Explaining Bayesian Optimization by Shapley Values Facilitates Human-AI Collaboration for Exosuit Personalization

Abstract

Bayesian optimization (BO) has become indispensable for black box optimization. However, BO is often considered a black box itself, lacking transparency in the rationale behind proposed parameters. This is particularly relevant in human-in-the-loop applications like personalization of wearable robotic devices. We address BO’s opacity by proposing ShapleyBO , a framework for interpreting BO proposals by game-theoretic Shapley values. Our approach quantifies the contribution of each parameter to BO’s acquisition function (AF). By leveraging the linearity of Shapley values, ShapleyBO can identify the influence of each parameter on BO’s exploration and exploitation behaviors. Our method gives rise to a ShapleyBO -assisted human-machine interface (HMI), allowing users to interfere with BO in case proposals do not align with human reasoning. We demonstrate these HMI’s benefits for the use case of personalizing wearable robotic devices (assistive back exosuits) by human-in-the-loop BO. Results suggest that human-BO teams with access to ShapleyBO outperform teams without access to ShapleyBO . ( Open Science: ShapleyBO as well as code and data to reproduce findings available at https://github.com/rodemann/ShapleyBO . This work builds upon the master’s thesis of the second author supervised by the last author [ 18 ], and substantially extends and formalizes the results presented therein.).