A clean and robust Pytorch implementation of Categorical DQN(C51) ：

Render	Training curve

Other RL algorithms by Pytorch can be found here.

gymnasium==0.29.1
matplotlib==3.8.2
numpy==1.26.1
pytorch==2.1.0

python==3.11.5

python main.py

where the default enviroment is 'CartPole'.

If you want to train on different enviroments, just run:

python main.py --EnvIdex 1

The --EnvIdex can be set to be 0 and 1, where

'--EnvIdex 0' for 'CartPole-v1'  
'--EnvIdex 1' for 'LunarLander-v2'

Note: if you want to play on LunarLander, you need to install box2d-py first. You can install box2d-py via: pip install gymnasium[box2d]

python main.py --EnvIdex 0 --render True --Loadmodel True --ModelIdex 60 # Play with CartPole

python main.py --EnvIdex 1 --render True --Loadmodel True --ModelIdex 320 # Play with LunarLander

You can use the tensorboard to record anv visualize the training curve.

• Installation (please make sure Pytorch is installed already):

pip install tensorboard
pip install packaging

• Record (the training curves will be saved at '\runs'):

python main.py --write True

• Visualization:

tensorboard --logdir runs

For more details of Hyperparameter Setting, please check 'main.py'

Bellemare M G, Dabney W, Munos R. A distributional perspective on reinforcement learning[C]//International conference on machine learning. PMLR, 2017: 449-458.