An Optimisation Framework for Unsupervised Environment Design

TL;DR

This paper introduces an optimization framework for unsupervised environment design in reinforcement learning, providing theoretical guarantees and demonstrating improved robustness of agents across diverse environments.

Contribution

It offers a new optimization-based approach with provable convergence for UED, advancing beyond prior methods reliant on convergence guarantees.

Findings

Outperforms prior UED methods in various environments

Provides theoretical convergence guarantees for the proposed algorithm

Enhances agent robustness in high-risk settings

Abstract

For reinforcement learning agents to be deployed in high-risk settings, they must achieve a high level of robustness to unfamiliar scenarios. One method for improving robustness is unsupervised environment design (UED), a suite of methods aiming to maximise an agent's generalisability across configurations of an environment. In this work, we study UED from an optimisation perspective, providing stronger theoretical guarantees for practical settings than prior work. Whereas previous methods relied on guarantees if they reach convergence, our framework employs a nonconvex-strongly-concave objective for which we provide a provably convergent algorithm in the zero-sum setting. We empirically verify the efficacy of our method, outperforming prior methods in a number of environments with varying difficulties.