Particle Swarm Optimization

Particle Swarm Optimization (PSO) is an evolutionary algorithm based off of swarm intelligence and is used as a stochastic optimization technique. One major characteristic of PSO is many configuration parameters, which allow the algorithm to be adjusted to various problem landscapes. In the first half of this paper, we empirically benchmark performance of different parameterizations of PSO in a fractal-like test environment for convergence and robustness against local optima. Experiments were done with most salient parameters: dampening factor, neighborhood size and dimensionality. We evaluate a proposed technique of eliminating the known problem of dimensional collapse and position bias, as well as propose a novel optimization technique for restarting particles which have prematurely converged to increase robustness against local optima. PSO was then successfully applied to an image matching algorithm called Evolisa, using a combination of hill-climbing and PSO. PSOlisa also employs several optimization techniques to improve convergence efficiency.