sebastian-schuchmann / self-play-tictactoe-ai-ml-agents- Goto Github PK

View Code? Open in Web Editor NEW

A Self Play reinforcement learning Agent learns to play TicTacToe using the ML-Agents Framework in Unity.

C# 100.00%

self-play-tictactoe-ai-ml-agents-'s Introduction

Self-Play TicTacToe AI (ML-Agents)

Versions

Unity: 2020.1.0f1+
ML-Agents: 1.0.3+ (Release 5)
ML-Agents (Python): 0.18.1
TensorFlow: 2.3.0

Quickstart

Inference

A pretrained model is already included. Start up Unity, press play and try to win (It's possible).

Training

Make sure you have the ML-Agents Python package installed. For guidance, check out this page. This project was tested with Python 3.6.8 on MacOS Catalina with the ML-Agents Python Package v.0.18.1 installed.

First open up the terminal and cd into the "runs" folder of the repo.

cd Self-Play-TicTacToe-AI-ML-Agents-/ML-Agents\ runs

Then start the ML-Agents trainer referencing the trainer config located in this repo ("ML-Agents config/TicTacToe.yaml").

mlagents-learn ../ML-Agents\ config/TicTacToe.yaml --run-id="TicTacToe-0"

Now press play in the Unity Editor and the training should start. A script automatically reduces the animation and graphic settings to ensure a better training performance. Depending on your machine, you should achieve good results after 1 to 3 hours of training.

I want to use an updated version of ML-Agents

This project will probably work with the next few release versions of ML-Agents if no major changes occur. Make sure you have the latest ML-Agents Python package installed. Then, update the ML-Agents package through Unity's package manager if to prevent any version discrepancies between the package and the python communicator.