شبکه های عصبی یکپارچه و تجزیه و تحلیل پوششی داده ها برای ارزیابی تامین کنندگان تحت اطلاعات ناقص

Supplier evaluation and selection are critical decision making processes that require consideration of a variety of attributes. Several studies have been performed for effective evaluation and selection of suppliers by utilizing several techniques such as linear weighting methods, mathematical programming models, statistical methods and AI based techniques. One of the successful evaluation methods proposed for this purpose is data envelopment analysis (DEA), that utilizes techniques of mathematical programming to evaluate the performance of a set of homogeneous decision making units, when multiple inputs and outputs need to be considered. It is often complicated, costly and sometimes impossible to acquire all necessary information from all potential suppliers to attain a reasonable set of similar input and output values which is an essential for DEA. The purpose of this study is to explore a novel integration of neural networks (NN) and data envelopment analysis for evaluation of suppliers under incomplete information of evaluation criteria.

Tough rules of competition require quick response to rapid market changes while maintaining profitability. Effective management of collaborative supply chain networks (SCNs) is a fruitful solution to increase the efficiency of production and service performances and competitiveness of companies because of their direct impact on a variety of final product dimensions such as cost, product and design quality, and producibility (Talluri, 2002). Once a well-managed and established supply chain is developed, it has a lasting effect on the competitiveness of the entire supply chain (Chen, Lin, & Huang, 2006) in that selecting right suppliers has a significant effect on improving corporate competitiveness (Boer, Labro, & Morlacchi, 2001). As a result, supplier evaluation has become more critical and effective and robust evaluation methods are required for measuring the supplier performance. Supplier evaluation is a complicated and difficult process as a consequence of possibly conflicting multiple criteria, involvement of many alternatives and internal and external constraints imposed on the buying process (Jayaraman, Srivastava, & Benton, 1999). Therefore, a wide range of techniques which include conceptual, empirical, and modeling approaches have been applied in various studies to address the issue of supplier evaluation.

In order to cope with the drawbacks associated with homogeneity and accuracy assumption of DEA, we propose a supplier evaluation method by integrating neural networks with data envelopment analysis to reduce the general evaluation criteria set into a performance measures set. The aim of this method is to come up with a set of common attributes that all decision units share and are compared upon. This paper presents a novel approach to supplier evaluation problem with an integrated NN–DEA for supplier evaluation. The main contribution of the method is its ability of working with incomplete information of evaluation attributes, which is a common problem in real life situations. Also, depending on the Gallant’s (1993) heuristic of neural networks’ inference ability, an acceptability index is proposed for calculating the confidence measures of the results obtained from neural network systems. Also, to the knowledge of authors, this is the first study in literature in which neural networks are used as a data aggregation method for data envelopment analysis. An illustrative example is demonstrated as an application of supplier evaluation in an automotive manufacturer by the integrated NN–DEA method. It is observed that integrated system performs better than plain DEA analysis in reflecting management opinions for given example where some of the information on supplier evaluation criteria is incomplete. However, we can mention that the future line of investigation will consist of testing generalization capabilities of integrated system by extending the application range and increasing the number of examples for each case. Also, other substitution techniques for replacing unknown data and other methods for handling the missing inputs of neural networks, like interval input vectors or modified back-propagation, will be investigated for choosing best performing method. The future stage of this research is focused on analyzing biases in efficiency scores caused by input aggregation by neural networks to provide a complete basis for use of neural networks for input aggregation for DEA.