YangSkywalker/Papers-of-Offline-RL

Related papers for offline reforcement learning (we mainly focus on representation and sequence modeling and conventional offline RL)

Papers-of-Offline-RL

Related papers for offline reforcement learning (we mainly focus on representation and sequence modeling and conventional offline RL)

Statement

Anyone is welcomed to submit a pull request for the related and unlisted papers on offline RL, which are published on peer-review conferences (ICML/NeurIPS/ICLR/CVPR etc.) or released on arXiv.

Contents

• Representation Learning (for RL)
  
• Sequence Modeling
  
• Conventional Offline-Reinforcement-Learning
  

Representation Learning (for RL)

• Provable Representation Learning for Imitation with Contrastive Fourier Features

  提出用contrastive learning方法从offline dataset表征state，可以证明在使用该方法优化表征state的函数和强化学习过程时，等价于将学习的policy与收集offline dataset的target policy的距离的upper bound降低，不需要考虑target policy表征的形式，理论上可以将error降到0。相较Behavior Cloning方法（Bisimulation-based）性能大幅提升，实验在tabular case和Atari2600上训练的DQN均做了实验。（learning dynamic and reward）BC对target policy采集数据的形式有限制，且由于数据未全覆盖且空间大sample efficiency比较低

• Pretraining Representations for Data-Efficient Reinforcement Learning

• Representation Matters: Offline Pretraining for Sequential Decision Making

Sequence Modeling

• Reinforcement Learning as One Big Sequence Modeling Problem
• Decision Transformer: Reinforcement Learning via Sequence Modeling[code]

Conventional Offline-Reinforcement-Learning

• Conservative Q-Learning for Offline Reinforcement Learning [code]

  CQL：提出在offline-rl中Q learning过程中加入正则项来缓解over-estimate的问题。通过正则项可以学习到真实Q的lower bound，这样扩大估计值和真实值Q的距离，缓解offline RL中存在的distribution shift问题

• Off-policy deep reinforcement learning without exploration

  BCQ: constrains the mismatch between the state-action visitation of the policy and the state-action pairs contained in the batch by using a state-conditioned generative model to produce only previously seen actions.

• D4RL: Datasets for Deep Data-Driven Reinforcement Learning) [code]

• Stabilizing Off-Policy Q-Learning via Bootstrapping Error Reduction [code]

  BEAR：为环境offline RL设定中从静态数据集中学习存在的数据偏移问题，提出在Bellman方程迭代更新Q过程中，由于action distribution shift导致的累积误差越来越大可能会导致最终不收敛的现象。文章提出了如何限制动作的选择来减轻此问题，提出算法bootstrapping error accumulation reduction (BEAR)，借助support-set matching来避免此问题

• Offline Decentralized Multi-Agent Reinforcement Learning

  MABCQ: 通过对conditional VAE模型的隐向量的约束，实现对Q learning网络从offline学习到online的约束，方法在MA-Mujoco上验证

• Believe What You See: Implicit Constraint Approach for Offline Multi-Agent Reinforcement Learning

  ICQ: 本质上依然是对Q-network做约束，加上mixing网络为ICQ-Ma作为offline MARL的方法

• Boosting Offline Reinforcement Learning with Residual Generative Modeling

  同样使用conditional VAE来约束offline learning过程中学到的policy

• Behavior Constraining in Weight Space for Offline Reinforcement Learning

• Offline Reinforcement Learning with Fisher Divergence Critic Regularization

• Offline Meta-Reinforcement Learning with Online Self-Supervision

• Offline-to-Online Reinforcement Learning via Balanced Replay and Pessimistic Q-Ensemble

• Policy Finetuning: Bridging Sample-Efficient Offline and Online Reinforcement Learning

• S4RL: Surprisingly Simple Self-Supervision for Offline Reinforcement Learning in Robotics

• Mopo: Model-based offline policy optimization

• Combo: Conservative offline model-based policy optimization

• Weighted model estimation for offline model-based reinforcement learning