Dynamic Multi-Swarm Particle Swarm Optimization Based on Elite Learning

This paper presents a dynamic multi-swarm particle swarm optimization based on an elite learning strategy (DMS-PSO-EL). In DMS-PSO-EL, the whole evolutionary process is divided into a former stage and a later stage. The former and later stages are focus on the exploration and the exploitation, respectively. In the former stage, the entire population is divided into multiple dynamic sub-swarms and a following sub-swarm according to the particles' fitness values. In each generation, the dynamic sub-swarms evolve independently, which is beneficial for keeping population diversity, while particles in the following sub-swarm choose elites in the dynamic sub-swarms as their learning exemplars aiming to find out more promising solutions. To take full advantages of the different sub-swarms and then speed up the convergence, a randomly dynamic regrouping schedule is conducted on the entire population in each regrouping period. In the latter stage, all the particles select the historical best solution of the entire population as an exemplar aiming to enhance the exploitation ability. The comparison results among DMS-PSO-EL and other 9 well-known algorithms on CEC2013 and CEC2017 test suites suggest that DMS-PSO-EL demonstrates superior performance for solving different types of functions. Furthermore, the sensitivity and performance of the proposed strategies in DMS-PSO-EL are also testified by a set of experiments.