Quantifying Generalization in Reinforcement Learning
This addresses the issue of poor generalization in RL for researchers and practitioners, though it is incremental as it applies known techniques from supervised learning to a new benchmark.
The paper tackles the problem of overfitting in deep reinforcement learning by introducing a new procedurally generated benchmark called CoinRun, and finds that agents overfit even to large training sets, with improvements from deeper architectures and supervised learning methods like regularization and data augmentation.
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.