AlphaZero stronger than Stockfish


Back to forum

Thibault de Vassal    (2017-12-07)
AlphaZero stronger than Stockfish

It looks like there's no more month without news from Google Deepmind... This time again, this is quite stunning!

AlphaZero would have been able to beat (crush) the most recent version of Stockfish, that is also the world champion program and of course the free engine well known by correspondence chess players.

But most important is that actually AlphaZero would have outperformed Stockfish after only 4 hours of training (if I understood well), while it took 8 hours to outperform AlphaGo Lee and only 120 minutes to outperform Elmo at Shogi! However it seems much much harder for the neural program to improve at chess after this stunningly fast auto-learning.

100 games played (25 wins & 25 draws with white! 3 wins with black... no loss, either with white or black, which is an incredible performance)

All details available (must read) here:

A few games played by AlphaZero against Stockfish are included in the arxiv article.

Ilmars Cirulis    (2017-12-07 18:08:10)
AlphaZero stronger than Stockfish

Let's ask Stockfish to make their own "Zero" version. :) Get some 1000 donors of computing power, that should be enough to reproduce this result in a year or two, I believe.

Arturas Drozdovas    (2017-12-07 20:42:55)
AlphaZero stronger than Stockfish

Stockfish was like a beginer compared to alphazero :) brilliant games

Thibault de Vassal    (2017-12-08 00:59:20)
AlphaZero stronger than Stockfish

What do you mean by "Zero" version Ilmars? Would this be about the computer hardware according to you?

Thibault de Vassal    (2017-12-08 03:47:56)
AlphaZero stronger than Stockfish

According to Chessbase, there was 24 hours of self-learning before the match.

Also, "AlphaZero was calculating roughly 80 thousand positions per second, while Stockfish, running on a PC with [32-cores, was calculating] about 70 million positions per second.", about 900 times slower that is quite significant.

Ilmars Cirulis    (2017-12-08 10:00:57)
AlphaZero stronger than Stockfish

No, just do some collective, distributed training of the neural network (multiple machines doing self-play, one doing update sof the network), with public weights of the resulting neural network and open-source "Stockfish-Zero" for running it.

The Leela author is doing something similar with Leela-Zero project, for example.

Ilmars Cirulis    (2017-12-08 11:01:15)
AlphaZero stronger than Stockfish

The result could be run on any machine - either with possibly beefy video cards, or using only the slower CPU.

On correspondence time controls that would great analysis tool in any case.

Garvin Gray    (2017-12-08 15:07:01)
AlphaZero stronger than Stockfish

Thanks for the extra information Thibault about processing speed available to both engines.

Also, I wonder how AlphaZero would go against the latest Houdini and Komodo with equal equipment?

The question is not one of curiousity, but is more that I have a suspicion that also the AlphaZero programmers had the opportunity to gain quite an advantage because Stockfish is open course and so they could really work out how the engine 'works'.

This process would be much harder against Houdini or Komodo. So a long time control match against either of those engines with equal processing power would be a fairer test.

Ilmars Cirulis    (2017-12-08 18:10:02)
AlphaZero stronger than Stockfish

That too close to conspiracy theory. If one can't trust Deepmind, well, there are even worse ways how they could cheat - including completely falsified games by some bored grandmasters with engine and lots of free time.

Timofey Denisov    (2017-12-08 20:58:14)
AlphaZero stronger than Stockfish

Then WHY AlphaZero isn't in CCRL?

I think it was such method of cheating: different hardware. So AlphaZero has much more CPU time. It's explains all...

Timofey Denisov    (2017-12-08 21:03:17)
AlphaZero stronger than Stockfish

And UCI protocol don't require "per se" alpha-beta search algorythm. Engine can send to analysis NOTHING, and respond only MOVE with any score.

Ilmars Cirulis    (2017-12-08 21:12:35)
AlphaZero stronger than Stockfish

Btw, AlphaZero and the match is proof of concept, for research purposes. And that was done excellently.

If one remembers that Deepmind priorities isn't being computer chess or Go champion, but advancing Artificial Intelligence research (and earning money for Google - using those advances and gained experience) then any accusations of cheating or not participating in CCRL seems meaningless.

Arturas Drozdovas    (2017-12-08 21:16:36)
AlphaZero stronger than Stockfish

Just look how alphazero plays, strategic moves that lead to a win. Its impossible for houdini, komodo or stockfish to find these moves with any of the hardware.

Arturas Drozdovas    (2017-12-08 21:19:34)
AlphaZero stronger than Stockfish

Probably not a single correspondence player would find a win as black in a game posted above :)

Ilmars Cirulis    (2017-12-08 21:23:22)
AlphaZero stronger than Stockfish

There will be some more matches, I hope. I expect that at least some criticism about settings and machine of chess engine will be heard.

Also more serious paper about AlphaZero (chess/shogi/go) will be publised.

And then Deepmind are going to leave chess in the past, in the same way as it did with go. At the best, it will be used for testing some of next research ideas, and we will get few more games to look at.

And then it will make a superhuman Starcraft player, then maybe some AI that can do math research like the humans do (I would like to live so far :D), then maybe computer will learn languages properly... :)

Chess is just random checkpoint for Deepmind. We will have to make our own AlphaZero to play with, anyway. :)

Ilmars Cirulis    (2017-12-08 21:25:17)
AlphaZero stronger than Stockfish

* published

(Demis Hassabis said that on Twitter - i mean, about more serious publication.)

Ilmars Cirulis    (2017-12-08 21:34:20)
AlphaZero stronger than Stockfish


Ilmars Cirulis    (2017-12-08 22:56:07)
AlphaZero stronger than Stockfish

Here is also a Reddit thread with few comments/answers from Stockfish author Tord Romstad:

One quote from it.
<< But please note that I don't really blame Google. They didn't do anything wrong. They were running a scientific experiment, not a sporting event, and their experiment convincingly demonstrates what they set out to prove. >>

Thibault de Vassal    (2017-12-09 00:08:28)
AlphaZero stronger than Stockfish

Many thanks.

With all these informations, and now that I've seen the games, I must say that I have more and more doubts about the significance of this match... To be continued.

Garvin Gray    (2017-12-09 09:09:47)
AlphaZero stronger than Stockfish

Just to be clear, or a clarification for those who need it. I never stated that anyone from AlphaZero cheated, or did anything of the kind.

All I stated was that because Stockfish is an open source program, they had the opportunity to study how the program works and tune AlphaZero to take advantage of that.

Any opponent should take advantage of any small advantage they can gain.

I also made clear mention of the processing speed differences. This presented AlphaZero a clear advantage. The issue of same hardware being used is not a new topic. It is brought up in almost every tournament when chess engine tournaments are being played. That for a fair and level competition, the engines need to compete on similar strong hardware.

Reading some of the other replies and thinking further about my first post, I wonder how AlphaZero would go if a four engine event was held, with one game per day between Houdini, Komodo, Stockfish and AlphaZero with equal hardware being used?

Little opportunity then to tune AlphaZero to one specific engine and it would be more a test of the overall strength of the engine in long time control play.

Ilmars Cirulis    (2017-12-09 11:24:21)
AlphaZero stronger than Stockfish

Tuning against specific opponent = 100% cheating, in this context. In that case Deepmind simply lied in their own article.

Herbert Kruse    (2017-12-09 17:04:48)
AlphaZero stronger than Stockfish

i am not impressed at all, in corr chess AZ would be no help

Ilmars Cirulis    (2017-12-09 21:49:56)
AlphaZero stronger than Stockfish

If you limit it to only one minute, then I
almost agree.

Ilmars Cirulis    (2017-12-09 21:51:51)
AlphaZero stronger than Stockfish

Although, with one minute for one correspondence game move... no engine is any help.

Thibault de Vassal    (2017-12-09 23:04:05)
AlphaZero stronger than Stockfish

Or... on 2000 cores, any engine can help :)

Ilmars Cirulis    (2017-12-10 00:51:56)
AlphaZero stronger than Stockfish

For 6000$ one can buy necessary hardware for running AlphaZero. :)
(In reddit comments there were recomendation of Titan-V:
Buy two and you supposedly have about the same big machine as AlphaZero used for playing match games.)

If I had the money, I would buy. Probably wouldn't use for chess, anyway, but that's option for true chess lovers (who already have fully trained AlphaZero somehow :D).

Ilmars Cirulis    (2017-12-10 01:03:35)
AlphaZero stronger than Stockfish

No idea how much would 2000 core machine cost. :)

Garvin Gray    (2017-12-14 11:47:49)
AlphaZero stronger than Stockfish

Even more I read about this result, the more I am inclined to put this result in the 'nice story' but the devil is in the detail.

The time control was 1 minute per move, no tablebases and limited hardware for Stockfish.

This is a major limiting factor for any of the major programs and not how they are designed to work.

Also reading the chessbase article, as I suspected, AlphaZero was able to play many, many games against the Stockfish program, but Stockfish had no such opportunity against AlphaZero.

And so with each game, AlphaZero and the programmers had the opportunity to learn about Stockfish's strengths and weaknesses. Stockfish had no such opportunity.

The only way to see if AlphaZero and its MonteCarlo system is any good is for it to enter the next Computer World Championship and play under equal equipment against all the best chess programs.

Same time control, same processing power , opening books tailored for each engine and tablebases available for each engine.

That is the only way to find out how good the next version of AlphaZero really is.

Ilmars Cirulis    (2017-12-14 12:09:18)
AlphaZero stronger than Stockfish

Conspiracy theories and not understanding of Deepmind motivation.

They tested if the concept works. It was success. They are satisfied and start working on other interesting/useful stuff, as they mostly don't care about chess.

The only training of AlphaZero happened when it played against itself. Stockfish was just an opponent to play against - to check how strong has AlphaZero became.

AlphaZero too had no opening book or endgame tablebases, so that's not relevant. Etc. etc., basically too much conspiracy theories and too much caring about which is the most strongest engine (at least in comparison to Deepmind, as they are totally chill about it, imho :D). :)

Ilmars Cirulis    (2017-12-14 12:12:53)
AlphaZero stronger than Stockfish

Sure, give each engine both CPUs and GPUs. :D

Also, look at suffering of Baduk/Go community and try to evaluate - what's the probability that Deepmind will release the trained Neural Network of chess playing AlphaZero. :) It's about zero.

They don't care about TCEC, they are in machine learning business, not chess business.

Ilmars Cirulis    (2017-12-14 12:20:51)
AlphaZero stronger than Stockfish

The best we can hope is:
1) release of other 90 games against Stockfish
2) some self-play games of AlphaZero, when it's fully trained
3) *maybe* some additional games against more powerful Stockfish or some other engine

The best thing in all cases all - possibility to see more AlphaZero games. More very interesting, strategic games.
I can run Stockfish on my computer and only wait 20-50 times longer to get highest quality games of it (sorry, slow computer :D), but I can't do that with AlphaZero, until chess community hasn't made their own version.

Arturas Drozdovas    (2017-12-14 20:42:02)
AlphaZero stronger than Stockfish

Nice games, especialy i liked a game where alphazero let capture of the knight (longterm piece sac) and won. No engine would do such a thing :)

Thibault de Vassal    (2017-12-14 23:06:08)
AlphaZero stronger than Stockfish

I quite agree with Garvin & Ilmars, if it's possible without any contradiction :)