Empirical study on learning in fuzzy systems

The authors examine the ability of fuzzy systems to function as approximators of nonlinear mappings by computer simulations on real-life data. The relation among six factors in a sensory test on rice taste is modeled by fuzzy systems with five input variables and a single output variable. Fuzzy if-then rules with nonfuzzy singletons in the consequent part are employed in fuzzy systems. A learning rule based on a descent method is applied to the consequent part of each fuzzy if-then rule. By a random subsampling technique, the performance of fuzzy systems for test data and training data is compared with that of multilayer neural networks. A simple method for specifying initial fuzzy if-then rules is proposed to improve the performance of fuzzy systems