Cell autophagy optimization algorithm based on cellular self-degradation and recycling for global optimization

Abstract Every metaheuristic has to balance exploring new regions of the search space against refining the solutions it has already found, and getting that balance right across very different landscapes is hard. We address this with the Cell Autophagy Optimization Algorithm (CAO), which draws on cellular autophagy, the stress-triggered process by which a cell breaks down and recycles its own damaged components, and whose molecular machinery was recognised by the 2016 Nobel Prize in Physiology or Medicine. CAO frames optimization as a cycle of autophagic flux built from seven cooperating operators: chaotic tent-map seeding, opposition-based learning, an adaptive stress threshold that routes each agent individually, lysosomal degradation of weak solutions, Lévy-flight reconstruction guided by an elite archive, opposition-based population renewal, and a mitophagy-inspired escape from stagnation. We evaluate CAO on the CEC2017 suite (29 functions at $$D = 10$$ , $$D = 30$$ and $$D = 50$$ ) and the CEC2022 suite (12 functions, $$D = 20$$ ), comparing it against fifteen algorithms, among them the state-of-the-art DE variants jSO and L-SHADE and three recently proposed methods (LAOA, FGO, OOA), over 30 independent runs per setting. CAO places 3rd of 16 at $$D = 50$$ (mean Friedman rank 4.66) and 4th of 16 at $$D = 30$$ (rank 5.66), and its standing improves as dimensionality grows, beating jSO on 9 of 29 functions at $$D = 50$$ . On three constrained engineering design problems, solved with both penalty and $$\\varepsilon $$ -constrained formulations, it records the best mean on the seven-variable Speed Reducer, ahead of every competitor including jSO and L-SHADE. An ablation study points to Lévy-flight recycling as the single most important component, removing it degrades 25 of 29 functions, while a sensitivity analysis finds CAO largely indifferent to its hyperparameters (four of six are low-sensitivity). All comparisons are supported by Wilcoxon signed-rank and Friedman tests at $$\\alpha = 0.05$$ .