Quantifying Generalization in Reinforcement Learning

TL;DR

This paper investigates overfitting in deep reinforcement learning by using procedurally generated environments, introduces CoinRun as a new benchmark, and demonstrates that certain architectures and regularization methods improve generalization.

Contribution

It introduces CoinRun, a novel benchmark environment for assessing RL generalization, and evaluates various architectures and regularization techniques for improving generalization.

Findings

Agents overfit to large training sets in CoinRun

Deeper convolutional architectures enhance generalization

Regularization methods like dropout and data augmentation improve RL generalization

Abstract

In this paper, we investigate the problem of overfitting in deep reinforcement learning. Among the most common benchmarks in RL, it is customary to use the same environments for both training and testing. This practice offers relatively little insight into an agent's ability to generalize. We address this issue by using procedurally generated environments to construct distinct training and test sets. Most notably, we introduce a new environment called CoinRun, designed as a benchmark for generalization in RL. Using CoinRun, we find that agents overfit to surprisingly large training sets. We then show that deeper convolutional architectures improve generalization, as do methods traditionally found in supervised learning, including L2 regularization, dropout, data augmentation and batch normalization.