Preselection and diversity in PSO

Abstract Particle swarm optimization (PSO) is a popular metaheuristic algorithm that is theoretically well understood, but its full potential does not yet seem to have been realized. While diversity is a critical issue in all population-based algorithms, it is measurable in a natural way in PSO. It is more easily interpretable in terms of performance, such that the option of diversity control arises. Fitness evaluations (FEs), in standard PSO, contribute to convergence towards optimality only if a personal or global improvement was actually made. Thus, some FEs can be avoided if a particle is sufficiently unlikely to achieve an improvement. The interesting outcome is a relation between diversity and the ratio of unnecessary FEs. While high diversity implies a highly exploratory regime where nearly all FE are necessary, we observe that with FE savings, better performance can also be reached for the same total number of FEs. All observations show a strong parameter dependence. However, parameter scans show that the effect is prevalent at parameters near the optima for a given problem. In contrast, the method discussed here is largely ineffective for parameters that entail poor performance.