AMPA-FS: An Adaptive Multi-strategy Marine Predators Algorithm for Wrapper-based Feature Selection

Feature selection (FS) is essential for high-dimensional classification. The Marine Predators Algorithm (MPA) couples L\'evy and Brownian random walks in a three-phase foraging model, yet its binary variant (BMPA) suffers from (i) over-aggressive step-size decay during exploration, (ii) uneven search-space coverage from uniform initialization, and (iii) no explicit mechanism for escaping local optima. This paper proposes AMPA-FS, integrating three synergistic modules: (1) Tent chaotic map initialization with elite opposition-based learning; (2) a phase-aligned adaptive step-size control factor matching MPA's three phases; and (3) a stagnation-triggered elite-guided differential mutation operator. Experiments on ten UCI datasets against nine binary metaheuristics show that AMPA-FS achieves 5 wins, 5 ties, and 0 losses against BMPA in classification accuracy with no regression on any dataset, ranks 3rd in wrapper-fitness Friedman rank among all ten algorithms, and selects 17--28% fewer features than swarm-based competitors on datasets with D ≥30. Vargha--Delaney effect-size analysis confirms that AMPA-FS provides small-to-medium practical improvements over the majority of competitors on the wrapper-fitness objective. A complete 2 ³ −1 ablation study quantifies each module's individual and synergistic contribution, showing that the full model achieves the best average fitness rank across all variants.