Core Concepts
Bayesian optimization with Shapley values enhances human-AI collaboration by providing interpretability and rationale behind optimization decisions.
Abstract
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.
Stats
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.