AlphaZero style distributed computing project for Go and Chess

MooMoo2 · 2018-03-24, 03:42

Someone started an AlphaZero style distributed computing project to play Go and Chess, which I find to be quite interesting. Unlike other programs, they learn how to play both games without any info besides the rules of the game. They start off playing random moves, and initially, the quality of their games is below a complete beginner. However, they learn from each game, and the more games are played, the better they get.

Leela Zero, the Go-playing program, started in late November 2017 and has played a bit under 6 million games at the time of this post. It's now the strongest free go-playing program and is approximately 7-dan strength (roughly the equivalent of a chess IM, but slightly lower than a GM). It beat two low-level professionals a few days ago, and its progression is at: http://zero.sjeng.org/

Leela Chess, the chess-playing program, is a more recent addition and started training earlier this month. After playing approximately 750,000 games, its ELO is around 1300, which will beat beginners but lose to anyone who plays chess seriously. Its progression is at: http://lczero.org/ , and you can play it at http://play.lczero.org/

I'm not affiliated with either project, but if you have any spare computing power that isn't used for prime hunting, run those self-improving programs!

The Carnivore · 2018-03-31, 17:59

Quote:

Originally Posted by MooMoo2

After playing approximately 750,000 games, its ELO is around 1300, which will beat beginners but lose to anyone who plays chess seriously. Its progression is at: http://lczero.org/ , and you can play it at http://play.lczero.org/

LCZero played over 1.5 million games now, so I figured I'd give it a shot. I'm white, LCZero is black. This was the first chess game I've played this year, so I'm happy I didn't lose. I think LCZero is around 1600 ELO now.

LaurV · 2018-04-03, 07:22

I also gave it a try some days ago, with few fast games, we had two draws and it beat me two times. I actually liked it, and the idea is very nice.

Edit: forgot to say, in all games I played with white pieces.

henryzz · 2018-04-04, 13:18

I managed to get to 25%(had taken a rook for a bishop) before making a mistake in the endgame.
Has anyone beaten it yet?

unconnected · 2018-04-04, 14:21

This engine is very weak in endgames. I've played several games in fast mode (say 3-5 min per game) and total score was smth like 10-5 in my favor.

ATH · 2019-05-28, 13:22

Now 16-17 months after the project began Lc0 has won the TCEC Superfinal, the TCEC Cup twice, and 1 tournament so far at chess.com/ccc

2018-03-24, 03:42	#1
MooMoo2 Aug 2010 1176₈ Posts	AlphaZero style distributed computing project for Go and Chess Someone started an AlphaZero style distributed computing project to play Go and Chess, which I find to be quite interesting. Unlike other programs, they learn how to play both games without any info besides the rules of the game. They start off playing random moves, and initially, the quality of their games is below a complete beginner. However, they learn from each game, and the more games are played, the better they get. Leela Zero, the Go-playing program, started in late November 2017 and has played a bit under 6 million games at the time of this post. It's now the strongest free go-playing program and is approximately 7-dan strength (roughly the equivalent of a chess IM, but slightly lower than a GM). It beat two low-level professionals a few days ago, and its progression is at: http://zero.sjeng.org/ Leela Chess, the chess-playing program, is a more recent addition and started training earlier this month. After playing approximately 750,000 games, its ELO is around 1300, which will beat beginners but lose to anyone who plays chess seriously. Its progression is at: http://lczero.org/ , and you can play it at http://play.lczero.org/ I'm not affiliated with either project, but if you have any spare computing power that isn't used for prime hunting, run those self-improving programs!

2018-04-03, 07:22	#3
LaurV Romulan Interpreter Jun 2011 Thailand 258B₁₆ Posts	I also gave it a try some days ago, with few fast games, we had two draws and it beat me two times. I actually liked it, and the idea is very nice. Edit: forgot to say, in all games I played with white pieces. Last fiddled with by LaurV on 2018-04-03 at 07:23

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Considering getting back into distributed computing	jasong	jasong	7	2016-03-28 14:36
GNFS distributed computing	andi314	Factoring	50	2005-05-11 00:56
Distributed Computing Survey	garo	Lounge	11	2004-09-01 03:31
A distributed-computing project to optimize GIMPS FFT? Genetic algorithms	GP2	Software	10	2003-12-09 20:41
The difference between P2P and distributed computing and grid computing	GP2	Lounge	2	2003-12-03 14:13

2018-04-04, 13:18	#4
henryzz Just call me Henry "David" Sep 2007 Cambridge (GMT/BST) 16F8₁₆ Posts	I managed to get to 25%(had taken a rook for a bishop) before making a mistake in the endgame. Has anyone beaten it yet?

2018-04-04, 14:21	#5
unconnected May 2009 Russia, Moscow 13×199 Posts	This engine is very weak in endgames. I've played several games in fast mode (say 3-5 min per game) and total score was smth like 10-5 in my favor.

2019-05-28, 13:22	#6
ATH Einyen Dec 2003 Denmark 3⁵·13 Posts	Now 16-17 months after the project began Lc0 has won the TCEC Superfinal, the TCEC Cup twice, and 1 tournament so far at chess.com/ccc