Summary of Q-value Regularized Transformer For Offline Reinforcement Learning, by Shengchao Hu et al.
Q-value Regularized Transformer for Offline Reinforcement Learning
by Shengchao Hu, Ziqing Fan, Chaoqin Huang, Li Shen, Ya Zhang, Yanfeng Wang, Dacheng Tao
First submitted to arxiv on: 27 May 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper proposes a new approach to offline reinforcement learning (RL) that combines the strengths of two existing techniques: Conditional Sequence Modeling (CSM) and Dynamic Programming (DP). The proposed method, Q-value regularized Transformer (QT), uses a value function to approximate optimal future returns for each state, which helps to stabilize learning in long-horizon and sparse-reward scenarios. QT learns an action-value function and incorporates a term that maximizes action-values into the training loss of CSM, aiming to seek optimal actions that align with the behavior policy. The authors evaluate their method on D4RL benchmark datasets and show that it outperforms traditional DP and CSM methods. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Offline reinforcement learning is important for machines to learn from past experiences without interacting with the environment. Researchers have been developing new algorithms to improve this area, especially Conditional Sequence Modeling (CSM) which learns action distributions based on history trajectories and target returns. However, CSM has limitations when stitching together optimal trajectories from sub-optimal ones. The paper proposes a solution called Q-value regularized Transformer (QT), which combines the strengths of two existing techniques: CSM and Dynamic Programming (DP). QT uses a value function to approximate optimal future returns for each state and learns an action-value function that aligns with the behavior policy. |
Keywords
» Artificial intelligence » Reinforcement learning » Transformer
