Automatic Reward Shaping from Confounded Offline Data

TL;DR

This paper introduces a robust deep reinforcement learning algorithm that effectively learns policies from biased, confounded offline data, outperforming standard methods in complex Atari games with unobserved confounders.

Contribution

It proposes a novel deep RL algorithm that is resilient to confounding biases in offline data, ensuring safer policy learning in high-dimensional environments.

Findings

Consistently outperforms standard DQN in confounded Atari games.

Effective in environments with unobserved confounders and input mismatches.

Demonstrates robustness to biases in complex, high-dimensional settings.

Abstract

A key task in Artificial Intelligence is learning effective policies for controlling agents in unknown environments to optimize performance measures. Off-policy learning methods, like Q-learning, allow learners to make optimal decisions based on past experiences. This paper studies off-policy learning from biased data in complex and high-dimensional domains where \emph{unobserved confounding} cannot be ruled out a priori. Building on the well-celebrated Deep Q-Network (DQN), we propose a novel deep reinforcement learning algorithm robust to confounding biases in observed data. Specifically, our algorithm attempts to find a safe policy for the worst-case environment compatible with the observations. We apply our method to twelve confounded Atari games, and find that it consistently dominates the standard DQN in all games where the observed input to the behavioral and target policies mismatch and unobserved confounders exist.