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 enable

Stats

• citations0
• S2 citations—
• github stars0
• HF likes0
• heat score0.00
• arxiv keybrown2020safe

Related papers

• Reward Learning From Human Preferences And Demonstrations In Atari (2018)0.00
• Bayesian Robust Optimization For Imitation Learning (2020)0.00
• A Bayesian Solution To The Imitation Gap (2024)0.00
• Learning Safe Policies With Expert Guidance (2018)0.00
• Backplay: "man Muss Immer Umkehren" (2018)0.00
• Basis For Intentions: Efficient Inverse Reinforcement Learning Using Past Experience (2022)0.00
• Monte Carlo Augmented Actor-critic For Sparse Reward Deep Reinforcement Learning From Suboptimal Demonstrations (2022)0.00
• Optimizing Neurorobot Policy Under Limited Demonstration Data Through Preference Regret (2026)0.00