Skip to content
Loic Nyssen edited this page Oct 14, 2018 · 8 revisions

This wiki details how to launch and run NEAT (Naturally Evolving Augmented Topology) experiments against Super Mario Bros - as well as providing a base testing framework for developing and running your own experiments and training algorithms.

While we have focused heavily on NEAT in our experimentation the setup is wholly based around OpenAI so it should be relatively easy to extend the system to use any form of network; provided it can leverage the OpenAI interface.

It is expected that the user is comfortable with and can install and use the following technologies:

  • Docker - required for setup
  • Python - language used for framework and training scripts
  • OpenAI Gym - base framework used to manage training scripts

Getting started

To get started you will need to follow the setup guide.

Note: This guide is geared against *nix based systems and has been tested mainly against Ubuntu Linux and macOS systems. Some extra work will likely be required to make this work on Windows or other versions of Linux.

Experiments

Within this project we have included four different experiments that can be launched within the Docker container.

Training scripts

The section on training networks can be found here.

Utility scripts

The section on utility scripts can be found here.

You can also view an example of a trained network playing World 1 Level 1 here.

Clone this wiki locally