Enhanced particle swarm optimization through external memory support

Abstract

A particle swarm optimization strategy using an external memory of partial position and velocity vectors for the purpose of achieving better and faster search capabilities is introduced. Partially complete position and velocity vectors stored in memory are segments cut from the two components of promising solutions over a number of previous iterations, where the size and location of segments are selected completely at random. Elements of external memory (segments) are also associated with their parents' fitness values that are used in retrieving the stored elements. After every iteration, the worst k% of the swarm population is considered and position and velocity vectors of each particle in this subpopulation are partly modified by memory elements retrieved using a fitness-based selection procedure. To update the memory, randomly-sized and randomly located segments cut from the best m% of the current swarm population replaces those memory elements with the worst fitness values. The proposed approach is used for the solution of several benchmark numerical optimization problems for which the obtained results demonstrate that both the speed and solution quality are improved compared to conventional PSO algorithms.