SJTU 2021Spring CS489 强化学习 Reinforcement Learning
This project contains three algorithms, DQN, PPO, and SAC.
For DQN, supported Atari environments:
- PongNoFrameSkip-v4
- BoxingNoFrameSkip-v4
For PPO and SAC, supported MuJoCo environments:
- Ant-v2
- Hopper-v2
- HalfCheetah-v2
You can run run.py like
python run.py --env_name ENV_NAME --method METHOD (--train)
where you can assign the environment to run by --env_name, and assign the algorithm to use in a MuJoCo environment by --method, like --method PPO or --method SAC (default).
The default use is to load the trained models and test the algorithms in environments, it will run 100 episodes and the mean and variance of the scores will be printed.
If you want to train a model by yourself, you can add --train.
The results can be find in the report.
https://github.com/jmichaux/dqn-pytorch
https://github.com/reinforcement-learning-kr/pg