Hyperparameter Tuning in Offline Reinforcement Learning

In this work, we propose a reliable hyper-parameter tuning scheme for offline reinforcement learning. We demonstrate our proposed scheme using the simplest antmaze environment from the standard benchmark offline dataset, D4RL. The usual approach for policy evaluation in offline reinforcement learning involves online evaluation, i.e., cherry-picking best performance on the test environment. To mitigate this cherry-picking, we propose an ad-hoc online evaluation metric, which we name "median-median-return". This metric enables more reliable reporting of results because it represents the expected performance of the learned policy by taking the median online evaluation performance across both epochs and training runs. To demonstrate our scheme, we employ the recently state-of-the-art algorithm, IQL, and perform a thorough hyperparameter search based on our proposed metric. The tuned architectures enjoy notably stronger cherry-picked performance, and the best models are able to surpass the reported state-of-the-art performance on average.