Policy-Gradient-PyTorch

Implementation of vanilla stochaistic (categorical) policy gradient algorithm to play cartpole.
Vanilla policy gradient takes longer but convergence is smoother than DQN for the cartpole, both of these properties as expected.

Training

python ./vpg_pytorch.py 

vpg_pytorch.py trains model, saves the checkpoint for every 1000 episode, and saves well-trained model's weights.

• path for checkpoint file ./Save/YY-MM-DD-hh:mm:ss/vpg_cp_ep[#ep].pth
• weight file: ./Save/YY-MM-DD-hh:mm:ss/vpg_weight_ep[#ep].pth

Testing

python test_vpg_pytorch.py <directory>
python test_vpg_pytorch.py "Save/2021-03-18-15:42:31"

This test every saved weight file (vpg_weight_ep*.pth) under the directory for 100 episodes.

Running with rendering

./run_learnt_model.py <file>
./run_learnt_model.py "Save/2021-03-18-15:42:31/vpg_weight_ep9970.pth"

This executes the cartpole env with rendering and show you how the learnt model actually works.