Safe Imitation Learning via Fast Bayesian Reward Inference from Preferences

Abstract

Bayesian reward learning from demonstrations enables rigorous safety and uncertainty analysis when performing imitation learning. However, Bayesian reward learning methods are typically computationally intractable for complex control problems. We propose Bayesian Reward Extrapolation (Bayesian REX), a highly efficient Bayesian reward learning algorithm that scales to high-dimensional imitation learning problems by pre-training a low-dimensional feature encoding via self-supervised tasks and then leveraging preferences over demonstrations to perform fast Bayesian inference. Bayesian REX can learn to play Atari games from demonstrations, without access to the game score and can generate 100,000 samples from the posterior over reward functions in only 5 minutes on a personal laptop. Bayesian REX also results in imitation learning performance that is competitive with or better than state-of-the-art methods that only learn point estimates of the reward function. Finally, Bayesian REX enables efficient high-confidence policy evaluation without having access to samples of the reward function. These high-confidence performance bounds can be used to rank the performance and risk of a variety of evaluation policies and provide a way to detect reward hacking behaviors.

In proceedings ICML 2020

References in corpus (21)

• Adam: A Method for Stochastic Optimization
• What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?
• Tutorial on Variational Autoencoders
• Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization
• Rainbow: Combining Improvements in Deep Reinforcement Learning
• On the Complexity of Solving Markov Decision Problems
• Cooperative Inverse Reinforcement Learning
• A Connection between Generative Adversarial Networks, Inverse Reinforcement Learning, and Energy-Based Models
• A Simple Baseline for Bayesian Uncertainty in Deep Learning
• Causal Confusion in Imitation Learning
• AI Safety Gridworlds
• Functional Variational Bayesian Neural Networks
• DART: Noise Injection for Robust Imitation Learning
• Safe Policy Improvement by Minimizing Robust Baseline Regret
• Fast and Scalable Bayesian Deep Learning by Weight-Perturbation in Adam
• A Divergence Minimization Perspective on Imitation Learning Methods
• Beyond Confidence Regions: Tight Bayesian Ambiguity Sets for Robust MDPs
• Impossibility and Uncertainty Theorems in AI Value Alignment (or why your AGI should not have a utility function)
• Risk-Sensitive Generative Adversarial Imitation Learning
• Safety Augmented Value Estimation from Demonstrations (SAVED): Safe Deep Model-Based RL for Sparse Cost Robotic Tasks
• Learning Safe Policies with Expert Guidance

Cited by in corpus (19)

• A Review of Robot Learning for Manipulation: Challenges, Representations, and Algorithms
• What are you optimizing for? Aligning Recommender Systems with Human Values
• Learning Submodular Objectives for Team Environmental Monitoring
• Deception in Social Learning: A Multi-Agent Reinforcement Learning Perspective
• Example-Driven Model-Based Reinforcement Learning for Solving Long-Horizon Visuomotor Tasks
• The MAGICAL Benchmark for Robust Imitation
• SIRL: Similarity-based Implicit Representation Learning
• Robust Imitation Learning from Noisy Demonstrations
• Autonomous Assessment of Demonstration Sufficiency via Bayesian Inverse Reinforcement Learning
• Inverse Decision Modeling: Learning Interpretable Representations of Behavior
• A Mapping of Assurance Techniques for Learning Enabled Autonomous Systems to the Systems Engineering Lifecycle
• Offline Preference-Based Apprenticeship Learning
• LazyDAgger: Reducing Context Switching in Interactive Imitation Learning
• Bayesian Robust Optimization for Imitation Learning
• Inverse Contextual Bandits: Learning How Behavior Evolves over Time
• Is the Rush to Machine Learning Jeopardizing Safety? Results of a Survey
• Safety Aware Reinforcement Learning (SARL)
• A Symmetric Loss Perspective of Reliable Machine Learning
• Value Alignment Verification