A Dynamic State Cluster-Based Particle Swarm Optimization Algorithm

Fitness landscape (FL) is an effective tool for describing and analyzing the real-time dynamics of the search process, offering valuable insights into the population’s varying states. In particle swarm optimization for complex optimization challenges, parameter selection significantly influences performance across various population states. However, current methods for constructing fitness landscapes demonstrate insufficient theoretical analysis of state parameters and involve high construction time costs. To address these limitations, this paper introduces a dynamic state cluster-based particle swarm optimization (DSCPSO) algorithm, which employs population phenotypic entropy based on clustering technique. (1) The algorithm provides theoretical splitting points by mathematically analyzing the population into four states: convergence, exploitation, escape, and exploration, enabling more effective parameter adaptive mechanisms. (2) DSCPSO incorporates sinusoidal chaos mapping to dynamically adjust inertia weights, allowing particles to better align with the population’s evolutionary state. (3) During the convergence state, an intelligent particle migration strategy (IPMS) enhances search efficiency within the solution space, preventing unnecessary computational resource consumption. Eventually, comparative analysis with 10 advanced existing algorithms on the CEC2017 and CEC2022 benchmark suites demonstrates that DSCPSO achieves competitive performance across over 70% of the functions, validating the algorithm’s effectiveness and superiority. In addition, the Wilcoxon-test of the algorithm verifies the validity of the algorithm, and also applies the algorithm to a high-dimensional feature selection problem, which demonstrates the ability of the proposed algorithm to solve real-world problems.