The development of neural network models by revised particle swarm optimization

Abstract

A novel training paradigm for artificial neural networks had been developed and presented in this article. In the proposed approach, a revised version of particle swarm optimization (PSO) had been employed to find out the optimal connection weights of feed-forward artificial neural networks for given training sets. Literatures reported that conventional particle swarm optimization could easily get stuck at local optima, especially in problem domains with high dimensionality. In our scheme, a re-seeding mechanism will be invoked when the system is under the risk of converging to pre-mature solutions. The incorporation of the concept of mutation had endowed the systems with better capability in escaping local optima and approaching to the global optimum. A series of experiments were conducted to verify the feasibility and effectiveness of the proposed approach, and optimistic results were obtained as expected. In additions, the impact and influence of different parameter settings on system performance was investigated through comprehensive empirical study, as reported in this paper.