Human-Aware Reinforcement Learning for Fault Recovery Using Contextual Gaussian Processes

This work addresses the iterated nonstationary assistant selection problem, in which over course of repeated interactions on a mission, an autonomous robot experiencing fault must select single human from among group assistants to restore it operation. The our problem have level performance that changes as function their experience solving problem. Our approach uses reinforcement learning via multi-arm bandit formulation learn about capabilities each potential and decide task. study, is built past work, evaluates for Gaussian-process-based machine method effectively model complex dynamics associated with forgetting. Application simulation shows capable tracking human-like Using novel policy called proficiency window, shown technique can outperform baseline strategies while providing guarantees use. offers effective alternative dedicated supervisors, application any human–robot system where set humans responsible overseeing operations.