Optimization of Recurrent NN by GA with Variable Length Genotype

Abstract

The gradient based learning algorithms for complex hybrid neurofuzzy architectures have a lot of local minima and a seriously time consumption complexity is involved as a consequence. Genetic Algorithms with variable length genotypes are successfully used in getting better performances for systems with complex structure or, at the same performances, a less complex structure of the system. We propose a sophisticated algorithm that solves simultaneously the optimization objectives of learning algorithms in fuzzy recurrent neural networks: both regarding the fuzzy NN performances (by its fuzzy weights matrix) and regarding the architecture (number of fully connected neurons). In this paper we developed a genetic algorithm with variable length genotypes that offers a systematic way of getting a minimal neuro-fuzzy structure satisfying the above mentioned requested performance. This advantage is not to be neglected when a complex hybrid intelligent architecture must be designed without any previous details regarding it requested architecture.