NeurIPS · 2025

Reinforcement Learning with Imperfect Transition Predictions: A Bellman-Jensen Approach

Chenbei Lu, Zaiwei Chen, Tongxin Li, Chenye Wu, Adam Wierman

Advances in Neural Information Processing Systems

How can reinforcement learning use imperfect forecasts beyond one step?

A Bayesian value function and Bellman Jensen gap quantify the value of imperfect transition forecasts. BOLA separates offline value learning from online adaptation, with sample efficiency analysis and experiments in synthetic environments and wind energy storage control.

Research themes

Connections

Algorithms guide learned decision makers

Q value advice and imperfect transition forecasts provide structured information for reinforcement learning. LEAD applies a related principle by supporting language agents with dueling bandit algorithms.

Open questions

Prediction quality as a decision resource

Which prediction errors actually matter for the decision being made?

Explore open directions

Cite this paper

Chenbei Lu, Zaiwei Chen, Tongxin Li, Chenye Wu, Adam Wierman. Reinforcement Learning with Imperfect Transition Predictions: A Bellman-Jensen Approach. Advances in Neural Information Processing Systems, 2025. https://papers.nips.cc/paper_files/paper/2025/hash/940a7634dab556b67af15bacd337f7db-Abstract-Conference.html

BibTeX
@inproceedings{tongxin-bellman-jensen,
  title = {{Reinforcement Learning with Imperfect Transition Predictions: A Bellman-Jensen Approach}},
  author = {Chenbei Lu and Zaiwei Chen and Tongxin Li and Chenye Wu and Adam Wierman},
  year = {2025},
  booktitle = {Advances in Neural Information Processing Systems},
  url = {https://papers.nips.cc/paper\_files/paper/2025/hash/940a7634dab556b67af15bacd337f7db-Abstract-Conference.html}
}
2025ACL Findings

Beyond Numeric Rewards: In-Context Dueling Bandits with LLM Agents

Fanzeng Xia, Hao Liu, Yisong Yue, Tongxin Li

The study identifies a gap between quick preference discovery and sustained exploitation by language agents. LEAD combines LLM reasoning with dueling bandit algorithms to obtain weak and strong regret guarantees, with evaluations under noisy and adversarial prompts.

Can language agents learn reliably from pairwise preferences? large language models LLM agents contextual control world models reinforcement learning dueling bandits learning augmented algorithms algorithms with predictions competitive analysis robustness consistency online optimization
2023NeurIPS

Beyond Black-Box Advice: Learning-Augmented Algorithms for MDPs with Q-Value Predictions

Tongxin Li, Yiheng Lin, Shaolei Ren, Adam Wierman

Q value predictions expose more information than an opaque policy recommendation. For a single trajectory MDP, the analysis characterizes consistency and robustness tradeoffs and shows how structured advice improves the guarantees available from a robust baseline.

Does the structure of advice improve robust decision making? learning augmented algorithms algorithms with predictions competitive analysis robustness consistency online optimization large language models LLM agents contextual control world models reinforcement learning dueling bandits