An Approach Toward Multiobjective Optimization Problems in Hysteresis Neural Networks

Shinya Kujirai, Toshimichi Saito

Abstract: This paper studies a multiobjective optimization problem in continuous-time recurrent neural networks. For simplicity, we use a simple recurrent neural network: a hysteresis associative memory characterized by a binary hysteresis activation function and ternary cross-connection parameters. The optimization problem is based on two objectives. The first objective evaluates the memory accuracy and the second objective evaluates connection sparsity. In order to analyze the optimization problem, we present a simple evolutionary algorithm with growing connection structure. Applying the algorithm to typical examples, we have obtained a Pareto front that guarantees existence of a trade-off between the two objectives. The trade-off provides basic information for the system dynamics and becomes a criterion to optimize the system parameters.