Genetic Algorithm for Fuzzy Neural Networks using Locally Crossover

Fuzzy feed-forward (FFNR) and fuzzy recurrent networks (FRNN) proved to be solutions for "real world problems". In the most cases, the learning algorithms are based on gradient techniques adapted for fuzzy logic with heuristic rules in the case of fuzzy numbers. In this paper we propose a learning mechanism based on genetic algorithms (GA) with locally crossover that can be applied to various topologies of fuzzy neural networks with fuzzy numbers. The mechanism is applied to FFNR and FRNN with L-R fuzzy numbers as inputs, outputs and weights and fuzzy arithmetic as forward signal propagation. The a-cuts and fuzzy biases are also taken into account. The effectiveness of the proposed method is proven in two applications: the mapping a vector of triangular fuzzy numbers into another vector of triangular fuzzy numbers for FFNR and the dynamic capture of fuzzy sinusoidal oscillations for FRNN.