An Architecture Design Method of Modular Dynamical Neural Networks Using Genetic Algorithms

Abstract

In this paper, we propose an evolutionary approach to architecture design of modular dynamical neural networks. As one of modular dynamical neural networks, we adopt Cross-Coupled Hopfield Nets (CCHN) in which plural Hopfield networks are coupled to each other. The architecture of CCHN is represented by some structural-parameters such as the number of modules, the numbers of units per module, the module connectivity, and so forth. In the proposed design method, these structural-parameters are treated as phenotype of an individual, and suitable modular architecture is searched through the evolution of its genetic representation (genotype) by using genetic algorithms. Based on a simple direct coding method, the order of length of genetic representation for the structural-parameters can be estimated to be O(N²) where N is the total number of units. On the other hand, the order of genetic representation proposed here is O(N). To verify the usefulness of proposed method, we apply a CCHN to associative memories. Here, the fitness of an individual is defined so as to be larger when a CCHN has a simpler architecture as well as when the association performance is higher. As the result of simulations, we certify that the proposed design method can find high-performance CCHN with simple modular architectures.