Enforcing ethical goals over reinforcement-learning policies

Abstract Recent years have yielded many discussions on how to endow autonomous agents with the ability make ethical decisions, and need for explicit reasoning transparency is a persistent theme in this literature. We present modular transparent approach equip comply prescriptions, while still enacting pre-learned optimal behaviour. Our relies normative supervisor module, that integrates theorem prover defeasible deontic logic within control loop of reinforcement learning agent. The operates as both an event recorder on-the-fly compliance checker w.r.t. external norm base. successfully evaluated our several tests using variations game Pac-Man, subject variety “ethical” constraints.