Deep Reinforcement Learning Methods for Google Research Football RL Environment

Update: Auto-Encoder Architecture for the Deep Q-Network

I improved my model focusing on the DL part of this project for the Deep Learning course taught by prof. Alessio Ansuini from the Data Science & Artificial Intelligence M.Sc Course at the University of Studies of Trieste (UNITS).
It started as a project for the Reinforcement Learning course taught by Prof. Antonio Celani.

This repo is a fork from the Google Research Football Repository and aims to implement a Deep Q-Network to train an Agent to play on its own.

For further information, please refer to the two PowerPoint presentation in this repo, which provides a comprehensive overview of the theoretical framework, algorithmic explanations, training structure and results achieved (unfortunately is not possible to see the videos inside of the powerpoint, but you can find them in the folder DQN/episode_observations or DQN/episode_observations_DL)

Configuration

For Linux machines, a bash script is provided to facilitate the cloning of this repository. It is hoped that this script will function as intended, although it is possible that some packaging installations may have been overlooked in the conda environment. This is unlikely to present a significant issue.