Direct Policy Gradients - RL with A* sampling and direct optimization

Pytorch impelentation for MiniGrid and DeepSea experiments from the paper "Direct Policy Gradients: Direct Optimization of Policies in Discrete Action Spaces"

Installation

clone repository, create new virtualenv and install dependencies:

git clone https://github.com/GuyLor/reinforcement_learning.git
python3 -m venv direct_rl
source direct_rl/bin/activate
cd reinforcment_learning
pip3 install -r requirements.txt

Usage

train from scratch:

python run.py --train

let the trained policy to "play" after training:

python run.py --train --play 

save and/or load the model after training:

  python run.py --train --play --save_path my_policy_model_new.pkl --load_path my_policy_model.pkl

open tensorboard:

  tensorboard --logdir logs

License

MIT