Parallel PPO-PyTorch

February 6, 2020 ยท View on GitHub

A parallel agent training version of Proximal Policy Optimization with clipped objective.

Usage

  • To test a pre-trained network : run test.py
  • To train a new network : run parallel_PPO.py
  • All the hyperparameters are in the file, main function

Results

CartPole-v1LunarLander-v2
cartpolelander

Dependencies

Trained and tested on:

Python 3.6
PyTorch 1.3
NumPy 1.15.3
gym 0.10.8
Pillow 5.3.0

TODO

  • implement Conv net based training

Setting up Conda Environment

  • conda env export | grep -v "^prefix: " > environment.yml to export the file environment.yml
  • conda create -f environment.yml to create the conda environment used for training

References