Chaotic Lévy and adaptive restart enhance the Manta Ray foraging optimizer for gene feature selection

Swarm-based optimization algorithms often face challenges in maintaining an effective exploration–exploitation balance in high-dimensional search spaces. Manta Ray Foraging Optimization (MRFO), while competitive, is hindered by static parameter settings and premature convergence. This study introduces CLA-MRFO, an adaptive variant incorporating chaotic Lévy flight modulation, phase-aware memory, and an entropy-informed restart strategy to enhance search dynamics. On the CEC’17 benchmark suite, CLA-MRFO achieved the lowest mean error on 23 of 29 functions, with an average performance gain of 31.7% over the next best algorithm; statistical validation via the Friedman test confirmed the significance of these results ( $$p < 0.01$$ ). To examine practical utility, CLA-MRFO was applied to a high-dimensional leukemia gene selection task, where it identified ultra-compact subsets ( ≤ 5% of original features) of biologically coherent genes with established roles in leukemia pathogenesis. These subsets enabled a mean F 1 -score of 0.953 ± 0.012 under a stringent 5-fold nested cross-validation across six classification models. While highly effective in a binary classification setting, the method’s performance in a multi-class diagnostic context revealed constraints in generalizability, indicating that the identified biomarkers are highly context-dependent. Overall, CLA-MRFO exhibited consistent behavior (<5% variance across runs) and provides an adaptable framework for high-dimensional optimization tasks with applications extending to bioinformatics and related domains.