Keras Implementation of Alpha Zero for Chess

#1

Post by atroyn » Sat, Mar 03, 2018, 6:53 PM UTC

Keras Implementation of Alpha Zero for Chess
github.com

Re: Keras Implementation of Alpha Zero for Chess

#2

Post by bradknowles » Sat, Mar 03, 2018, 9:46 PM UTC

In the repo, they suggest that you instead go support another project. Namely, https://github.com/glinscott/leela-chess

Re: Keras Implementation of Alpha Zero for Chess

#3

Post by waterhouse » Sun, Mar 04, 2018, 8:13 AM UTC

bradknowles wrote: post #2

In the repo, they suggest that you instead go support another project. Namely, https://github.com/glinscott/leela-chess

Oh man, it's like BOINC. (In fact, perhaps one could incorporate it into BOINC.) I wonder if, after a while, the most powerful ML platforms in the world will be volunteer-run distributed systems. That would be an interesting situation.

Re: Keras Implementation of Alpha Zero for Chess

#4

Post by thope » Sun, Mar 04, 2018, 9:43 AM UTC

If anybody is interested, you will find ~300 millions PGN to train your NN models on chess games here : https://database.lichess.org/

Re: Keras Implementation of Alpha Zero for Chess

#5

Post by atroyn » Sun, Mar 04, 2018, 10:12 AM UTC

thope wrote: post #4

If anybody is interested, you will find ~300 millions PGN to train your NN models on chess games here : https://database.lichess.org/

One interesting thing about the 'zero' approach is that it learns through self-play. So you can use a database to bootstrap the learning, but in principle you'd only need the game rules and enough training time.