روش جدید برای مدیریت موجودی تصادفی چند پله ای و محیط های فازی عصبی

Managing inventory in a multi-echelon supply chain is considerably more difficult than managing it in a single-echelon one. A strategy that optimizes inventory one echelon at a time results in excess inventory without necessarily improving service to customer. In this paper, a methodology for effective multi-echelon inventory management is proposed. Subsequently; a neural network simulation of the model is then presented with the support of neuro-fuzzy demand and lead time forecasting, and finally its performance is calculated using performance metrics selected from the SCOR model. The results show that, the inventory is efficiently deployed and uses realistic breakdowns. The proposed methodology aims to provide an important tool for the management of general N-echelon tree-structured supply chains that overcomes some of the deficiencies of competing methodologies.

A manufacturing supply chain is a network of suppliers, factories, subcontractors, warehouses, distribution centers, and retailers, through which raw materials are acquired, transformed, produced, and delivered to the end customers. Such a supply chain network must meet customers’ service goals at specified service levels at the lowest possible cost (Bhaskaran and Leung, 1997, Tsiakis et al., 2001 and Taskin and Guneri, 2005). Supply chain inventory management (SCIM) is an integrated approach to the planning and control of inventory throughout the entire network of cooperating organizations from the source of supply to the end user. SCIM goals are to improve customer service, increase product variety, and decrease costs (Giannoccaro et al., 2003). Fig. 1 shows a multi-echelon system consisting of a number of suppliers, plants, warehouses, distribution centers and customers (Axsater, 1990, Andersson and Melchiors, 2001 and Axsater, 2003). Full-size image (36 K) Fig. 1. A multi-echelon inventory system. Figure options An important issue in supply chains (SCs) is the need to make decisions in the face of uncertainty. In addition to demand and lead time fluctuations for a product, the uncertainty is compounded by information delays associated with the manufacturing and distribution processes that characterize SCs (Giannoccaro et al., 2003). To incorporate uncertainty in a supply chain model a suitable means of representing it must be found (Gupta and Maranas, 2003). Three methods are frequently used for representing uncertainty (Gupta and Maranas, 2003 and Hameri and Paatela, 2005): First, the uncertain parameters are assumed to have normal distribution with a specified mean and standard deviation; second, the forecast parameters are described as fuzzy numbers defined by their degree of membership; and third, expected outcomes are captured by several discrete scenarios and their associated probability levels (Chen and Lee, 2004). Examples of all three approaches can be found in the literature (Taskin Gümüs and Güneri, 2007). For example, Gupta and Maranas (2000), Gupta et al. (2000) incorporate uncertain demand in their supply chain model via a normal distribution function, and propose a two-stage solution to the restocking problem. They have recently published a generalization of their model that can handle multi-period and multi-customer problem (Gupta and Maranas, 2003). Tsiakis et al. (2001) use the scenario planning approach to describe demand uncertainties. Petrovic et al. (1999) use fuzzy sets to handle uncertainties in customer demand and the supply of raw materials for manufacturing products. Giannoccaro et al. (2003) also apply fuzzy set theory to model the uncertainties associated with both market demand and inventory costs. In this paper, Mitra and Chatterjee’s (2004) deterministic and stochastic models of supply chain inventories are improved by the addition of neuro-fuzzy demand and lead time forecasting, the incorporation of variable costs and generalization to N-echelons. The goal of the work is to improve supply chain performance by minimizing uncertainty. Mitra and Chatterjee (2004) examine De Bodt and Graves’ model (1985), that they developed this model in their paper namely “Continuous-review policies for a multi-echelon inventory problem with stochastic demand”, for fast moving items from the implementation point of view. The proposed modification of the model leads to a reduction of the expected total cost of the system under certain conditions. They assume that the end-item demand is normally distributed and lead time is deterministic. Also their model has a restriction of being appropriate for only two echelon supply chains. Their model can be extended to multi-stage serial and two-echelon assembly systems (Taskin Gümüs and Güneri, 2007).

The supply chain model developed in this paper remedies several deficiencies of similar models found in the literature. The basic deficiency is that demand and/or lead times are assumed to be constant or to fit a probabilistic distribution. In our model, demand and lead time are determined by neuro-fuzzy calculations, a method that gives realistic results (Section 4.2). Another deficiency of other models is that they apply only to shallow chains with a few echelons and assume that the connections between units at different echelons are serial. The model we have developed is generalized for N-echelons and a tree-structured supply chain. Another assumption of published models is that late orders are delayed until the next order cycle arrives. Our model allows orders that arrive out of phase to be expedited. In this sense it more accurately reflects the real world, where expediting costs are tolerated to provide high service levels. Another innovation of our paper is that we employed a neural network to simulate our neuro-fuzzy forecast based supply chain model. Conventional simulation software (ARENA, SLAM II, etc.) could not be used because our model incorporated neuro-fuzzy forecasting (to see the summary of the innovations and difference of our paper, please look at Table 1, as mentioned in Section 2). Many published models have been subjected to little or no performance analysis. In contrast, we used an industry reference model, the SCOR model, to test the performance of our supply chain model. By building accurate forecast data and realistic cost figures into a general N-echelon tree-structured supply chain, our model and methodology eliminate several deficiencies in published models. In future research, we plan to eliminate the identity assumption for retailers and distributors and to expand the model to handle more than one product type. Finally, we plan to calculate more variables calculated by neuro-fuzzy approximation to further increase the model’s fidelity to the real world.