Accelerating Safe Reinforcement Learning with Constraint-mismatched Policies

2020/06/20 by Tsung-Yen Yang, Justinian Rosca, Yang, Tsung-Yen +5 · 1 citation
Computer Science · Engineering · Mathematics · #Adversarial Robustness in Machine Learning #Artificial Intelligence (cs.AI) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Reinforcement Learning in Robotics #Robot Manipulation and Learning #Robotics (cs.RO) #cs.AI #cs.LG #cs.RO #stat.ML

paper · pdf · doi:10.48550/arxiv.2006.11645

International Conference on Machine Learning (ICML) 2021

openalex publication_date 2020/06/20 · openalex created_date 2020/06/25 · arxiv created 2021/07/10 · arxiv updated 2021/07/13 · openalex updated_date 2026/07/28

Abstract

We consider the problem of reinforcement learning when provided with (1) a baseline control policy and (2) a set of constraints that the learner must satisfy. The baseline policy can arise from demonstration data or a teacher agent and may provide useful cues for learning, but it might also be sub-optimal for the task at hand, and is not guaranteed to satisfy the specified constraints, which might encode safety, fairness or other application-specific requirements. In order to safely learn from baseline policies, we propose an iterative policy optimization algorithm that alternates between maximizing expected return on the task, minimizing distance to the baseline policy, and projecting the policy onto the constraint-satisfying set. We analyze our algorithm theoretically and provide a finite-time convergence guarantee. In our experiments on five different control tasks, our algorithm consistently outperforms several state-of-the-art baselines, achieving 10 times fewer constraint violations and 40% higher reward on average.

Citations

• Constrained Markov Decision Processes
• Playing Atari with Deep Reinforcement Learning
• Trust Region Policy Optimization
• Generative Adversarial Imitation Learning
• Agent-Agnostic Human-in-the-Loop Reinforcement Learning
• Leveraging Demonstrations for Deep Reinforcement Learning on Robotics Problems with Sparse Rewards
• Safe Exploration in Continuous Action Spaces
• Reinforcement Learning from Imperfect Demonstrations
• Lyapunov-based Safe Policy Optimization for Continuous Control
• Extrapolating Beyond Suboptimal Demonstrations via Inverse Reinforcement Learning from Observations
• A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning
• Deep Q-learning from Demonstrations
• Benchmarking Batch Deep Reinforcement Learning Algorithms
• Learning to be Safe: Deep RL with a Safety Critic

Cited by

• Learning safe, constrained policies via imitation learning: Connection to Probabilistic Inference and a Naive Algorithm

Related

• Incentivizing Safer Actions in Policy Optimization for Constrained Reinforcement Learning
• A Survey of Constraint Formulations in Safe Reinforcement Learning
• Constrained Reinforcement Learning Under Model Mismatch
• Safe and Balanced: A Framework for Constrained Multi-Objective Reinforcement Learning
• Constrained Policy Optimization