A Cluster-Driven Dual-Population Differential Evolution Algorithm and Its Applications

Abstract To overcome the limitations of traditional differential evolution (DE) in highdimensional, multimodal, and constrained continuous optimization problems, This paper proposes ClusterDual-DE, a cluster-driven dual-population DE framework. ClusterDual-DE maintains two complementary subpopulations: (i) a cluster-guided exploitation subpopulation that performs adaptive niching/clustering to intensify local search within promising basins, and (ii) an explorationoriented subpopulation that promotes global search via large-step mutation and a dimension-adaptive restart strategy. The two subpopulations cooperate through three mechanisms: success-history-based parameter adaptation, elite migration with shared memory, and function-evaluation-budget-driven population reduction, which jointly regulate information exchange and allocate computational resources between exploration and exploitation. Extensive experiments on the CEC 2014 and CEC 2017 benchmark suites (10-100 dimensions) show that ClusterDual-DE achieves the lowest average rank in the Friedman test and under the Wilcoxon signed-rank test (e.g., p< 0.05), with multiple-comparison correction if applicable. ClusterDual-DE outperforms nine representative stateof-the-art DE variants in solution quality, robustness, and convergence efficiency. The generality of ClusterDual-DE is further demonstrated in two classical engineering applications (pressure vessel design and automotive side-impact optimization). In addition, This paper evaluates a cooperative unmanned underwater vehicle (UUV) path-planning task. The task involves coupled constraints from 3-D seafloor obstacles, wave effects, and sonar conditions. ClusterDual-DE achieves a lower objective cost and a shorter planning time than nine competing algorithms. The results indicate good adaptability in complex coupled environments.