Fuzzy rules generation using new evolutionary algorithms combined with multilayer perceptrons

Based on evolutionary algorithms (EAs) and multilayer perceptrons (MLPs), a fuzzy rules generation method inclusive of two main learning stages is presented in this paper. In the primary stage, a new EA is developed to generate numerical control rules from input-output data without the help of experts, which increases the diversity of individuals to reduce the opportunities of falling into local optima. Every generated numerical rule is accumulated in a lookup table called a numerical-rule-based controller (NRC). In the secondary stage, both antecedent and consequent variables of the numerical rules are fuzzified by training MLPs with the backpropagation algorithm. All training data are directly derived from the NRC with simple manipulations. Consequently, a linguistic-rule-based controller (LRC) consisting of the generated fuzzy rules is completed. Two illustrative experiments are successfully made on the computer simulation and hardware implementation of the NRCs and LRCs of different types using the new EA combined with the MLPs. The experimental results reveal that the proposed EA-MLP MLP approach is efficient and effective to generate fuzzy rules which control nonlinearly dynamical systems exceedingly well.