Escape Local Minima with Improved Particle Swarm Optimization Algorithm

Abstract

Particle Swarm Optimization (PSO) is a powerful meta-heuristic technique which has been maneuvered to solve numerous complex optimization problems. However, due to its characteristics, there is a possibility to trap all particles in a local minimum in the solution space and then they cannot find the way out from the trap on their own. Therefore, we modify the traditional PSO algorithm by adding an extra step so that it helps PSO to find a better solution than the local minimum that they undesirably found. We perturb all the particles by adjusting parameter values in the traditional algorithm when there is no improvement of the objective value over the training iterations, assuming that particles have stuck in a local minimum. In this research, we mainly focus on adjusting the learning factors. However, the parameter values have to be used in an effective way to perturb the particles. The behavior of the proposed modification and its parameter adjustments are studied using a function which has a large number of local minima - Schwefel’s function. Results show that 2 out of 3 PSO attempts trap in local minimum and slight changes on learning factors do not help them to get out from the traps. However, perturbances made with large learning factors can find better solutions than the local minima that they stuck in and help to find the global minimum eventually.