Correct Me If I’m Wrong: Using Non-Experts to Repair Reinforcement Learning Policies
summary
Reinforcement learning policies get patched by synthesizing runtime shields from short queries to non-experts and retraining with those constraints. Implemented in a 2D kitchen, three correction types are tested: action refinement (salad converges ~30% faster), alternative-item (cake converges after ~104 episodes vs fails without a shield), and forbidden-action (coffee converges after ~24 episodes). An online study with 18 non-experts mostly produces usable rules, while wrong shields sometimes create infinite correction loops.
AI-written summary · about