With exponential forms for the consequences in neural-network-based fuzzy logic controllers

In order to construct a fuzzy control system, the membership functions for linguistic variables must be identified and the fuzzy control rules must be derived. However, such information heavily depends on expert knowledge which is not always available. Our research intended to develop a general methodology for designing fuzzy logic controller. In our previous work, a methodology of designing fuzzy logic controllers is proposed. This methodology uses the linear combination concept for fuzzy control rules in multilayer feedforward neural networks to implement fuzzy reasoning and computing processes. However, the behavior of fuzzy rules may not always be truly represented by a linear function of the input variables, and thus in certain cases, the learning process may become slow, or even diverge. The basic idea of this paper is to use linear combinations of exponential functions of input variables, instead of only linear functions of input variables, in the consequences of fuzzy rules. In so doing, not only can exponential functions introduce nonlinearity into the representation of fuzzy rules, but also a new parameter freedom will be included to identify the major behavior of fuzzy rules. The result shows that such an approach indeed has better learning ability than an original neural-network-based fuzzy logic controller does. A fuzzy car system is used to demonstrate the robust performance of the proposed control scheme. Computer simulation shows that the resultant neural-network-based fuzzy logic controller can work well not only for the learned patterns but also for unfamiliar patterns