跨维协方差适应在障碍约束搜索中引发向欺骗吸引子的不可逆漂移
Cross-Dimensional Covariance Adaptation Induces Irreversible Drift Toward Deceptive Attractors in Barrier-Constrained Search
研究概述
研究障碍约束优化中 CMA-ES 的协方差诱导不可逆漂移失效模式。
原文摘要（英文）
We report covariance-induced irreversible drift, a failure mode of CMA-ES in barrier-constrained optimization. Full CMA-ES systematically transitions from feasible to infeasible solutions through cross-dimensional covariance adaptation (73% drift rate vs 43% for sep-CMA-ES, Fisher p=0.018). The underlying landscape contains empirically disconnected feasible basins linked by directed, optimizer-dependent transitions. Population scaling experiments show complete separation (100% vs 0%) at lambda=200. Adaptive mitigation via reactive switching fails; preemptive variant selection is required.