An asynchronous and steady state update strategy for the particle swarm optimization algorithm

Abstract

This paper proposes an asynchronous and steady state update strategy for the Particle Swarm Optimization (PSO) inspired by the Bak-Sneppen model of co-evolution. The model consists of a set of fitness values (representing species) arranged in a network. By replacing iteratively the least fit species and its neighbors with random values (simulating extinction), the average fitness of the population tends to grow while the system is driven to a critical state. Based on these rules, we implement a PSO in which only the worst particle and its neighbors are updated and evaluated in each time-step. The other particles remain steady during one or more iterations, until they eventually meet the update criterion. The steady state PSO (SS-PSO) was tested on a set of benchmark functions, with three different population structures: lbest ring and lattice with von Neumann and Moore neighborhood. The experiments demonstrate that the strategy significantly improves the quality of results and convergence speed with Moore neighborhood. Further tests show that the major factor of enhancement is the selective pressure on the worst, since replacing the best or a random particle (and neighbors) yields inferior results.