Hao Jin, Yang Peng, Wenhao Yang, Shusen Wang and Zhihua Zhang. Federated Reinforcement Learning with Environment Heterogeneity. AISTATS, 2022.
- PyTorch: version 1.1.0 is required
- OpenAI Gym:
pip install gym gym[atari] - TensorBoardX:
pip install tensorboardX - PTAN: install from sources (delete torch==1.7.0 which is unnecessary in requirements.txt)
- MuJoCo: only for HalfCheetah and Hopper environment
The customized environments RandomMDPs and WindyCliffs are implemented in utils.py and GridWorldEnvironment.py respectively.
Some utility functions and standard tabular reinforcement learning algorithms are included in utils.py.
The algorithms QAvg, SoftPAvg and ProjPAvg are implemented for both the environments in the python files with the corresponding prefixes.
The python files with prefix heter are for the experiments in Table 1 in the paper.
The customized environments are implemented in MyCartPole.py, MyAcrobot.py, MyHalfCheetah.py and MyHopper.py.
Some utility functions and standard deep reinforcement learning algorithms are included in DeepRLAlgo.py.
The algorithms DQNAvg and DDPGAvg are implemented in DQNAvg.py and DDPGAvg.py respectively, we apply these methods in the customized environments in the python files with the corresponding prefixes.
The personalized version of the above is implemented in the python files with prefix Per.