核心概念
Bayesian optimization with Shapley values enhances human-AI collaboration by providing interpretability and rationale behind optimization decisions.
摘要
Bayesian optimization (BO) with Gaussian processes is crucial for black box optimization problems. ShapleyBO framework interprets BO's proposals using game-theoretic Shapley values, aiding in human-machine interaction. The method disentangles contributions to exploration and exploitation, enhancing trust and efficiency in AI systems.
統計資料
BO is optimized with confidence bound (CB) as an acquisition function.
ShapleyBO quantifies each parameter's contribution to the acquisition function.
Contributions are dissected into mean optimization and uncertainty reduction.
Epistemic and aleatoric uncertainties are disentangled using Shapley values.