SLDALA: Enhancing the Artificial Lemming Algorithm Through Anatomy-aware Integration of Social Learning and Differential Evolution

The Artificial Lemming Algorithm (ALA) demonstrates strong energy-controlled search dynamics but suffers from a precise structural limitation: its random candidate selection ($$\:{x}_{rand}$$ in exploration) and spiral center placement ($$\:{x}_{center}$$ in exploitation) carry no directional intelligence, discarding the population’s accumulated search history at every step. This paper proposes SLDALA, an anatomy-aware integration framework whose central contribution is identifying these stochastic degrees of freedom within ALA’s movement equations and redesigning Social Learning (SL) and Differential Evolution (DE) to inhabit them - rather than replacing ALA’s proven dynamics. In the exploration phase, a multi-elite weighted social compass biases ALA’s inherent random candidate selection without replacing it; in the exploitation phase, DE differential information displaces ALA’s spiral center toward more promising regions while the spiral geometry remains intact. This anatomy-aware injection preserves ALA’s convergence properties while substantially enhancing its directional intelligence. The framework instantiates into two problem-specific variants: SLDALA104 for discrete optimization, where SL and DE provide directional guidance that augments ALA’s movements under a continuous relaxation encoding shared by all comparison algorithms; and SLDALAH for continuous optimization, which extends the shared injection architecture with strategy-specific adaptive machinery (dual SHADE memory banks, JADE-inspired archive, CR-controlled dimension-level displacement). A static problem-type identifier activates the appropriate variant at initialization. On the full CEC2017 benchmark suite at 50D and 100D, SLDALA improves upon ALA by an average of 17% across all functions, with improvements exceeding 50% on multimodal and hybrid functions; on complex high-dimensional functions, improved search path quality reduces convergence time to match or undercut ALA’s runtime, and the gap to SHADE and L-SHADE narrows to within 10% on a subset of complex functions. On feature selection tasks spanning 9 to 12,600 features, SLDALA achieves 91.06% average accuracy and 98.8% feature reduction, ranking first on all five ultra-high-dimensional biomedical datasets against 20 competitors. These results establish SLDALA as an effective anatomy-aware enhancement of ALA with particular strengths in complex high-dimensional optimization.