Evolving Fuzzy Neural Networks by Particle Swarm Optimization with Fuzzy Genotype Values

Particle swarm optimization (PSO) is a well-known instance of swarm intelligence algorithms and there have been many researches on PSO. In this paper, the author proposes an extension of PSO for solving fuzzy-valued optimization problems. In the proposed extension, genotype values (i.e. values in particle position vectors) are not real numbers but fuzzy numbers. Search processes in PSO are extended so that PSO can handle genotype instances with fuzzy numbers. The proposed method is experimentally applied to evolution of neural networks with fuzzy weights and biases. Experimental results showed that fuzzy neural networks evolved by the proposed method could model hidden target fuzzy functions despite the fact that no training data was explicitly provided.