Contrastive Example-Based Control

TL;DR

This paper introduces a novel offline, example-based control method that learns an implicit multi-step transition model to estimate Q-values, outperforming reward-based approaches in various tasks.

Contribution

It proposes a new approach that bypasses reward function learning by modeling multi-step transitions implicitly for offline control.

Findings

Outperforms reward-based methods on multiple offline control tasks.

Demonstrates improved robustness and scalability with dataset size.

Effective on both state-based and image-based tasks.

Abstract

While many real-world problems that might benefit from reinforcement learning, these problems rarely fit into the MDP mold: interacting with the environment is often expensive and specifying reward functions is challenging. Motivated by these challenges, prior work has developed data-driven approaches that learn entirely from samples from the transition dynamics and examples of high-return states. These methods typically learn a reward function from high-return states, use that reward function to label the transitions, and then apply an offline RL algorithm to these transitions. While these methods can achieve good results on many tasks, they can be complex, often requiring regularization and temporal difference updates. In this paper, we propose a method for offline, example-based control that learns an implicit model of multi-step transitions, rather than a reward function. We show that this implicit model can represent the Q-values for the example-based control problem. Across a range of state-based and image-based offline control tasks, our method outperforms baselines that use learned reward functions; additional experiments demonstrate improved robustness and scaling with dataset size.