Safe Reinforcement Learning via Shielding

About

Reinforcement learning algorithms discover policies that maximize reward, but do not necessarily guarantee safety during learning or execution phases. We introduce a new approach to learn optimal policies while enforcing properties expressed in temporal logic. To this end, given the temporal logic specification that is to be obeyed by the learning system, we propose to synthesize a reactive system called a shield. The shield is introduced in the traditional learning process in two alternative ways, depending on the location at which the shield is implemented. In the first one, the shield acts each time the learning agent is about to make a decision and provides a list of safe actions. In the second way, the shield is introduced after the learning agent. The shield monitors the actions from the learner and corrects them only if the chosen action causes a violation of the specification. We discuss which requirements a shield must meet to preserve the convergence guarantees of the learner. Finally, we demonstrate the versatility of our approach on several challenging reinforcement learning scenarios.

Mohammed Alshiekh, Roderick Bloem, Ruediger Ehlers, Bettina Könighofer, Scott Niekum, Ufuk Topcu · 2017

Related benchmarks

Task	Dataset	Result
Autonomous Driving	CARLA	Collision Rate26.7	8
Reinforcement Learning	Minepump 1.0 (train)	Reward-806.73	7
Reinforcement Learning	Minepump 1.0 (evaluation)	Reward-669.16	7
Replay Suppression	250-node graphs Policy-frozen RSD	RAG0.78	7
Reinforcement Learning	ZonesEnv	Reward15.92	6
Safe Reinforcement Learning	LTC controller Full oracle access	Σδ0	5
Reinforcement Learning	Continuous Cartpole	Average Final Reward537.76	5
Atari game reinforcement learning	Atari Seaquest Expected Environment (train)	Training Reward224.03	5
Navigation	Two Dimensional Navigation environment (10 seeds)	Final Reward (Avg)782.23	5
Atari game reinforcement learning	Atari Seaquest Unexpected Environment (evaluation)	Reward0	5