The classifier judged this item not relevant to AI security, so it is left out of lists, feeds, the API and dataset releases. If that is wrong, report a correction on this page.
InfoResearchPeer-reviewed
Model-Based Offline Reinforcement Learning With Adversarial Data Augmentation
- Published
- Record updated
Summary
The paper introduces MORAL, a model-based offline reinforcement learning method that replaces fixed-horizon rollouts with adversarial data augmentation. An adversarial process selects ensemble models against the policy for biased sampling, and a differential factor regularizes extrapolation error. The authors report that MORAL outperforms other model-based offline RL methods on the D4RL benchmark in policy learning and sample efficiency.