In recent years many investigations have shown that Particle Swarm Optimization (PSO) is a very competitive global optimization heuristic. However, in very complex state spaces the classical PSO algorithm converges too fast and hence provides only suboptimal results. Looking at swarm robotics it seems natural to adopt a repulsive force to avoid this undesired behavior as suggested in Charged PSO but the downside of this is the problem of final convergence in static applications. The contribution of this paper is to introduce a dynamic charge reduction over time defining particle groups which are iteratively merged, reducing the number of charged particles during the optimization run. A visualization of this process shows spontaneous formation of independent particle groups, redolent very much of swarm movement in nature. Optimization results are superior compared to other PSO approaches especially in very complex high dimensional search spaces.
展开▼