Вы находитесь на странице: 1из 6

AN ORDER ENCODING GENETIC ALGORITHM FOR LOT-SIZING PROBLEM WITH MULTIPLE SUPPLIERS Timur Keskintrk 1 zlem Akay Kasapoglu 2

Istanbul University, Faculty of Business Administration, Department of Quantitative Techniques, 34320 Avcilar / ISTANBUL Tel: +90 212 473 70 70 (18245), Fax: +90 212 590 40 00, E-mail: tktirk@istanbul.edu.tr 2 Istanbul University, Faculty of Business Administration, Department of Production, 34320 Avcilar / ISTANBUL Tel: +90 212 473 70 70 (18259), Fax: +90 212 590 40 00, E-mail: ozlemak@istanbul.edu.tr

Abstract: This paper presents a new method in genetic algorithm (GA) with order encoding (OE) for solving the traditional lot sizing problem with multiple suppliers. There are many types of encoding depending on the structure of the problem in GA, which is a heuristic optimization technique. The most used way of encoding is binary. OE is inspired by binary encoding and developed especially for lot sizing problems. Results from computational experiments on a set of problems show that the proposed technique produces high-quality solutions.
Keywords: Lot- sizing, Order Encoding, Genetic Algorithms.

1-Introduction MRP is a field of study which determines the production schedules and net requirements. Although it is an old field of study, it is still daily used by manufacturers for deciding on the order quantity, order period and suppliers. In order for a manufacturing firm to match its demand with supply, it has to deal with issues well beyond its own organizational boundary. For instance, the firm has to take into account suppliers capability as well (Kim, Leung, Park & Zhang, Lee, 2002). Products can be sourced from a set of approved suppliers. A supplier dependent transaction cost applies for each period in which an order is placed on a supplier. A product dependent holding cost per period applies for each product in the inventory that is carried across a period in the planning horizon (Basnet, Leung, 2005). We propose a lot sizing model that considers suppliers capacity in the period when the order is placed. If the capacity of the supplier is not available at that period, another approved supplier is taken into account. The basic philosophy is to ensure that the right number of products is planned at the right time. Our problem category is a single level with constrained resources. Determining a proper lot sizing policy is a key dimension of inventory control, as placing proper batches can allow significant reduction in inventory control. (Dellaert, Jeunet & Jonard, 2000). In determining the lot sizing problem with multiple suppliers, the objective is minimizing the total cost by optimizing the order and holding cost. In this problem, the decisions are when and how much to order from which suppliers, in which periods. In this paper, a genetic algorithm method is proposed which represents all the decision variables in its chromosomes. In the OEGA the aim is to have chromosomes which are more efficient and faster than those of Binary GA. 2-Related work The lot sizing problem has been widely studied under different aspects of demand rates (constant, time varying), demand characteristics (deterministic or stochastic), products (single or multiple), quantity discount or not, etc. Wagner Whitin developed an optimal solution for dynamic multistage production systems under some special assumptions, but it still suffers from computational complexity, and thus several heuristics approaches have been proposed such as Silver- Meal (SM), Lot for Lot (L4L), Part Period Balancing (PPB) heuristics. However, these methods guarantee only a local optimal solution. For other cases of lot sizing problem some interesting results are obtained under additional assumptions. (Van, Tabucanon, 2004). For instance, Pujawan (2003) focused on analytical and simulation models where order variability could also be 35th International Conference on Computers and Industrial Engineering

1135

affected by lot sizing techniques applied by a supply chain channel in determining the quantity of orders to be placed to its upstream channel. Gupta and Magnusson (2005), in their article are concerned with a CLSP (capacitated lot sizing and scheduling problem) with a single resource over multiple periods that span the planning horizon. Basnet, Leung (2005), presents a multi-period inventory lot sizing scenario, where there are multiple products and multiple suppliers. Dellaert et al developed a binary encoding genetic algorithm for Multi-Level Lot-Sizing Problem (MLLS). Khouja, Michalewicz & Wilmot (1998) proposed a paper to investigate the use of genetic algorithms for solving Economic Lot Size Scheduling Problem (ELSP). 3-Problem Description 3-1 Assumptions The following assumptions apply to the classical economic lot sizing problem 1. Demand is known and constant. 2. Lead time, that is the time between the placement of the order and the receipt of the order, is known and constant. 3. Receipt of inventory is instantaneous. In other words, inventory from an order arrives in one batch, at one time. 4. Quantity discounts are not possible. 5. The only variable costs are the costs of setting up or placing an order (set up cost) and the cost of holding or storing inventory over time ( holding or carrying cost). 6. Stockouts (shortages) can be completely avoided if orders are placed at the right time (Heizer, Render, 1995). 7. The item is treated entirely independently of other items; that is, benefit from joint review or

replenishment does not exist or it is simply ignored. 8. The entire replenishments of each period must be available at the beginning of periods.
9. The cost factors do not change appreciably with time; in particular, inflation is out of consideration. 10. The entire order quantity is delivered at the same time. (Silver, Pyke & Peterson, 1998). 11.In our problem there are multiple suppliers for the same product, if the capacity of that supplier is not available at that period, the order is placed from another approved supplier, the batch in one period can not be shared between suppliers, only one supplier should place the order. 3-2 Notations We formulate the lot-sizing problem with multiple suppliers using the following notation: Indices i i=1,2N Index of suppliers. t t=1,2.K Index of time periods. Cost parameters H Holding cost of product per period OCi Ordering cost for supplier i Quantity variables Dt Demand for product in period t Ci Capacity of supplier I It Level of inventory at the end of period t Decision variables Qt X,t Order quantity in period t if an order is placed on supplier i in period t, i otherwise 0. (X0 = 0)

1136

35th International Conference on Computers and Industrial Engineering

3-3 Mathematical Model

Minimize :

( OC( X ) + HI t )
N t =1
t

subject to : I t = I ( t 1) + Qt Dt ; t = 1,..., N , Qt 0; t = 1,..., N , I t 0; t = 1,..., N , Qt C( X t ) ; t = 1,..., N , 0 if Xt = i if Qt = 0, Qt > 0,

where

4- Genetic Algorithm Genetic algorithms are search and optimization methods which were developed initially by J. Holland and his associates at the University of Michigan in the 1960s and 1970s (Reeves, 1995). These algorithms encode a potential solution to a specific problem on a simple chromosome and apply genetic operators to these structures so as to preserve critical information. Better individuals which are represented by chromosomes, are likely to be the winners in a competing environment and tend to yield good quality offspring, and this means better solutions to the problem. (Man, Kwong, 1996). The procedure of GA is given as follows:

Figure 9 Flowchart of a simple GA

Encoding, fitness function, selection, crossover, and mutation are considerations of GA. Proposed genetic algorithm model is flexible which could be used for any period length and any number of suppliers. 5- Design of OEGA for lot-sizing problem with multiple suppliers. In an application of GA to a problem, possible solutions are represented in the form of chromosomes. Chromosomes are composed of genes which represent variables (in our problem periods). Chromosomes lengths change according to period number in the problem. Every gene has a value 0 if there is no order at that period, if there is, the gene takes the indices (i=1,..., N ) from the supplier that the order will be placed. N is the number of suppliers. In figure 2 there are 3 chromosomes of different alternative solutions for 10 periods, 5 suppliers and demands. For example in the first chromosome and in the first period, while there is an order from the supplier 1, there is no order in period two and three, there is also an order in period four from supplier 5 and so on. The quantity of the first replenishment from supplier 1, with a fixed lead time one,

35th International Conference on Computers and Industrial Engineering

1137

at the first period is, Q1 = 50+75+120= 225. The goal of this encoding is to improve the cost effectiveness and execution speed.
Demand Chromosome 1 Chromosome 2 Chromosome 3
50 1 5 3 75 0 4 0 120 0 0 0 100 5 0 0 65 0 0 1 80 2 4 4 90 0 1 0 50 0 0 0 40 1 0 2 75 2 2 0

Gene 1 Gene 2 Gene 3 Gene 4 Gene 5 Gene 6 Gene 7 Gene 8 Gene 9 Gene 10

Figure 10 Coding scheme

The genetic algorithm starts with a large commune of chromosomes known as the initial population (Haupt, Haupt, 1998). After the initial population is created, chromosomes are rated in terms of their fitness. Fitness function is developed according to the problem. In OEGA the fitness function contains a penalty function for the circumstances where the fitness function becomes distant from the target value. Penalty function method is well known for solving constrained optimization problems for decades. The approach converts the problem into an equivalent unconstrained problem and then solves it using suitable search algorithm (Sarker, Newton, 2002). In our problem the penalty function value is added to the fitness function value when there is an understock position. The fitness function becomes distant from the optimum when there is an unresponded demand. In figure 4 there is an algorithm for the fitness function of our problem.

After chromosomes fitness values are evaluated, next step is selection. The selection operator chooses the individuals that will survive. Better (fitter) individuals have larger chances of survival (Kamrani, Rong & Gonzalez, 2001). There are many methods for the selection. Roulette wheel selection, tournament selection, rank selection, steady state selection are some of them. We use roulette wheel selection in our OEGA. Crossover along with selection reproduction and mutation is one of the main operators of any GA. It provides an exchange of design characteristics between paired individuals (designs). Several crossover techniques have been devised to accomplish this task, amongst them are single-point, 2-point, multipoint (involving more than two cut points), variable to variable and uniform crossovers (Hasanebi, Erbatur, 2000). According to the size of the problem, single-point, two-point or multi-point crossover could be used. (Eroglu, Keskintrk, 2005). 1,0 crossover rate has been used. There is a single point crossover in the crossover point 3, for OEGA chromosome in Figure 3.
C h rom o s o m e 1 C h rom o s o m e 2 O ffspring 1 O ffspring 2
1 5 1 5 0 4 0 4 0 0 0 0 5 0 0 5 0 0 0 0 2 4 4 2 0 1 1 0 0 0 0 0 1 0 0 1 2 2 2 2

Figure 3 Single -point crossover for OEGA

1138

35th International Conference on Computers and Industrial Engineering

function lotsizefitness(pop,ds,hc,oc,sc); % pop=population,ds=demands,hc=hold cost %oc=order cost, sc=suppliers'capacities [row,column]=size(pop); lsf=zeros(1,row); for i=1:row suppliers=nonzeros(pop(i,:)); periods=find(pop(i,:)); for j=1:size(suppliers,1) lsf(i)=lsf(i)+oc(suppliers(j)); end for j=1:size(periods,2)-1 order(periods(j))=min(sum(ds(periods(j):periods(j+1)-1)),sc(suppliers(j))); end order(periods(end))=min(sum(ds(periods(end):end)),sc(suppliers(end))); suppliers(column)=0; suppliers(1:size(order,2))=order; stocks=zeros(1,column); stocks(1)=suppliers(1)-ds(1); for j=2:column if stocks(j-1)>0 stocks(j)=stocks(j-1)+suppliers(j)-ds(j); else stocks(j)=suppliers(j)-ds(j); end end lsf(i)=lsf(i)+(sum(stocks)*hc)+(abs((sum(ds)-sum(order)))*100); order=[]; end lotsizefitness=lsf;

Figure 4 Fitness function.

Mutation models bring a random change in genetic information of creatures and are inspired by random change of genetic information in living organisms, e.g. through the effects of radiation or chemical mismatch (Schmitt, 2001). Mutation plays decidedly secondary role in the operation of GA. In artificial genetic systems the mutation operator protects against such an irrecoverable loss (Goldberg, 1989). In binary encoding, when a gene in a string is selected to be mutated, we just flip the bit from 0 to 1 and from 1 to 0. This means that only the selected parameter in the chromosome will change its value and the rest will remain unchanged. In our OE encoding, a unique gene undergoes mutation too. However changing of the gene which is selected to be mutated, is different from binary encoding. Gene can change its value from 0 to i or opposite, where i is an index number of suppliers and can possibly mutate from i to another i. Let us consider the single-point mutation for OE encoding in Figure 5, the mutation point is 3 and the genes value is changed from 0 to 3.

5 5

4 4

0 0

0 3

0 0

4 4

1 1

0 0

0 0

2 2

Figure 5 OE Encoding single-point mutation

0,4-0,6 mutation rate has been used because of the structure of the problem in this paper. Some experimental results show that increasing the rate of mutation is improving the performance of the OEGA 6. Experimental Results and Conclusion The aim in the lot sizing problem where the demand is known and constant, is to keep the total relevant costs as low as possible satisfying the demand for that period. We generated some test problems to compare the efficiency of our OE encoding and binary encoding on the PC (Pentium 4 CPU 3.00 GHz, 512 MB RAM, Windows XP). The GA code is developed in MATLAB. Demands for different periods were randomly generated from integer [0,100], and the suppliers capacity from integer [100,500], order costs for every

35th International Conference on Computers and Industrial Engineering

1139

supplier were generated between integers [50, 150]. We assumed the holding cost and the lead time as one. The average of 50 test problems for each size is shown in Table 1. This study reveals that the Cpu times for our OEGA encoding is clearly shorter than those for binary encoding.
Table 1 Computational Results
Problem Size Supplier 3 5 5 5 10 Period 10 10 15 20 50 Population Size 20 20 40 40 50 Best Fitness 550 614 631 871 2800 Average CPU Times OE Encoding 1,4181 1,6194 2,3491 3,4009 114,746 Binary Encoding 1,508 2,1237 4,0028 7,4744 278,1928 % Difference in CPU time 6,34% 31,14% 70,40% 119,78% 142,44%

We have presented a different type of encoding OE, for the lot sizing problem with multiple suppliers. We also considered the suppliers capacity. In this study the model is developed for one item only but it could be improved for multi items and multi level problems by using OEGA. This type of encoding could be used in some other applications of genetic algorithms, transportation problems could be an example. Although there are lots of papers for traditional lot sizing problem there are not many quantitative models considering the lot sizing problem with capacity constraints. References 1. Basnet, C.,Leung, J., M., Y.(2005). Inventory lot sizing with supplier selection. Computers& Operations Research, 32, 1-14. 2. Dellaert, N., Jeunet., J, Jonard., N.(2000). Genetic algorithm to solve the general multi level lot sizing problem with time varying costs. International Journal of Production Economics, 68, 241-257. 3. Eroglu, E., Keskintrk, T. (2004). In: Proceedings of the International Logistics Congress, (pp. 265-272), Izmir. 4. Goldberg, D.E. (1989). Genetic Algorithms in Search Optimization and Machine Learning, USA, Addison Wesley Publishing Company. 5. Grupta, D., Magnusson., T. (2005). The capacitated lot-sizing and scheduling problem with sequencedependent setup costs and setup times. Computers & Operations Research, 32, 727-747. 6. Hasanebi, O., Erbatur, F..(1999).Evaluation of Crossover Techniques in Genetic Algorithm Based Optimum Structural Design. Computers and Structures, 78, 435-448. 7. Haupt, R.L., Haupt, E., S.(1998). Practical Genetic Algorithms. John Willey & Sons. 8. Heizer, J., Render, B.(1996). Production & Operations Management, Prentice Hall ,New Jersey,1996.

9. Huang, Y.P, Huang C.H.(1997). Real-valued genetic algorithms for fuzzy grey prediction system, Fuzzy Sets and Systems, 87, 265-276.
10. Kamrani, A., Rong, W., Gonzalez R.(2001) A genetic Algorithm Methodology for data mining intelligent knowledge acquisition. Computers& Industrial Engineering, 40, 361-377. 11. Khouja, M., Michalewicz, Z., Wilmot, M.(1998). The use of genetic algorithms to solve the economic lot sizing scheduling problem. European Journal Of Operational Research.110, 509-524. 12. Kim B., Leung, J. M. Y., Park, T., K., Zhang, G., Lee, S.(2002) Configuring a manufacturing firms supply network with multiple supplier. IIE Transactions,34, 663-667. 13. Man, T., Kwong, S. (1996). Genetic algorithms: concepts and applications. IEEE Transaction on Industrial Electronics, 43(5), 519534. 14. Pujawan ,N.(2004). The effect of lot sizing rules on order variability. European Journal of Operational Research, 159, 617-635. 15. Sarker, R., Newton C. (2002) A genetic algorithm for solving economic lot size scheduling problem. Computers and Industrial Engineering, 42, 189-198. 16. Schmitt, L.M. (2001). Theory of genetic Algorithms. Theoretical Computer Science, 259,1-61. 17. Silver, E.A., Pyke, D.F., Peterson, R.(1998). Inventory Management and Production Planning and Scheduling, John Willey & Sons , New York. 18. Van H., N., Tabucanon M., T. (2004). Adaptive genetic algorithm for lot-sizing problem with selfadjustment operation rate. International Journal of Production Economics, Article in press.

1140

35th International Conference on Computers and Industrial Engineering

Вам также может понравиться