Reproduce “Reinformer: Max-Return Sequence Modeling for Offline RL” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zifeng Zhuang, Dengyun Peng, Jinxin Liu, Ziqi Zhang, Donglin Wang
Reinformer introduces a max-return sequence modeling approach for offline reinforcement learning that directly integrates the objective of maximizing returns into the transformer framework. The method, built upon expectile regression, predicts in-distribution maximum returns at each step, enabling significantly enhanced trajectory stitching on D4RL benchmarks, such as a 37.07% improvement on Maze2d-v1 over prior sequence models and strong gains on Antmaze datasets.
Key insights
- Explicitly optimizing for return maximization within sequence models, rather than just conditioning on specified returns, is crucial for effective offline RL.
- Expectile regression provides a theoretically grounded method to predict in-distribution maximum returns, preventing out-of-distribution issues common with naive high-return targets.
- Dynamically predicting the maximum achievable return at each step during inference, rather than using a fixed initial target, significantly improves trajectory stitching and overall performance in offline RL tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.