CLA-MRFO: A Synergistic Chaotic Lévy and Adaptive Restart Variant of Manta Ray Foraging Optimizer for Gene Feature Selection

Swarm-based optimization algorithms often struggle to balance exploration and exploitation in complex, high-dimensional search spaces. While Manta Ray Foraging Optimization (MRFO) has demonstrated competitive performance in recent studies, its effectiveness remains constrained by rigid parameter settings and phase-specific stagnation. This paper introduces the synergistic Chaotic Lévy and Adaptive Restart variant of MRFO (CLA-MRFO), which integrates chaotic dynamics, Lévy-driven step-size modulation, and phase-aware memory into a cohesive framework. Unlike in-cremental hybrids, CLA-MRFO enables dynamic interaction among its components: chaotic maps adaptively regulate Lévy flight (LF) behavior to improve global exploration; memory mechanisms preserve diverse elite solutions across search phases; and an entropy-informed restart strategy injects diversity when stagnation is detected, without disrupting convergence stability.The proposed method is evaluated on both the CEC’17 benchmark suite and a real-world, high-dimensional gene expression dataset for leukemia classification. CLA-MRFO outperforms eight recent metaheuristic algorithms, including MRFO, PSO, and GBO variants, across the majority of benchmark functions. In the context of gene feature selection (FS), CLA-MRFO successfully identifies compact, discriminative gene subsets that yield robust classification performance across six machine learning models. Using a 5-fold nested cross-validation protocol, the selected features achieved average F1-scores exceeding 0.95. Statistical analyses confirm that these improvements are both stable and significant.These results position CLA-MRFO as a robust, generalizable, and interpretable optimizer for both synthetic benchmarks and real-world biomedical FS tasks.