Diagnosing Unconditional Migration in the Snow Ablation Optimizer Through Controlled Retention Strategies

Abstract

The Snow Ablation Optimizer (SAO) uses ranked solutions to generate candidates but accepts every evaluated candidate, even when it degrades its source state. We examined whether this unconditional migration rule created a systematic risk for the individuals that contributed most to search progress. On 12 representative CEC2017 functions, current leaders produced a median 82.51% of generation-best improvements. Their median candidate-degradation rate was 64.26%, compared with 35.72% for followers, and the difference remained significant on all functions after Holm correction. This contribution–risk concentration identifies candidate acceptance and protection target as the relevant correction points. A unified retention rule compared fitness-ranked leader retention (LR), random retention (R), full retention (ALL), and the original SAO without changing candidate generation or the evaluation budget. A 30-seed, five-percentage-point diagnostic grid selected LR-70% and nominal R-95% for the complete benchmark. Across 84 CEC2017 and CEC2022 function–dimension blocks and 19 algorithms, R-95% achieved the lowest overall mean-error rank (3.0476). Its global rank was better than every comparator after Holm correction, including LR-70% and ALL. Median-error ranking and exclusion of CEC2017 F2 preserved this CEC result.