Evolutionary canonical particle swarm optimizer - A proposal of meta-optimization in model selection

Abstract

We proposed Evolutionary Particle Swarm Optimization (EPSO) which provides a new paradigm of meta-optimization for model selection in swarm intelligence. In this paper, we extend the technique of online evolutionary computation of EPSO to Canonical Particle Swarm Optimizer (CPSO), and propose Evolutionary Canonical Particle Swarm Optimizer (ECPSO) for optimizing CPSO. In order to effectually evaluate the performance of CPSO, a temporally cumulative fitness function of the best particle is adopted in ECPSO as the behavioral representative for entire swarm. Applications of the proposed method to a suite of 5-dimensional benchmark problems well demonstrate the effectiveness. Our experimental results clearly indicate that (1) the proper parameter sets in CPSO for solving various optimization problems are not unique; (2) the values of parameters in them are quite different from that of the original CPSO; (3) the search performance of the optimized CPSO is superior to that of the original CPSO, and to that of RGA/E except for the result to the Rastrigin's benchmark problem. © Springer-Verlag Berlin Heidelberg 2008.