Deepmind releases AlphaZero code

[Madeleine Birchfield]

It seems that Deepmind has released the source code of a reimplementation of AlphaZero on github:

https://github.com/deepmind/open_spiel/ ... alpha_zero

Documentation: https://github.com/deepmind/open_spiel/ ... ha_zero.md

MCTS search: https://github.com/deepmind/open_spiel/ ... ms/mcts.cc

Thanks to Daniel Shawul for finding this information and sharing it with the Leela discord.

[Collingwood]

So has anyone tested this yet?

[Nay Lin Tun]

1.It is not full details of codes

2. It is a very good documentary record on revolution of early neural networks

3. Current Lc0 running on TCEC is 5x times larger NN weights than A0 ( 30x386 vs 20x256) with much more advanced features than A0 ( e.g TB probe, TB scoring of value head, MLH etc) and likely minimal of +180 elo stronger than A0 now.

[Collingwood]

1.It is not full details of codes

What does this mean? Does the github dump not compile and run?

[Nay Lin Tun]

1.It is not full details of codes

What does this mean? Does the github dump not compile and run?

It doesn't mention about all details of A0 training. You can get a running version of weak A0.

[Collingwood]

It doesn't mention about all details of A0 training. You can get a running version of weak A0.

Okay. That's not very useful then, esp. since Lc0 passed full A0 long, long ago.

[supersharp77]

1.It is not full details of codes

What does this mean? Does the github dump not compile and run?

It doesn't mention about all details of A0 training. You can get a running version of weak A0.

Well fellow Citizens...AlphaZero is now a "Chess Legend" and I am sure along with many others would like to see some 'working Windows compiles' produced for some "Testings"....Thx in Advance......AR

[jonkr]

I haven't tested it out yet, but looks like they made a very generalized framework for learning to play games
https://github.com/deepmind/open_spiel
Then they re-implemented some of the algorithms used by alpha zero in that framework.

I am curious about strength but I wouldn't expect too much (at least compared to the real alpha-zero or the strongest currently available programs). I will probably play around with the framework just to have an idea of what it is and how well it works.

I was actually working on something like this. I stalled a bit and went back to using chess to test when I realized how convenient having something like CuteChess-cli is, just auto-run a match on multiple cores with any opponent or options. I kept finding more and more features I'd need to add make working with my general framework nice. I had a cobbled together 8x8 Checkers, 8x8 Othello, and support for different board sizes (but not trained well like the 8x8 versions.)

DeepMind's is certainly much more general and much more complete than I had gotten. Not sure if I will feel like working on mine to releasable state, the number of things to do to make it good and general kept adding up. I am curious if anyone has tested Open Spiel performance/strength.