Proceedings of the 2004 Winter Simulation Conference R. G. Ingalls, M. D. Rossetti, J. S. Smith, and B. A. Peters, eds. SIMULATION-BASED OPTIMIZATION FOR MATERIAL DISPATCHING INA RETAILER NETWORK Ganesh Subramaniam Abhijit Gosavi Department of Industrial Engineering University at Buffalo, The State University of New York Buffalo, NY 14260, U.S.A. ABSTRACT This paper presents preliminary work done on simulation- based optimization of a stochastic material-dispatching sys- tem in a retailer network. The problem we consider is one of determining the optimal number of trucks and quantities to be dispatched in such a system. Theoretical solution models for versions of this problem can be found in the literature. Unlike most theoretical models, we can ac- commodate many real-life considerations, such as arbitrary distributions of the governing random variables, and all important cost elements, such as inventory-holding costs, stock-out costs, and transportation costs. We have used two techniques, namely, neuro-response surfaces and simulated annealing, for optimizing our system. We have also used a problem-specific heuristic, known as the mean demand heuristic, to provide us with a good starting point for sim- ulated annealing and a benchmark for our other methods. Some computational results are also provided. 1 INTRODUCTION Typically, a supply chain of consumer goods, such as gaso- line, food products, and clothing items, consists of distri- bution centers, warehouses, and retailers. The distribution industry focuses on transporting goods from the manufac- turer to the customers. A goal of this industry is to make the distribution process “lean", and thereby achieve cost bene- fits. The problem we have considered here is geared towards reducing the inventory in the distribution network and en- suring a satisfactory service level. Generally, a warehouse serves multiple retailers where customers arrive randomly to buy products. An optimization problem commonly-faced by managers is to determine (1) the number of trucks to be dispatched, and (2) the amount of goods to be dispatched. Associated with this problem are the costs of holding excess inventory (inventory-holding costs), not being able to meet customer demand (stock-out costs), and transporting goods from the manufacturer to the retailers (transportation costs). The cost function that we have developed in this paper accounts for all of these elements. When there are multiple retailers and each retailer has unique random characteristics, such as arrival rate of customers and size of the demand, one has a large-scale and complex stochastic optimization problem on which it is not easy to construct an exact theo- retical model. In this paper, we study a complex problem with multiple retailers and a large number of governing random variables, and use a simulation-based approach for solving it. Seminal work on this problem is from Clark and Scarf (1960). In their paper, they have assumed a holding and shortage cost but ignored the setup cost or the reorder cost. Jönsson and Silver (1987) have suggested the use of a “redistribution" strategy in which the inventories at the retailers are pooled and redistributed to standardize the inventory at each retailer. Their model is for systems in which demand variation is low and in which it can be shown that the stockouts can occur only in the last periods of an order cycle. McGavin, Schwarz, and Ward (1993) have suggested a so-called “between-replenishment," “risk-pooling" policy for this problem. They have a two-interval alloca- tion policy in which the stock is withdrawn from the warehouse at two (unequal) intervals in the same order cycle. Their model ignores the inventory holding costs. Nahmias and Smith (1994) have developed a model for demands that have the negative-binomial distribution. Federgruen and Zipkin (1984) have modeled an extension of the previous work by Eppen and Schrage (1981) overcoming some of the limitations like normal distribution of the demand and identical holding and penalty costs across all the retailers. But they develop a myopic model, i.e., a model in which the system is optimized in the period during which the actual allocation occurs, ignoring the 1412