Chess reinforcement learning by AlphaGo Zero methods

Discussion of chess software programming and technical issues.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Post Reply
bhamadicharef
Posts: 31
Joined: Fri Nov 25, 2016 9:14 am
Location: Singapore
Contact:

Chess reinforcement learning by AlphaGo Zero methods

Post by bhamadicharef » Sun Mar 04, 2018 10:38 am

Hi all

Chess reinforcement learning by AlphaGo Zero methods
https://github.com/Zeta36/chess-alpha-zero

Enjoy !
Brahim HAMADICHAREF
Singapore

CheckersGuy
Posts: 273
Joined: Wed Aug 24, 2016 7:49 pm

Re: Chess reinforcement learning by AlphaGo Zero methods

Post by CheckersGuy » Sun Mar 04, 2018 11:03 am

This github repo was already posted ( if not in multiple times). Still, I would really like chess-alpha-zero to be as popular as the go variant (LeelaZero) :cry:

brianr
Posts: 443
Joined: Thu Mar 09, 2006 2:01 pm

Re: Chess reinforcement learning by AlphaGo Zero methods

Post by brianr » Sun Mar 04, 2018 1:18 pm

That repo is largely dormant, although I am still fiddling with it (not as the owner).

Most are looking here, which is becoming ready for widespread use:

https://github.com/glinscott/leela-chess

Henk
Posts: 6480
Joined: Mon May 27, 2013 8:31 am

Re: Chess reinforcement learning by AlphaGo Zero methods

Post by Henk » Fri Mar 09, 2018 8:57 am

Maybe better forget it. It's too slow for normal hardware.

brianr
Posts: 443
Joined: Thu Mar 09, 2006 2:01 pm

Re: Chess reinforcement learning by AlphaGo Zero methods

Post by brianr » Fri Mar 09, 2018 12:16 pm

Depends on your point of view. I've been doing computer chess since 1971, but not very well. Tinker is consistently in the top of the bottom third of engines. Hah. For me, it is about the journey.

Trying to understand the AZ approach and watching my gpu chew on the NN until things train a bit more is satisfying enough. The Leela Chess team is far ahead, like Fishtest with Stockfish, but holds real promise with a more crafted NN and tuning approach fueled by crowd-sourced horsepower.

Henk
Posts: 6480
Joined: Mon May 27, 2013 8:31 am

Re: Chess reinforcement learning by AlphaGo Zero methods

Post by Henk » Fri Mar 09, 2018 12:31 pm

4673 output nodes and (8? *) 19 * 64 input nodes makes each network slow.

brianr
Posts: 443
Joined: Thu Mar 09, 2006 2:01 pm

Re: Chess reinforcement learning by AlphaGo Zero methods

Post by brianr » Fri Mar 09, 2018 2:11 pm

Yup, which is why I started looking at tic-tac-toe.
That NN is very fast with the AZ approach.
Then, I looked at Othello.
Thanks to https://github.com/suragnair/alpha-zero-general

Its NN is considerably slower, but the game is far more complex.
Of course, it is another major complexity jump to chess.

Mangling: Don't bring a knife NN brain to a gunfight (chess or go) :

Post Reply