Multiple-Strategies dung beetle optimizer and its applications in engineering optimization and bankruptcy prediction

Engineering optimization and hyperparameter tuning in machine learning represent distinct classes of complex optimization problems: the former is characterized by numerous constraints and multimodal landscapes, while the latter demands identifying optimal configurations from a vast combinatorial space. This paper proposes a novel algorithm based on the Dung Beetle Optimizer (DBO), termed Multiple-Strategies DBO (MSDBO), which integrates three targeted improvement strategies: (1) an adaptive ball-rolling and spawning strategy that dynamically balances exploration and exploitation, (2) an optimal boundary control strategy that constrains search agents within feasible regions, and (3) a foraging enhancement strategy that strengthens local search to escape premature convergence. To comprehensively evaluate MSDBO, we benchmark it against the original DBO, two enhanced DBO variants, and eight well-established metaheuristic algorithms on the CEC2017 suite (30, 50, and 100 dimensions) and the CEC2022 suite (10 and 20 dimensions). The Friedman mean rank test validates the statistically superior performance of MSDBO, and the Wilcoxon rank-sum test confirms its significant superiority at the 0.05 significance level. Ablation studies on 23 classical benchmark functions verify that each integrated strategy contributes critically to the overall performance gain. For real-world engineering applications, MSDBO is applied to three constrained design problems-speed reducer design, pressure vessel design, and step-cone pulley design-and consistently yields lower-cost solutions than 11 competing algorithms. To further validate its efficacy in machine learning hyperparameter optimization, we develop an MSDBO-tuned Kernel Extreme Learning Machine (MSDBO-KELM) model for corporate bankruptcy prediction. On this task, MSDBO-KELM achieves a classification accuracy of 82% and a specificity of 85%, outperforming all comparison algorithms across four metrics:accuracy, Matthews correlation coefficient, sensitivity, and specificity. These results collectively demonstrate that MSDBO is an efficient, robust, and broadly applicable optimizer for both constrained engineering design and data-driven prediction tasks. The source code is publicly available at https://github.com/Dedai-Wei/MSDBO.git.