Jiajun Fan

Critic PI2: Master Continuous Planning via Policy Improvement with Path Integrals and Deep Actor-Critic Reinforcement Learning

In brief

What the paper does.

Critic PI2 is a model-based reinforcement learning framework for continuous control, where tree-based planners such as AlphaGo and MuZero struggle. It combines trajectory optimization through policy improvement with path integrals, deep actor-critic learning and a learned model, and shows that planning with a critic improves sample efficiency and real-time performance on inverted-pendulum and other continuous domains.

Published in
arXiv preprint
Paper
arxiv.org/abs/2011.06752

Cite

Cite the paper.

@article{fan2020criticpi2,
  title={Critic PI2: Master Continuous Planning via Policy Improvement with Path Integrals and Deep Actor-Critic Reinforcement Learning},
  author={Fan, Jiajun and Ba, He and Guo, Xian and Hao, Jianye},
  journal={arXiv preprint arXiv:2011.06752},
  year={2020}
}