Hello Chess Lovers and Technology Enthusiasts.
Today I want to show you a very cool Open Source Project that I am following for the last few weeks.
A few months back, DeepMind (an AI company owned by Google) has published a Paper about AlphaZero, a Chess AI that was able to beat the best Chess Engine (Stockfish) with an incredible result of 28 wins and 0 losses with 72 draws.
Obviously there was a huge interest in the chess community about AlphaZero.
Unfortunately they only published 10 games and the Source Code hasn't been published.
But now, some clever developers have coded Leela Chess Zero (LCZero), a Chess AI that works like AlphaZero, but it's Open Source and everyone can contribute to this Project.
How does it work?
Unlike conventional Chess Engines, LCZero is learning Chess on her own by constantly playing against herself.
Every move she makes, every checkmating pattern and every opening that she plays has been discovered by her own.
Nobody has teached her tactics and strategies. She is figuring out everything by herself.
LCZero is using a neural network that has recently been increased from 64 to 128 filters.
Every X amount of Selfplay Games, a new neural network is created, which then plays against the older version.
Right now, this happens multiple times per day and LCZero is constantly getting better and better every day.
The fast growth is only possible, because hundreds of people are commiting their CPU and GPU power to support the project. Because of that, LCZero can play almost 500k Selfplay games every day.
Meanwhile, some users letting LCZero compete against some conventional Chess Engines to document the progress.
You can check this or that Google Spreadsheet to see the current strength.
LCZero has already reached Grandmaster Level.
The estimated rating for Neural Network ID 160 is already over 2600 CCRL at 1+1 time control.
Today, Grandmaster Andrew Tang played against LCZero ID 125 on lichess.org
The Match was streamed on twitch.tv and was watched by over 4000 live viewers on 2 channels.
Andrew Tang has a FIDE rating of 2517 and is currently rank 1 in bullet and ultrabullet time control on lichess. Rank 61 in blitz and rank 2 in rapid.
He played 2 rapid games with a time control of 15+2, which he both lost, 4 blitz games at 5+2, which he also lost.
He got 3 draws and 12 losses at bullet tc and he managed to get 1 win in 23 games of ultrabullet with 3 draws and 19 losses.
This was a very good result for LCZero. Much better than expected.
And it was only ID 125. The newer IDs are even stronger.
It is expected that LCZero will reach 3000 Elo within the next few weeks and it will probably be able to beat Stockfish within the next 3 months.
If you are interrested in the Project, you can check the official Website and the Source Code on GitHub.
They also have a very active Discord Server.
If you want to play against Leela, you can either use this guide and use a Chess GUI like Arena on your local Computer, or you can try to beat a weaker version on http://play.lczero.org/