A reinforcement learning-based communication topology in particle swarm optimization

Recently, a multitude of researchers have considered the fully connected topology (Gbest) as a default communication topology in particle swarm optimization (PSO). Despite many earlier studies of this issue indicating that the Gbest might favor unimodal problems, the topology with fewer connections, e.g., Lbest, might perform better on multimodal problems. It seems that different topologies make PSO a problem-related algorithm, while in this paper a problem-free PSO which integrates a reinforcement learning method has been proposed, referred to as QLPSO. In the new proposed algorithm, each particle acts as an agent independently, selecting the optimal topology under the control of Q-learning (QL) during each iteration. Two variants of QLPSO consider the different dimensions of the communication topology, respectively. In order to investigate the performance of QLPSO, experiments on 28 CEC 2013 benchmark functions are carried out when comparing with static and dynamic topologies. The reported computational results show that the proposed QLPSO is more superior compared with several state-of-the-art methods.