Lanctot, S. Dieleman, D. Grewe, J. Nham, N. Kalchbrenner, I. Sutskever, T. Lillicrap, M. Leach, K. Kavukcuoglu, T. Hassabis Mastering the game of Go with deep neural networks and tree search.

We use optional third-party analytics cookies to understand how you use GitHub. You can always update your selection by clicking Cookie Preferences at the bottom of the page.

For more information, see our Privacy Statement. We use essential cookies to perform essential website functions, e.

We use analytics cookies to understand how you use our websites so we can make them better, e. Skip to content. MIT License. Dismiss Join GitHub today GitHub is home to over 50 million developers working together to host and review code, manage projects, and build software together.

Sign up. GitHub is where the world builds software Millions of developers and companies build, ship, and maintain their software on GitHub — the largest and most advanced development platform in the world.

Sign up for free Dismiss. Go back. Launching Xcode If nothing happens, download Xcode and try again. Latest commit. Bump superagent from 1.

Git stats commits. Failed to load latest commit information. Nov 18, Fix TF graph issue Sep 10, Mar 29, Update demo bot model for Keras 2.

Mar 18, May 5, Mar 1, Mar 13, Install tensorflow in travis environment. Mar 4, Apr 20, Apr 5, MIT license.

Apr 11, May 6, Oct 28, Apr 14, Created a conda environment file for a working copy Sep 12, Apr 13, Update model saving to use a temp file.

Include Demo Script. Feb 24, Add six to deps in setup. May 4, Sep 9, Mar 20, View code. Getting started Test BetaGo by running the following commands.

Consider contributing by: Adding new models to the model zoo. Writing new Go data processing functionality. Adding more complex models and bots.

How can I run my own bot? Motivation Being both a passionate and mediocre Go player and programmer, this project is a matter of honor to me. Anyway, here you go: Literature [1] A.

Releases 3 tags. Sponsor this project. Packages 0 No packages published. It downloads Go games for you, preprocesses them, trains a model on data, for instance a neural network using keras, and serves the trained model to an HTML front end, which you can use to play against your own Go bot.

Test BetaGo by running the following commands. It should start a playable demo in your browser! This bot plays a reasonable moves, but is still very weak.

You can modify and extend any of the steps outlined above and help decrease the gap between AlphaGo and BetaGo, tear down walls and disrupt the establishment.

Consider contributing by:. Training and serving a bot can be done in just a few steps. The following example uses a convolutional neural network implemented in keras, but you are free to choose other libraries as well.

The code for this example can be found in the examples folder. We start by defining a Go data processor, which downloads an preprocesses Go games. A regular Go board consists of 19 times 19 fields.

Next, we train a neural network to predict moves. If you insist, you may call it a policy network. This example is just one of many possible architectures to tackle this problem and by no means optimal.

Feel free to add or adapt layers and come up with your own experiments. We use the new Keras 1. With processor and model we can initialize a so called KerasBot , which will serve the model for us.

Alright, alright. BetaGo consists of a just few components, all of which you have already seen. First, to load and process data into memory, we use a GoDataProcessor.

BetaGo comes with two such processors out of the box, namely SevenPlaneProcessor and the simpler ThreePlaneProcessor but it's relatively straight forward to add new ones.

The processor loads an index of zip files containing. There's a lot of Go games on KGS, so if you are not careful and try to load too many files this way, your application may crash.

This is where GoFileProcessor comes in, which stores data in a lean, binary format to be picked up later on. The work on processors originated from hughperkins kgsgo-dataset-preprocessor project, which deserves a lot of credit.

Next, to actually predict moves on data processed by any of the above processors, we provide a default implementation of a GoModel , called KerasBot , which trains a deep network of your choice and exposes it to a Flask REST API, whereas IdiotBot simply makes random moves.

KerasBot will try to place the best move, but will take inferior moves if predicted values turn out to be illegal moves. Notably, it is very handy to use keras here, but creating a new GoModel from scratch is not that hard.

In particular, it should be possible to extend the simple approach of KerasBot to something more sophisticated, e. The UI uses a fork of jokkebk awesome jgoboard, and the current Go board front end is just a plain JavaScript client for the above Flask server.

Being both a passionate and mediocre Go player and programmer, this project is a matter of honor to me.