RL agent for the Google Football environment

This code implements a bare-bones version of the Proximal Policy Optimization (PPO) algorithm for the purpose of training an AI bot to play the game of football. The tutorial series for learning step-by-step implementation of this algorithm can be found in video format here or in a blogpost here.

Setup instructions

Tested on Ubuntu 18.04 and a single NVIDIA GPU.

1. Get the Google Research Football environment up and running using these instructions. This repository uses the gpu version of tensorflow/gfootball.
2. Install Keras using pip3 install Keras.

Training

1. Execute python3 train.py script to start the PPO training loop.

Render on remote display server

To render the game screen on a remote display (eg. if using Google Colab), execute the instructions in display_server.sh. For more information, check out this thread.

Acknowledgements

1. PPO-Keras
2. PPO-PyTorch
3. Roboschool environment PPO tutorial