Abstract: The Marine Predators Algorithm (MPA), while promising for complex optimization, suffers from limited solution precision, imbalanced exploration–exploitation, and premature convergence. To address these shortcomings, this paper proposes a phased- enhancement variant named PEMPA, which integrates three novel strategies into distinct phases of MPA: 1) embedding historical best positions in the high-velocity ratio phase to refine solution quality; 2) introducing an adaptive inertia weight based on an inverted Sigmoid function in the unit-velocity ratio phase to systematically balance exploration and exploitation; and 3) designing a two-stage opposition-based learning operator in the low-velocity ratio phase to prevent premature convergence. The performance of PEMPA is comprehensively evaluated across 23 classical benchmark functions, the IEEE Congress on Evolutionary Computation (CEC) 2017 test suite, 21 feature selection tasks, and a real-world medical insurance fraud detection problem. Experimental results confirm that the proposed strategies significantly enhance the efficiency and robustness of MPA. Furthermore, PEMPA demonstrates highly competitive performance compared with several state-of-the-art metaheuristic algorithms, validating its effectiveness and scalability for diverse optimization challenges.

Key words: Marine predators algorithm')">Marine predators algorithm, Opposite-based learning, Inertia weight, Numerical optimization, Featureselection