Multi-swarm Particle Swarm Optimization with Multiple Learning Strategies


Inspired by the division of labor and migration behavior in nature, this paper proposes a novel particle swarm optimization algorithm with multiple learning strategies (PSO-MLS). In the algorithm, particles are divided into three sub-swarms randomly while three learning strategies with different motivations are applied to each sub-swarm respectively. The Traditional Learning Strategy (TLS) inherits the basic operations of PSO to guarantee the stability. Then a Periodically Stochastic Learning Strategy (PSLS) employs a random learning vector to increase the diversity so as to enhance the global search ability. A Random Mutation Learning Strategy (RMLS) adopts mutation to enable particles to jump out of local optima when trapped. Besides, information migration is applied within the intercommunication of sub-swarms. After a certain number of generations, sub-swarms would aggregate to continue search, aiming at global convergence. Through these learning strategies and swarm aggregation, PSO-MLS possesses both good exploration and exploitation abilities. PSO-MLS was tested on a set of benchmarks and the result shows its superiority to gain higher accuracy for unimodal functions and better solution quality for multimodal functions when compared to some PSO variants.

Genetic and Evolutionary Computation Conference, GECCO 2014, Vancouver, BC, Canada