/PRR

Meta-Reinforcement Learning with Policy Residual Representation

Primary LanguagePython

PRR

PRR code for the paper

WenJi Zhou, Yang Yu, Yingfeng Chen, Kai Guan, Tangjie Lv, Changjie Fan, Zhi-Hua Zhou. Reinforcement Learning Experience Reuse with Policy Residual Representation. In: Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI'19), Macao, China.

Requirement

  • python 3.6
  • rllab
  • numpy == 1.16
  • tensorflow==1.8.0
  • gym==0.13.0
  • scipy==1.3.0
  • matplotlib==3.1.1
  • theano==1.0.4
  • cached_property==1.5.1

Files

  • main.py is the PRR demo on three FetchTheKey environments in the paper
  • envs contains test environment
  • nets network structures (modified from https://github.com/pat-coady/trpo)
  • rl contains agent and policy codes

Run Demo

  • To run experiment on FetchTheKey environment, just running:
python main.py