A Transfer Learning-Based Adaptive Particle Swarm Optimization for Global Optimization Problem

Particle swarm optimization (PSO) is a popular evolutionary algorithm and widely used to solve practical engineering problems. However, most of the existing methods search for the optimal solution starting from zero initial information, and do not make use of the historical information obtained when solving similar problems previously. This will cause the method to waste lots of computing resources to a certain extent. Recently, the idea of transfer learning has received widespread attention. Transfer learning is a humanized machine learning method that aims to transfer knowledge from one domain (source domain) to another domain (target domain) so that the target domain can achieve better learning results. Therefore, in this study, the idea of transfer learning is extended to the field of evolutionary optimization, and a transfer learning-based adaptive particle swarm optimization framework is proposed (TAPSO). Firstly, the adaptive clustering model matching strategy (ACMS) is proposed to find the historical problems matching the target problem. In ACMS, density-based clustering strategy and maximum mean discrepancy work together to find historically similar problems for the current problem. Secondly, the adaptive knowledge transfer strategy (AKTS) is used to transfer knowledge from the original problem to the target problem. Finally, the comprehensive learning particle swarm optimization algorithm is embedded into the transfer learning framework proposed in this study. Extensive experiments have confirmed the effectiveness of the proposed transfer learning-based adaptive particle swarm optimization framework.