Identification of Nonlinear System Based on a New Hybrid Gradient-Based PSO Algorithm

Abstract

Because the number of iterations necessary to locate the global best solution is not known a priori, it's problematic to make a proper choice of inertial weights omega and constriction coefficients l of particle swarm optimization (PSO) algorithm in advance. The existing PSO algorithms are sensitive to the above two parameters. In addition, standard PSO algorithms convergence slowly and coarsely in the latter period. A new hybrid PSO algorithm is proposed to overcome the above shortcomings. The new algorithm utilizes original PSO algorithm for locating approximately a good local minimum, and then a conjugate gradient based local search is done with the best solution found by the PSO algorithm as its starting point for finding local minimum accurately. A new optimization circle begins with the accurate local minimum as global best particle. The simulation results show that the new algorithm convergences more fast and accurately than GA. It also shows better performance than GA in identifying the parameters of RBF neural networks.