Memory-enhanced artificial rabbits optimization for solving global optimization and feature selection

Modern artificial intelligence (AI) systems are increasingly challenged by high-dimensional and multimodal data streams, which exacerbate the curse of dimensionality and significantly increase the difficulty of feature selection. To address the limitations of the original Artificial Rabbits Optimization (ARO) algorithm—such as insufficient population diversity, loss of high-quality historical solutions, inadequate local search accuracy, and susceptibility to premature convergence—this paper proposes a Memory-enhanced Artificial Rabbits Optimization (MARO) algorithm. The proposed method incorporates an individual memory archive mechanism, an adaptive differential evolution operator, and a directed local search strategy, thereby effectively balancing global exploration and local exploitation. Furthermore, MARO is integrated with the K-nearest neighbor (KNN) classifier to construct a wrapper-based feature selection framework, termed MARO-KNN. Experimental results on the CEC2017 benchmark functions under 30- and 100-dimensional settings show that MARO achieves Friedman mean ranks of 1.40 and 1.20, respectively, significantly outperforming eight mainstream optimization algorithms, including GWO, PSO, and the original ARO. In feature selection experiments conducted on 22 publicly available datasets of varying scales, MARO-KNN attains an average rank of 2.83 in classification accuracy and 2.67 in the number of selected features, both ranking first. These results demonstrate that the proposed framework not only achieves superior overall performance, but also realizes optimal feature dimensionality reduction. Overall, MARO exhibits clear advantages in convergence speed, solution accuracy, and stability, providing an efficient and robust approach for complex high-dimensional global optimization and feature selection tasks.