Alpha Zero vs Stockfish (8) vs Leela Zero (Lc0)

[supersharp77]

Well Fellow Citizens it seems the "Deep Mind Group" may have given up on their 'Alpha Zero Chess Project' with the release of their coding for the aforementioned chess engine (Alpha Zero Basic Code) With the release of that code..it is the wish of many Chess enthusiasts (myself included) that the code be used to create a chess engine as close to the original Alpha Zero as possible so that we can check those original Wild Claims made back in 2017-2018 that Alpha Zero in the space of a 24 hour learning period was able to train a Chess Engine to Defeat the Legendary Stockfish 8 in a 100 Game Match..(A claim many in the Chess World.. Including myself thought to be a quite outrageous boast at the time). Not much has been heard from the mysterious 'Deep Mind Group' except for the occasional release of a theoretical paper..and when challenged to 'recreate the results of the Stockfish 8 match conditions they never responded and to the present day No one including the Leela Zero /LC0 team has been able to recreate those results. There were questions about CPU vs TPU's and opening books and time controls But I am quite sure that from time to time they have checked to see what Alpha Zero can do against LC0 or Stockfish Development or even Stockfish NNUE...and I am of the opinion that that is why they may have released their code at this time. Some of this reminds me of the IBM moves made after the historic Karparov vs Deep Blue matches with the Deep Blue project being discontinued and the software secrets hidden away for many years (recently viewed the youtube movie on the Deep Blue Project)...All that being said their contribution has been tremendous and we luckily got LC0 out of the deal with the help of a huge number of enthusiasts and programmers... Still the questions linger and the mystery remains.."Does the Alpha Zero engine still exist? Have they been testing against the current crop of chess engines? If so what were their results? Perhaps they have something even better? I am sure someone knows something but at present we have heard little to nothing...Let us know your opinion(s) on these matters...Happy New Year.. AR

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.

If Leela had access to the same number of games for training as Alpha Zero did then would it be as strong as the Alpha Zero on comparable hardware (something that matches 4 tpu).?

If this is unknown then
Maybe someone can run a 100 game match against stockfish 8 (with the same configuration as alpha zero played against). Leela uses hardware comparable to 4 tpu.

I would love to see alpha zero verses leela zero with unlimited hardware (e.g alpha zero with 100 TPU cards funded by google verses leela zero with 100 top GPU cards funded by Nvidia).
But if wishes were horses then beggars would ride.

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.

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.

DeepMind published a new paper.
It might be interesting for the experts. It´s a new approach to learning, which they evaluated on several board games.
https://arxiv.org/pdf/1911.08265.pdf

Hi,

What do you think about the claim that "Fat fritz" is stronger than Stockfish. Here is ChessBase article on this issue.

[Dann Corbit]

Without a giant rack of Google TPUs you will never be able to test their claims.
However, it is likely that the LC0 team can glean some valuable things by careful scrutiny of whaty the A0 team did.

[M ANSARI]

I think the claims of A0 are quite accurate. I also doubted their initial claims, but looking at how Lc0 and NNUE has changed the landscape of chess engines ... it is hard to not accept that A0 and AI has given chess engines a monumental leap in chess engine strength ... and the mighty SF8 (at the time) would also be cannon fodder against today's NNUE engines if given equivalent hardware to what A0 used. Remember that the hardware used was and still is difficult to reproduce outside of the Google sphere. I do think that Lc0 is a much more advanced version of A0 today if using equal hardware. If you look at some of the games of A0, it is clear that it suffered from the same weaknesses that Lc0 suffered with initially (terrible tactical mistakes in some positions). Lc0 seems to have progressed beyond that and thus I would consider it a much more polished version of A0 if using the same hardware.