zolastro / DeepGamer

Development of automated players by deep reinforcement learning for videogames

DeepGamer

This project presents a deep learning model able to learn how toperform several tasks using the 1993 game DOOM as environment.The agent is trained using raw pixels from the game screen and uses a deep learning variant of the Q-learning algorithm. Several optimizations techniques were applied in order to maximize performance and results.

Clone the repo

You can clone this repo via https using the following command:

https://github.com/zolastro/DeepGamer.git

Run this project

Get dependencies

The only major dependencies that you have to manually install are used for VizDoom. Follow their guide on how to install all the required dependencies.

Set up environment

The best way to get all needed dependencies is to use Anaconda. To install and initialize the environment, run the following commands:

conda env create -f src/environment.yml
conda activate deep-gamer

Run the agent

There are two scenarios to train the agent. To run them, go into the src folder and run:

python basic.py

or

python defend_the_center.py

Results

This graph shows the results of the agent in both scenarios.

Basic

Defend the center