Neuro-fuzzy approaches to decision making: A comparative study with an application to check authorization

Abstract

This paper describes and compares several nonlinear decision-making systems, including multilayer perceptrons, wavelet neural networks, polynomial neural networks, and fuzzy decision models. The applicability of these systems is illustrated through the problem of check authorization from incomplete data. A benchmark is established in terms of classical linear discriminant analysis and Bayes quadratic classification, in order to assess the need for the neuro-fuzzy strategies. An overall improvement of around 10 percentage points in classification accuracy on an independent test set is demonstrated for each of the neuro-fuzzy models over conventional statistical techniques. In addition to classification accuracy, five performance measures are reported: accuracy in dollar terms, robustness, parametric efficiency, training computational expense, and classification balance. Even though each system performs differently on these measures, any neuro-fuzzy model is recommended over traditional techniques in problems such as check authorization, where the improvement in reliability warrants the added cost of implementation.