Query-Based Learning for Dynamic Particle Swarm Optimization

Abstract

In recent years, many researchers have examined dynamic optimization problems (DOPs). The key challenge lies in the fact that the optimal solution of a DOP typically changes over time. This paper focuses on using query-based learning dynamic particle swarm optimization (QBLDPSO) to solve DOPs. QBLDPSO is mainly used for improving multi-population-based PSO; our QBL mechanism includes two learning strategies that integrate the concepts of diversity and memory into PSO. The first learning strategy, QBL quantum parameter adaptation (QBLQPA), is used to apply the concept of diversity to the multi-population based algorithm. This is different from typical diversity-based PSO approaches, which passively maintain the diversity of particles in the solution space. We actively adapt the ratio of quantum particles and neutral particles to achieve diversity without analyzing the distribution of optima in the solution space. The second learning strategy is query-based learning optima prediction (QBLOP). Although QBLOP exploits the concept of memory, we do not need to analyze the history of all particles. We select the $k$ nearest particles to the current best solution and use a minimum encompassing circle as the possible prediction region. Our experimental results are based on the generalized dynamic benchmark generator (GDBG), which is adopted as a benchmark for the DOP. The proposed method outperforms two state-of-the-art multi-population-based PSO methods with the average improvements of 11.37% and 8% using QBLQPA. In particular, for the recurrent problems in GDBG, our method improves performance by 35.06%.