Evolving metaheuristic: MsPSO with dynamic parameters and heterogeneous sub-swarms

Abstract Particle Swarm Optimization (PSO) often suffers from premature convergence when addressing complex or multimodal optimization problems, limiting its effectiveness in classical engineering applications. To overcome these shortcomings, this study introduces the Enhanced Multi-Swarm Adaptive and Cooperative Particle Swarm Optimization (EMsPSO). The algorithm partitions the population into four heterogeneous sub-swarms, each employing distinct and time varying control parameters inertia weights and acceleration coefficients to achieve a dynamic balance between exploration and exploitation. Three complementary mechanisms are integrated: a self adaptive strategy for dynamic search behavior, a cooperative information-exchange strategy enabling bidirectional interactions, and a diversity-enhancing operator to prevent stagnation. The EMsPSO algorithm is first evaluated on the CEC2014 benchmark test suite, using a set of 12 representative functions (including unimodal, multimodal, and hybrid types) in 30 dimensions, and its performance is rigorously compared with six state of the art multi-swarm PSO variants. Furthermore, to ensure a comprehensive validation, EMsPSO is also tested alongside basic metaheuristic algorithms and other recent excellent metaheuristics on 10 challenging functions from the CEC2021 benchmark suite, with statistical significance verified using Wilcoxon and Friedman tests. According to the results, EMsPSO achieves the best mean fitness values across all tested functions and demonstrates superior stability . Moreover, the EMsPSO algorithm is applied to three classical engineering design problems: vehicle side impact design, piston lever design, and tension/compression spring design. The results confirm that EMsPSO delivers optimal designs with objective values of 22.8429713803, 8.41269832311, and 0.0126653057 for VSID, PLD, and TCSD respectively and consistently outperforms all compared methods in solution quality and robustness.