases which will allow expert or knowledge-based systems to be developed, that include optimization capabilities [5]. However, difficulties in capturing the full essence of the demand forecast problem into a rule-based procedure with a capability for sensitivity analysis appear to be enormous. The objective of this research paper, therefore, is to propose the development of an interactive computer-based decision support system for optimal lot sizing that allows for the incorporation of an information-feedback mechanism and provides a user-friendly environment for resolving the symbolic nature of demand that can arise in specific situations. An implicit accompanying objective is to emphasize the opportunity that may be available within an organization to facilitate and promote a cross-functional approach in making critical managerial decisions. In the particular context of the lot-sizing problem this may call for integration within the production and marketing functions.
2.

THE ALGORITHMIC STRUCTURE OF THE LOT-SIZING PROBLEM

The algorithmic structure for this type of an inventory planning problem can be closely associated with the stage-state concept typical of many sequential decision problems where the optimal policy needs to be identified right across the planning horizon. The stage in a system corresponds to the particular time period , while the state of the system at the beginning of a period can be described just by the stock level at that time. The number of different stock levels

there may be at the beginning of a period can be significantly restricted by imposing two important conditions1 introduced by Wagner and Whitin. These are: (1) Batches should only be produced in periods, which start with zero inventories. (2) Batches should consist of an exact number of periods demand. The number of exact requirements policies is given by 2n-1, where n is the number of periods. Since the number of states and options at each stage is relatively limited we can think of the problem as that of finding a least-cost path on a directed network. The cost of moving from one state to another is given by the production and set-up costs incurred at that stage, plus any holding cost. The whole concept can be Table 1 illustrated by considering a simple Demand data for the Wagner-Whitin method four-period example, data for which is given in Table 1. The Period 1 2 3 4 holding cost is Rs.1 per unit per Demand 10 20 25 15 period, the set-up cost is Rs.25 and the cost of producing one unit is Rs.1, none of these quantities vary with time. We wish to find the optimal production plan. The suggested network for the problem can be developed using the arguments presented above. This is shown in Figure 1 overleaf. Nodes represent the states while the arcs represent the decisions, i.e., the number of units manufactured. Numbers in parenthesis indicate the corresponding costs of these decisions. For a larger number of periods, however, drawing the network becomes increasingly tedious and somewhat complex. Beginning with an initial inventory of zero (or any other integer value) and depending upon the quantity manufactured, the problem transforms to a particular state in the next stage where the state is defined by the difference between the sum of the above two quantities and the demand for the current stage. In moving from one state in the current stage to another state in the succeeding stage, there is a cost involved. This is defined by the sum of the production cost and set-up cost, and the holding cost if a quantity greater than that required for the current stage is manufactured. Since more than one feasible decision alternatives may be involved, the optimum cost for that state should be the minimum of the respective decision costs. But such a cost will be influenced by similar costs further down since an overall optimal policy across the planning horizon is being visualized. Thus it will be necessary to evaluate all possible decision alternatives for each and every state and thus identify the optimal policy for every state before an optimal policy could be identified across the planning horizon. 3. THE SOLUTION METHODOLOGY

The solution to this problem is approached through dynamic programming (DP), an operations research technique. In general, DP attempts to reduce a single problem involving a relatively large number of decision variables into a series of smaller problems, each with a small number of decision variables. At the core of DP is a recursive relationship that identifies the optimal policy for stage (n), given the optimal policy for stage (n + 1) is available. When we use this recursive relationship, the solution procedure starts at the end and moves backward stage by stage each time finding the optimal policy for that stage until it finds the optimal policy starting at the initial stage.
1

At a later stage in this paper the first condition will have to be relaxed to address the lot-sizing problem in its broader context.

The precise form of recursive relationship differs somewhat among dynamic programming problems and it has to be developed for specific applications. Section 3.1 elaborates on this procedure. Stock position at start of period
10 (35)

Cost during this period


25 (50) 15 (40)

0
30 (75)

20 (45) 45 (95) 60 (125) 0 (80)

0
40 (80) 0 (40)

0 (0)

55 (125)

20
0 (25)

25
0 (15)

15

70 (155)

45

40

0 (40)

60 Period Deman d 1 10 2 20 3 25 4 15

Figure1: Network representation of Example of Table 1. 3.1 DP Formulation of the Problem The first step in the development of a computer-based system would be to identify all the feasible states. We may begin by observing (a) The total number of stages in the problem are N = 4 where the current stage is denoted by n. (b) The possible states in each stage are denoted by sn and these correspond to the possible beginning inventory levels at each stage. (c) The decision xn corresponds to the number of units manufactured in one batch. Batches should consist of an exact number of periods demand. The possible states, then, at Stage 2, for example, may be given as shown in Table 2. Table 2 Batch size and beginning inventory levels Stage 1 Stage 2 Batch size States (Cumulative demand) (Beginning inventory levels) s n = x n 1 d n 1 x1 10 0 30 20 55 45

70 60 A recursive relationship that identifies the optimal policy for stage n, given the optimal policy for stage n+1 is now required to be established. Since this involves a minimum cost decision, an appropriate cost function needs to be developed. The elemental costs associated with each decision are pn = sn = hn = production cost per unit set up cost (assumed to be independent of the number of units manufactured) holding cost per unit per time period

The cost function may then be expressed as:


c sxn = p x x n + u n + hn * ( s n + x n d n )

where the holding cost hn corresponds to the number of units those are not consumed during a given period(s). The term hn, therefore, can be expressed as a function of s n + x n d n . The recursive relationship for any state, then, is expressed as

f n* ( s) = min c sxn + f n*+1 ( x n ) , where


f n* ( s ) is the optimal policy for the current state and f n*+1 ( x n ) is the optimal policy for the

transformed state in stage n + 1, which is known. This concept may be further illustrated by referring to Figure 2 below.
* f n ( s ) may be

Stage n xn

Stage n+1 sn + 1(j) sn + 1(j) sn + 1(j) Contribution of xn

obtained by identifying the minimum of the sum of the cost of a particular decision xn, and the optimal policy at the transformed state in stage n+1, given * by f n +1 ( x n )

State:

sn(i)

Figure 2: Identifying the optimal policy at the state sn. The computer routine for this recursive algorithm can now be developed by treating it as a three-dimensional array. This has been developed in Visual Basic 6.0. It could be subsequently

determined that the optimal policy is x1 = 30, x2 = 0, x3 = 40, and x4 = 0 with a corresponding cost of Rs. 155. This path (least-cost path) can be easily located on the network of Figure 1. 3.2 Information Feedback and Principle of Optimality Although the optimal solution provides the decision policy in a neat way, it may be appreciated that the decision environment is typically fluid in nature and at some future point in time market factors may change necessitating a review or revisions in essentially predetermined strategies. For the lot-sizing problem in particular, the market dynamics or fluidity may be reflected in terms of: 1. Availability of raw materials may increase due to quantity discounts or any other reason, thereby allowing a larger batch size to be manufactured. 2. Temporary shortages or price escalations in raw materials may force the company to opt for smaller batch sizes. 3. Change in workforce levels may also force decisions similar to (1) and (2) above. 4. Any other similar contingent situation. In the context of the problem network shown in Figure 1, implementing any of the nonoptimal decision options at any stage (identified by dashed lines) would mean that the system transforms to a state which is non-optimal. However, the principle of optimality for dynamic programming, clearly suggests that an optimal policy for the remaining stages is independent of the policy decisions adopted in the previous stages. This means that an optimal policy exists for any state regardless of how it was achieved. To illustrate this argument, let us suppose the system is in stage 2, and the immediate decision is not to manufacture anything (x2 = 0), thereby giving an optimal policy f2(s2) = 80. In this case the system transforms to a beginning inventory level of 0 in stage 3 (s3 = 0), from where f3(s2) = 80. However, if the decision at stage 2 had been to manufacture 25 units, or 40 units (both being feasible options there, though relaxing Condition 1), the system would transform to either (s3 = 25) or (s3 = 40) in stage 3. But the optimal policies at these states are already known. These are f3(s) = 40, x3 = 0, and f3(s) = 15, x3 = 0, respectively.
4

AN APPROACH TO RESOLVING SYMBOLIC DEMAND

For the decision situation identified, i.e., a pricing and a promotion action, information on demand exists in a symbolic form. This arises because experts have no reliable way of quantifying demand. Rather, they can only reason in terms of abstractions, specifically in the form of certain rules of the type IF price_cut is medium, THEN demand is high, for example. One approach towards resolving the problem, therefore, is a rule-based or a knowledge-based system. Rule-based systems (RBS) are programs that use preprogrammed knowledge to solve problems. The facts are generally stored in the form of IF-THEN rules. The RBS can use these rules as they are needed to solve problems when presented with data. The idea is that by using enough of these rules, one can cover the scope of the reasoning that an expert uses in a particular problem area. Rule-based systems need not always involve abstractions or symbolic information. However, abstraction simplifies thinking about the relationship among variables by reducing the combinations one has to worry about. If for a demand manager, high demand means anything between 25,000 and 50,000 units, then the category high demand encompasses the entire range of values between 25,000 and 50,000 and thus reduces the complexity of the problem by

reducing the need for a rule for every significant value. By using these subjective categories, things in the real world are evaluated by only the degree to which they relate to such categories. 4.1 Linguistic Ambiguity and Fuzzy Logic The use of subjective categories, however, brings problems of a unique kind. For example, we might believe implicitly that highly paid is upwards of Rs 50,000 per month. This is a generalization. But even though a person who made Rs 55,000 and a person who made Rs 78,000 would both be considered highly paid, the latter is clearly so to a greater degree. And what of the person who makes Rs 49,990 a month? Isnt that person highly paid? Fuzzy logic supports reasoning about these kinds of situations. It is based on gradations rather than sharp distinctions [2]. The power of fuzzy logic comes from the ability to describe particular phenomena or process linguistically and then to represent that description in a small number of very flexible rules. The knowledge in a fuzzy system is carried both in its rules and in fuzzy sets, which hold general descriptions of the properties of phenomena. This distribution further carries an advantage for sensitivity analysis. If a system produces an incorrect answer, the cause can be either a rule or the fuzzy sets it uses, both of which may be modified. 4.2 Modeling Linguistic Ambiguity To implement this human logic in system solutions, a mathematical model of fuzzy logic is required. The structure of such a system is shown in Figure 3.

Linguistic Variables Linguistic Level Numerical Level

Fuzzy Inference

Linguistic Variable

Fuzzification Numerical Variables

Defuzzification Numerical Variable

Figure 3: Structure of a fuzzy logic system. The fuzzification step essentially requires identifying all the relevant linguistic variables and their respective subjective categories or terms. However, input information is given numerically and the output is required to be numerical, too. Once all input variable values are translated into respective linguistic variable values, the fuzzy inference step evaluates the set of IF-THEN rules that define the evaluation. The result of this is again a linguistic value for the particular linguistic variable. The so-called defuzzification step translates this linguistic result into a numerical value that represents the variable as a number. These two translation steps are the cost involved in reaping the benefits of having rules that are interpreted in a human-like way. For both translation steps, a link between the numerical representation and the linguistic representation of a variable are needed. This link is the set of membership functions contained in each linguistic variable definition, articulated by the degree of membership concept. Both fuzzification and defuzzification use the membership functions to compute the translation step [7].

4.3

A Case Study On Demand Estimation Using Fuzzy Logic To demonstrate and validate the application of fuzzy logic for estimating demand in the wake of a pricing and promotion action, a real-life situation was considered. An example of a specific product of a prestigious company was taken up. Table 3 contains the relevant data. Table 3 Four input variables that will interact Linguistic variables and their possible numerical with the output variable demand are values identified. The possible ranges of 1. Product_age (in days) 0 120 numerical values that are associated 2. Price_sensitivity, (ratio) 05 with each of these values in the given 3. Promotion (percent of cost) 0 15 situation are also estimated. Each 4. Price_cut (rupees) 0 - 10 variable, in turn, contains two or more 5. Demand, incremental increase (in 30 - 70 than two terms (the so-called subjective units x 100) categories) identifiable in the given context. A set of appropriate rules to evaluate the impact of changes in the input variables on demand was next established. This is shown in Table 4. Table 4 Rule Set for the Product The Knowledge Base product is new AND IF promotion is THEN high price_sensitivity is high AND IF promotion is THEN high price_cut is very low AND IF promotion is THEN high price_cut is low AND IF promotion is THEN high price_cut is medium AND IF promotion is THEN medium price_cut is medium AND IF promotion is THEN high price_cut is high ----------------------------THEN price_cut is high AND IF promotion is THEN high price_cut is very high ----------------------------THEN

1. 2. 3. 4. 5. 6. 7. 8. 9.

IF IF IF IF IF IF IF IF IF

demand is medium demand is high demand is low demand is medium demand is medium demand is high demand is low demand is very high demand is high

Using the above input data and the knowledge base, the application was developed on a limited version of a fuzzy logic development environment fuzzyTECH, proprietary software supplied by Prentice Hall PTR, New Jersey. This system is fully compatible with Visual Basic 6.0 and has been integrated with the dynamic programming application module discussed in Section 3.1. It may be invoked directly from the menu screen of that application. 5. CONCLUSION

This study is strictly meant to be a research effort motivated by the need to to advocate for scientific and rational approaches towards managerial decision problems within the realm of production and operations management through techniques based on (but not limited to) operations research and intelligent decision support methods. Such approaches have been successfully integrated into the decision system that has been developed. At the research level, limitations of time and budget severely preclude the development of an elaborate system that is comparable with typical commercial packages for similar applications. However, this researchers experience with code development in Visual Basic and the software fuzzyTECH has been substantial enough to guarantee that there is all the potential available that may be further harnessed to develop fully functional customized applications. Moreover, the techniques may be used to address relevant problems in similar areas. 6. ACKNOWLEDGEMENTS

The authors wish to record and acknowledge the assistance and cooperation extended by the Institute of Business Administration, Karachi in relation to developing the business and cross functional theme of this study. 7. 1. 2. 3. 4. 5. 6. 7. 8. REFERENCES Anderson, Edward J., The Management of Manufacturing: Models and Analysis, Addison-Wesley, (1994) Dhar, Vasant and Stein, Roger, Intelligent Decision Support Methods, Prentice Hall, (1997). Fordyce, J. and Webster, F., The Wagner-Whitin method made simple, Production and Inventory Management, (1984), 25(2),. Hillier, F. S. and Lieberman, G. J., Introduction to Mathematical Programming, McGraw Hill, (1995). Labadie, John W., Knowledge-Based Optimization via Symbolic Dynamic Programming, Final Research Report, Center for Computer-Assisted Engineering, Colorado State University, Fort Collins, Colorado, (1990). Silver, Edward A., Pyke, David F. and Peterson, Rein, Inventory Management and Production Planning and Scheduling, John Wiley, (1998). Van Altrock, C., Fuzzy Logic and Neuro-Fuzzy Applications in Business and Finance, Prentice Hall, (1997). Zimmerman, H. J., Fuzzy Set Theory And Its Applications, Kluwer Academic Publishers, (2000).

You might also like