Leela Zero

Project Description

Leela Zero is a distributed, community effort to replicate the Alpha(Go) Zero results. Volunteers run a software client to produce self-play games between strong computer-go players, which are used to train a new, stronger generation of such players.
As of this writing (06-04-2018), the program is of pro strength and still improving.

Aim of the Project

Target groups: all go players that have access to a computer.

Goals:

• to make the resulting >=pro level software, data, networks and games freely available to the public.
• to provide a good codebase for further research into machine learning and go.
• to encourage the development of suitable graphical interfaces for the study of go games with >=pro-level analysis.

Location

Worldwide effort.

Timeline

2017-10-24: Public release of the Leela Zero source code.

2017-11-10: First network launched on the distributed effort.

End of 2017: The program reaches 5 kyu level.

2018-03: The program reaches pro level.

Results

• GTP compatible Go program.
• Series of neural networks that play Go at a wide variety of levels.
• Collection of all played games.
• Collection of all training data from those games.

Exposure

The project has been widely commented on most online Go forums, including internationally, and some articles in Chinese newspapers. Pro players have played games and commented them on YouTube.

Means of the project

C++, OpenCL, TensorFlow, Boost, CMake, Github, Node.js
Our website is http://zero.sjeng.org/

Human Resources

Gian-Carlo Pascutto: project lead, client implementation, training software implementation.

Jonathan Roy: server implementation.

Worldwide Go community: software improvements, running the clients.

Financial Resources

Own budget for hosting expenses. Storage for training data sponsored by online-go.com (OGS).

Perspective

The project will likely run at least through 2018, but at some point progress will inevitably slow down and people may drop off of the distributed effort.
The knowledge and data gathered is likely to lead to spin-off approaches trying new techniques. A fork for chess was already produced.

Wishes

People to keep running the client software.
People that not just have ideas but want to implement and test them in a scientific manner.
People to create analysis and study tools that work with the Go engine.
Some hosting support for the ever-growing amount of data.

If you wish to participate in the project, please visit https://github.com/gcp/leela-zero#i-want-to-help