Optimization of Product Mix Planning in High-Mix-Low-Volume Industries Using Genetic Algorithms SIEW CHIN NEOH1, NORHASHIMAH MORAD2, CHEE PENG LIM1, and ZALINA ABDUL AZIZ1 1 School of Electrical and Electronic Engineering, University Science Malaysia Engineering Campus, 14300 Nibong Tebal, Penang, MALAYSIA 2 School of Industrial Technology, University of Science Malaysia 11800, Minden, Penang, MALAYSIA Abstract: - This paper describes a product mix optimization model using Genetic Algorithms (GAs) for HighMix-Low-Volume (HMLV) industries. The main objectives of the proposed GA-based planner are to reduce the high conversion time due to frequent product changeovers and to allow a balanced workload throughout a particular planning horizon. In addition, the product mix model is developed to fulfill the customer demands as well as to prevent over-loaded capacity. The proposed approach is applied to data obtained from a semiconductor manufacturing plant and the results indicate that the optimized product mix planner performs better than those obtained from conventional planning methods. Key-Words: - Product Mix, Genetic Algorithms, High Mix Low Volume, Optimization 1 Introduction Product mix analysis is described as the process of determining the percentage of each product in different production and business activities [1]. This process is complex as it involves various inputs such as resource utilization, resource limitation, product variety, product characteristics, and market demand. Thus, product mix optimization plays an important role in capacity planning especially in High-MixLow-Volume (HMLV) manufacturing environments which need to produce more product variety with lower volume in the dynamic business climate. In product mix planning, the theory of constraints (TOC) and linear programming (LP) are two widely used approaches. Luebbe and Finch compared TOC and LP and clarify the differences between TOC philosophy and LP technique [2]. TOC heuristic is easily understandable and simpler to use than LP [2], but LP is a better planning tool than TOC when multiple constrained resources exist [3-4]. In [3], it is argued that traditional accounting methodologies cannot determine optimal product mix and that the dropping marginally profitable products based on the accounting rules will give less than optimal profitability. Beside TOC and LP, Genetic Algorithms (GAs) have been used in product mix planning problems. GA-based TOC has been proposed in [5] as it can easily find very high quality solutions for both small and large problems at reasonably small computation time. GA has also been applied to the optimization of product mix and material match in [6]. In this paper, a GA-based product mix optimization model is proposed to minimize conversion time, to fulfill predicted demand, to prevent overloaded capacity, and to achieve a balanced workload throughout a planning horizon. Since HMLV requires enhanced organization and process discipline to embrace a high degree of multitasking [7], it has always been a competitive manufacturing strategy for HMLV industries to obtain systematic product mix planning. The proposed approach is applied to a semiconductor manufacturing plant to solve practical, real-world product mix problems encountered by manufacturing firms. The problem of HMLV manufacturing systems that requires high conversion time has been addressed. Time means money, and the reduction in changeover and conversion time can allow the production line to maximize its resource utilization for more value-added processes. 2 Problem Formulation Multi-product production in the HMLV industries involves a large number of parameters and constraints. In Fig. 1, there are n types of products to be planned, Pi (i 1,2 ,...,n) represents the ith type of product in the product mix whereas Di (i 1,2 ,...,n) represents the ith product’s demand that must be fulfilled within the planning horizon. The planning time index is given as Wk (k 1,2 ,...,t) ,in which it determines the interval within the product mix planning horizon (e.g. weekly for a month or monthly for a year). In the proposed model, the product mix problem is represented by randomly generating product occurrence, Oik, and product volume, Vik, where i represents the product type and k represents the planning index. Product occurrence is used to indicate whether or not a product has been selected for production for that particular week or month (Equation 1). 1 if product i will be produce in week k Oik 0 if otherwise (1) After obtaining the product occurrence, product volume for each particular time index, Vik will then be generated randomly by fulfilling the product demand, Di and production loading rules. On the production floor, minimum production volume per product per week is a constraint. Mi (i 1,2 ,...,n) is used to represent the minimum production volume per product i per week. When the demand of product i, Di, is nonzero, the sum of product occurrence throughout the planning horizon has to be more than 1 but less than the total product demand, Di, divided by the minimum production volume per product i per week, Mi (Equation 3). Note that Mi ≠0. On the other hand, if Di is zero, the sum of product occurrence should be zero (Equation 4). In other words, there is no product needed to be produced if there is no demand for it. t 1 Oik k 1 t O ik Di , Mi 0 , k 1 Product type (Pi) P1 P2 P3 : : Pn Product type (Pi) P1 P2 P3 : : Pn W1 O11 O21 O31 : : On1 W1 V11 V21 V31 : : Vn1 Week (Wk) W2 O12 O22 O32 : : On2 W3 O13 O23 O33 : : On3 Week (Wk) W2 V12 V22 V32 : : Vn2 W3 V13 V23 V33 : : Vn3 …… …… …… …… Wt O1t O2t O3t : : Ont Predicted Demand (Di) D1 D2 D3 : : Dn Wt V1t V2t V3t : : Vnt Predicted Demand (Di) D1 D2 D3 : : Dn 2.1 Demand constraint The planning of product mix requires each product to fulfill the demand constraint of the whole planning horizon. Therefore, the sum of the generated product volume throughout the planning horizon has to be equal to the product demand, i.e., t ik Di , for i 1,2,3,..., n k 1 2.2 Production loading constraint (3) if Di 0 (4) 2.3 Capacity Constraint Capacity constraints refer to the resource ability and availability to process a particular set of product mix. Equation (5) shows that the available capacity of week kth, ACk is equal to the difference between total capacity of week kth, TCk and the down time capacity of that week, DCk which includes all capacity used for scheduled down time (e.g. preventive maintenance and engineering experiment ) and buffer time. ACk TCk DCk , for k 1,2,3,..., t (2) (5) The capacity usage based on the generated product mix in week kth, UCk (k 1,2 ,...,t) has to be always less than or equal to the available capacity in the production floor in week kth, ACk (k 1,2 ,...,t) as shown in Equation (6). UCk ACk Fig.1: Product mix structure in the proposed model V if Di 0 (6) 3 GA-based Product Mix Planning GAs are stochastic global search and optimization methods that mimic the metaphor of natural biological evolution [8]. Throughout the genetic evolution, better solution to the problems is obtained. Its effectiveness in industrial applications and ease of use has been demonstrated by a number of researchers [9]. Product type (Pi) P1 P2 P3 : : Pn W1 ….. ….. ….. : : ….. Week W2 ….. ….. ….. : : ….. (Wk) W3 ….. ….. ….. : : ….. W4 ….. ….. ….. : : ….. Product type (Pi) P1 P2 P3 : : Pn W1 ….. ….. ….. : : ….. Week W2 ….. ….. ….. : : ….. (Wk) W3 ….. ….. ….. : : ….. W4 ….. ….. ….. : : ….. Fig.2: Randomly Selected Crossover 3.1 Chromosome Representation In GA, the representation scheme determines how the problem is structured. In the proposed model, the product mix for a particular planning horizon is represented by a 2D matrix, and two types of encoding are used: 1) binary encoding and 2) real number encoding. Binary encoding is used to indicate the product occurrence Oik whereas real number encoding is used to indicate the weekly product volume, Vik that should be tested. The structure of the chromosome representation is shown in Fig.1. 3.2 Initialization and Selection Function The model will randomly generate solutions for the entire population. Stochastic Universal Sampling which is a single phase sampling algorithm with minimum spread and zero bias is used as the selection function to keep the expected number of copies of each chromosome into the next generation. 3.3 Genetic Operators 3.3.1 Crossover In the proposed model, Randomly Selected Crossover is used. The principle of Randomly Selected Crossover is depicted in Fig.2. The crossover probability is given as Pxovr, in which each individual has Pxovr chances to be selected for a crossover. For each crossover, number of rows (number of product) will randomly be selected from a 2D matrix (an individual) to be inter-exchanged with another individual to produce offspring (new individuals). 3.3.2 Mutation Two Mutation techniques are suggested in the model: Weekly Mutation (Fig.3) and Randomly Selected Mutation (Fig.4). Weekly Mutation is to swap the volume for randomly selected product in a particular week to another week whereas Randomly Selected Mutation is to regenerate random product volume for a randomly selected product. The probability for an individual to be selected to undergo mutation is Pmu for Weekly Mutation and Pm for Randomly Selected Mutation. Product type (Pi) P1 P2 P3 : : Pn W1 ….. ….. ….. : : ….. Week W2 ….. ….. ….. : : ….. (Wk) W3 ….. ….. ….. : : ….. W4 ….. ….. ….. : : ….. Fig.3: Weekly Mutation Product type (Pi) P1 P2 P3 : : Pn W1 ….. ….. ….. : : ….. Week W2 ….. ….. ….. : : ….. (Wk) W3 ….. ….. ….. : : ….. W4 ….. ….. ….. : : ….. Fig.4: Randomly Selected Mutation 3.4 Evaluation Function In this paper, an aggregating approach is applied to minimize conversion capacity and frequent changeover, to prevent overloaded capacity, and to achieve a balanced workload. Equation (7) shows the formulation of the objective value. (7) k 1 CVCk (k 1,2 ,...,t) refers to conversion capacity for week kth. It includes the hardware and software changing time, temperature waiting time, and lot set up time. The minimization of conversion capacity is needed to reduce the impact of rapid changeovers brought by HMLV systems. OVPk (k 1,2 ,...,t) is a penalty function given to product mixes that overload the weekly resource capacity for week kth. It indicates that the production floor is no longer capable of fulfilling the weekly demand based on the available resource capacity. In other words, OVPk is applied when UCk is more than ACk. UBPk (k 1,2 ,...,t) is a penalty function given to unbalanced workload throughout the planning horizon. Fig.5 shows an example of unbalanced workload in two weeks capacity consumption. The product mix distribution is not systematically planned: machine utilization may be too high in week 1 but too low in week 2. Thus, the introduction of UBPk is used to address unbalanced workload problems in product mix planning. GAP in Fig.5 refers to unused capacity or difference between available capacity and used capacity. Week 1 Week 2 20% 4 Results and Discussion As the proposed objective function is a minimization function, the lower the objective value, the better the product mix is. Fig.6 shows that the objective value decreases (better product mix is obtained) as the number of generation increases. Fig.6: Best Objective Value for each generation Weekly Capacity Comsumption (Original Product Mix) Capacity (%) t ObjectiveValue (CVCk OVPk UBPk ) 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 21.17 24.63 38.07 52.58 Available Weekly Capacity Time Frame 66.33 63.46 46.45 31.35 11.90 16.07 1 2 15.48 12.50 3 4 Week 90% DT Available Capacity GAP GAP Used Capacity Fig.7(a): Weekly capacity usage based on original product mix Weekly Capacity Usage (Proposed Model) Fig.5: Unbalanced workload in capacity usage. 3.5 Termination and Parameters Selection The termination criterion is met once the same solution has been continuously obtained by the GA for more than 40 reproduction trials. The population size used in the case study is 70 and the maximum generation used is 100. The probabilities for genetic operators are given as: Pxovr = 0.7, Pm = 0.7 and Pmu = 0.3. Capacity (%) Used Capacity 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 31.88 56.22 11.90 1 26.55 28.09 28.97 57.38 56.44 58.53 16.07 15.48 12.50 3 4 2 Down Time Week Gap Used Capacity Fig.7(b): Weekly capacity usage proposed product mix based on Using data from a semiconductor plant, weekly capacity consumptions based on the proposed GA product mix and the original product mix are compared. As shown in Fig.7(a) and 7(b), the proposed model produced a more balanced workload throughout the four weeks planning horizon as compared to the original product mix using predicted demand method. In Fig.7(a), the weekly workload is not distributed equally. More balanced workload can be observed in Fig.7(b) where the GAP is more or less the same for each week. The performance of the original and GA-based product mix are showed in Table 1. In comparison, the overall objective value has been improved by 65.93%. GA-based product mix shows better performance in conversion capacity (CVC) and unbalanced workload penalty (UBP) in which CVC has been reduced by 28.19% and UBP has been reduced to zero. Both original and GA-based product mix has zero value for OVP because there are no overloaded capacity and the production is still able to support the demand. The results show that the proposed model has managed to reduce conversion time caused by high product mix and, at the same time, is able to achieve a more balanced workload in the weekly product mix distribution. 5 Summary In this paper, chromosome representation using both binary and real-value encodings is suggested to represent the HMLV product mix structure. Based on the analysis and comparison of GA-based product mix planning and the original product mix planning, the GA approach shows its effectiveness in HMLV product mix optimization. The proposed GA-based model outperforms the original one in the reduction of conversion time and achievement of a balanced workload. Instead of having time lost due to frequent changeover, more time can be utilized for more value-added jobs. Besides, the model has also managed to fulfill each product demand throughout the planning horizon and, at the same time, prevent overloaded capacity. As a result, the GA-based model has proved its efficacy to be deployed as a product mix planning tool for HMLV manufacturing systems. Table 1: Comparisons of GA based product mix and original product mix using the proposed objective function evaluation Evaluation Function Original Product Mix GA based Product Mix Improvement (%) Objective Value 380.56 129.66 65.93 OVP (overloaded penalty) 0 0 0 CVC (conversion capacity) 180.56 129.66 28.19 UBP (unbalanced workload penalty) 200 0 100 References: [1] R. Al-Aomar, Product-Mix Analysis with Discrete Event Simulation, Proceedings of the Winter Simulation Conference, 2000, pp. 13851392. [2] R. Luebbe and B. Finch, Theory of constraints and linear programming: a comparison, International Journal of Production Research, Vol.30, No.6, 1992, pp.1471-1478. [3] T.N. Lee and G. Plenert, Maximizing product mix profitability – what’s the best analysis tool, Production Planning and Control, Vol.7, No.6, 1996, pp. 547-553. [4] G. Plenert, Optimization theory of constraints when multiple constrained resources exist, European Journal of Operation Research, Vol.70, No.1, 1993, pp. 126-133. [5] G.C. Onwubolu and M. Mutingi, Optimizing the multiple constrained resources product mix problem using genetic algorithms, International Journal of Production Research, Vol.39, No. 9, 2001, pp. 1897-1910. [6] Sk.A. Ali, R. de Souza, and Z. Hossain, Intelligent product mix and material match in electronics manufacturing, Neural, Parallel and Scientific Computations, Vol.11, 2003, pp. 97118. [7] B. Bilbrough, The right tool for the job, Circuits Assembly, November 2002, pp. 26-30. [8] J. Holland, Adaptation in Natural and Artificial Systems, The University of Michigan Press, Ann Arbor, 1975. [9] G. Mitsuo and R. Cheng, Genetic Algorithms and Engineering Optimization, John Wiley & Sons Inc.: New York, 2000.