Google Deep Mind Rating

Sort:
Martin_Stahl
CerealKilla13 wrote:

Wow, I've been wanting to see this match since Alpha Go beat Lee Sedol, Stockfish is like a Ferrari but AlphaZero is like an airplane....Doing something fundamentally different. Think of the wright brothers, They couldn't go faster than a modern day Ferrari, but with a long enough time-line Jets are the fastest things we got (outside of space travel). I don't think anyone can imagine what AlphaZero could do if it played chess for weeks or even months....5,000 Elo? 10,000? 15,000? Chess has been solved? Period. 

 

That really doesn't make any sense. Rating is based on the probability of winning over other opponents.  Unless it gets strong enough to win almost all games, then it could be 677 point higher rated than it's opposition.

 

If it is true that chess is most likely a draw, which the results seem to imply is still likely, then the ratings won't ever get that high.

Martin_Stahl
SmyslovFan wrote:

I still believe the highest possible rating is ~3600. Stockfish's rating is lowered by the fact it wasn't allowed to use its opening database.

 

But AZ didn't really either (it used its learned book). My guess is the results probably wouldn't be a lot different, even with the opening book set for the opponent (Stockfish in this instance). Better hardware on the Stockfish side would likely help some more (to offset the benefit of the Tensor Processing Units of AZ/DeepMind).

 

I agree on the top rating, which would be the max rating it could get by beating the highest rated human players 99% of the time.

SmyslovFan
Martin_Stahl wrote:
SmyslovFan wrote:

I still believe the highest possible rating is ~3600. Stockfish's rating is lowered by the fact it wasn't allowed to use its opening database.

 

But AZ didn't really either (it used its learned book). My guess is the results probably wouldn't be a lot different, even with the opening book set for the opponent (Stockfish in this instance). Better hardware on the Stockfish side would likely help some more (to offset the benefit of the Tensor Processing Units of AZ/DeepMind).

 

I agree on the top rating, which would be the max rating it could get by beating the highest rated human players 99% of the time.

 

When you have two completely different methods of analysis, the two sides should be able to use the optimal tools. AZ effectively created its own openings, while Stockfish was left defenseless in the opening.

Clavius

I seem to recall that the value of a good opening book for chess software is about 50 Elo.  That would give the Stockfish used against AlphaZero a rating of about 3170 based on what TCEC uses for a fully capable Stockfish.  AlphaZero's performance was about 110 Elo better than that or 3280.  However, Stockfish was also deprived of tablebases for endgames which are estimated to add 20-50 Elo so that would put A0 at about 3250.  But A0 was only given 4 hours to train.  The best Go program developed using the same self-learning process got 40 days.  

SmyslovFan

About a decade ago, the best cc players and OTB GMs could still compete against engines that didn't use opening databases. (See the Kramnik-Fritz match, 2006.) Stockfish really did need the database.

Pikelemi
I would like to see AlphaZero against some top GM and also see it play against it self.
MergedZamasu

I would love if AlphaZero played in a chess.com event or TCEC with better parameters and rules.

xropi
[COMMENT DELETED]
vickalan

Chess elo scores can't be calculated until after fair games are played. Currently Stockfish is the best chess engine because it showed up to play in this (recent tournament).

AlphaZero did not.😧

 

lfPatriotGames
CerealKilla13 wrote:

Wow, I've been wanting to see this match since Alpha Go beat Lee Sedol, Stockfish is like a Ferrari but AlphaZero is like an airplane....Doing something fundamentally different. Think of the wright brothers, They couldn't go faster than a modern day Ferrari, but with a long enough time-line Jets are the fastest things we got (outside of space travel). I don't think anyone can imagine what AlphaZero could do if it played chess for weeks or even months....5,000 Elo? 10,000? 15,000? Chess has been solved? Period. 

I dont really understand your analogy, but if this new alphazero is really like an airplane compared to a car isn't there already a way to do much better? a rocket is faster than a jet isn't it?

Elroch
Clavius wrote:

I seem to recall that the value of a good opening book for chess software is about 50 Elo.  That would give the Stockfish used against AlphaZero a rating of about 3170 based on what TCEC uses for a fully capable Stockfish.  AlphaZero's performance was about 110 Elo better than that or 3280.  However, Stockfish was also deprived of tablebases for endgames which are estimated to add 20-50 Elo so that would put A0 at about 3250.  But A0 was only given 4 hours to train.  The best Go program developed using the same self-learning process got 40 days.  

The value of an opening book is simply that it saves computing time. If you base your opening on the results of a hundred previous games, you are effectively making use of the computing (by humans perhaps) that was done in those games.

With AlphaZero, previous opening theory was irrelevant, because the way of playing the openings it developed by self-play appears to be slightly better (at least for itself: when forced to play other openings, it was no longer unbeatable and scored slightly less, but still did well).

Note that AlphaZero does not use an opening book, even one generated by itself. Rather its neural networks tend to like certain moves more than others, and these moves are the ones it plays in the opening like at other stages of the game.

SmyslovFan

Elroch, on the matter of an opening book, you do undersell the importance of the opening. 

 

The opening isn't just a matter of tactics, it's about setting up the pieces for the coming battle. Hundreds of years of human knowledge is worth something in this matter. AlphaZero came up with its own theory of the game and came up with openings that even strong GMs are impressed with and are studying.

No engine before AlphaZero has done as much to advance opening theory. Sure, computers have found holes in openings, and have even refuted entire variations. But AlphaZero has created some important theoretical breakthroughs in a single match. It has shown elite players some fascinating ideas in the Queen's Indian. It came up with a d5 line independent of human thought that humans have only started to explore seriously in the last decade or so. 

 

The opening database doesn't just provide a short-cut for computers, it provides them the blueprint for how to develop. AlphaZero has bypassed that entirely!

Elroch
SmyslovFan wrote:

Elroch, on the matter of an opening book, you do undersell the importance of the opening. 

 

The opening isn't just a matter of tactics, it's about setting up the pieces for the coming battle. Hundreds of years of human knowledge is worth something in this matter. AlphaZero came up with its own theory of the game and came up with openings that even strong GMs are impressed with and are studying.

What you may miss is that to AlphaZero, these are just moves it plays in a position that it doesn't have to remember having seen before. It uses the same approach to selecting a move on move 1 as it does on move 101. The games it played against Stockfish indicate that it incorporates some stochasticity in selection (as it did not stick rigidly to the same moves in the same positions).

No engine before AlphaZero has done as much to advance opening theory. Sure, computers have found holes in openings, and have even refuted entire variations. But AlphaZero has created some important theoretical breakthroughs in a single match. It has shown elite players some fascinating ideas in the Queen's Indian. It came up with a d5 line independent of human thought that humans have only started to explore seriously in the last decade or so. 

 

The opening database doesn't just provide a short-cut for computers, it provides them the blueprint for how to develop. AlphaZero has bypassed that entirely!

 

CerealKilla13

IfPatriotGames - well yes the AI is only going to get better, but it will take time. As for my comment about the Elo ratings, I said it tongue-in-cheek. I guess some people thought I was serious. Although GM Yasser Seirawan did make a  comment on one of the tournaments on the St. Louis Chess Club that you could be 5,000 Elo and there would be no way to calculate it if you won all your games ( you need to lose or draw at least one game before your Elo can be calculated; because mathematically, it could be infinite.)

SmyslovFan

Elroch wrote, "What you may miss is that to AlphaZero, these are just moves it plays in a position that it doesn't have to remember having seen before. it uses the same approach to selecting a move on move 1 as it does on move 101."

Elroch, you are making assumptions about how AlphaZero approaches openings compared to endgames that simply is not supported by any evidence, unless of course you have analyzed the paper that has yet to be published in extreme detail. 

 

From what I know about chess, the rules of development change in the endgame and the selection of moves to consider reflect that. I am pretty certain that AlphaZero has come to the same conclusions.

I believe that because of what I know about chess. I accept that AlphaZero may have turned those basic rules on its head, but I think it much more likely, based on its play, that it has discovered for itself truths that humans took centuries to develop. 

Elroch
SmyslovFan wrote:

Elroch wrote, "What you may miss is that to AlphaZero, these are just moves it plays in a position that it doesn't have to remember having seen before. it uses the same approach to selecting a move on move 1 as it does on move 101."

Elroch, you are making assumptions about how AlphaZero approaches openings compared to endgames that simply is not supported by any evidence, unless of course you have analyzed the paper that has yet to be published in extreme detail. 

I have read it carefully, having a considerable amount of experience in the subject myself, applied in a different area. I am certain that what I say is true. There is no role for an opening book in this system: there are two neural networks and these allow it to play moves in any position using the same procedure.

From what I know about chess, the rules of development change in the endgame and the selection of moves to consider reflect that. I am pretty certain that AlphaZero has come to the same conclusions.

For sure: the concepts AlphaZero encapsulates in its learned neural networks can make different parts of the networks important in different situations. This applies to almost any possible aspect of the game.

I believe that because of what I know about chess. I accept that AlphaZero may have turned those basic rules on its head, but I think it much more likely, based on its play, that it has discovered for itself truths that humans took centuries to develop.

The networks are used by the programmed algorithms in a totally uniform manner - evaluating, searching and selecting on every move. But the parts of the networks that are important on different moves may vary: this is exactly how most neural networks work. I think we probably agree but our words were not so clear to each other.

 

SmyslovFan

We are talking across purposes.

I accept that AlphaZero is the same machine and uses the same processes for determining the best move in the opening as opposed to the endgame. 

But, chess requires different priorities in the opening than the middle game or the endgame. Engines have traditionally had a very difficult time navigating the opening without a database, but have become quite literally perfect when the situation is reduced to 7 pieces or less. 

AlphaZero has proven that it can solve the problem of the opening in ways that previous engines could not. It showed that it can adapt far better than any previous engine. 

That does not disprove what I know about chess. It does prove that AlphaZero is a completely different sort of beast than Stockfish is. 

It also highlights Nakamura's complaint that Stockfish was crippled by the fact that it could not use databases. AlphaZero was fighting on its own turf. It won brilliantly, and it would have won even if Stockfish did have access to databases. But, Stockfish would not have lost by that lopsided margin. 

Elroch

Checking the comments of Nakamura, I assume he knows what he is talking about when he says Stockfish was unable to use an opening book -  this information is not confirmed by the DeepMind paper, and I can't locate a source for it.

This would be a fair point, but play in the opening book is really like correspondence play, using a vast amount of past computation to select moves. Personally, it if is true it was switched off, it would have got Stockfish some more draws and very likely no wins (the reason is that opening books are based on games against players significantly weaker than AlphaZero).

sammy_boi
Elroch wrote:

very likely no wins (the reason is that opening books are based on games against players significantly weaker than AlphaZero).

Note that in the 1200 game match where they played a variety of openings SF did score some wins, and the overall result was 61% (less than the 100 game match's 64%). An Elo difference of 100 or less isn't much.

In any case a good openings book would allow SF to play the opening better than it could on its own. We can't say for sure that it wouldn't have been able to score a win. A better mid game + more calculation + EGTB access... who knows.

DiogenesDue
SmyslovFan wrote:

We are talking across purposes.

I accept that AlphaZero is the same machine and uses the same processes for determining the best move in the opening as opposed to the endgame. 

But, chess requires different priorities in the opening than the middle game or the endgame. Engines have traditionally had a very difficult time navigating the opening without a database, but have become quite literally perfect when the situation is reduced to 7 pieces or less. 

AlphaZero has proven that it can solve the problem of the opening in ways that previous engines could not. It showed that it can adapt far better than any previous engine. 

That does not disprove what I know about chess. It does prove that AlphaZero is a completely different sort of beast than Stockfish is. 

It also highlights Nakamura's complaint that Stockfish was crippled by the fact that it could not use databases. AlphaZero was fighting on its own turf. It won brilliantly, and it would have won even if Stockfish did have access to databases. But, Stockfish would not have lost by that lopsided margin. 

I never worked in this area of development, but essentially, what the machine learning does/did here is to do almost exactly what human being did with engines:  take the accumulated knowledge of the game via massive numbers of games played, distill that down into a set of valuations, then apply those valuations.  The difference here is:

- Human beings have created valuations for human created engines based on centuries of accumulated chess knowledge.  They started with the rather basic queen = 9, pawn = 1 valuations and got a lot more complicated from there. 

- AlphaZero did this entire process in 4 hours by playing tens of millions of games against itself.

- Human beings had to add opening books, because they could not correctly evaluate those "valuations", ultimately, and also because engines would fudge the opening because of their event horizons.

- AlphaZero will create valuations completely by itself, without the biases and fallacies of human thinking, like the idea of attack vs. retreat, holding material vs. positional advantage, etc.  So, it stands to reason that the moves it is selecting are a surprise to GMs, because current engines use human-derived valuations for everything they do.

The results will be amazing...it's just too bad that they decided to "leak" their internal testing to the world for some publicity, because it was premature.  If they had waited, they never would have had to hamstring Stockfish the way they did in this match to make the results look so crushing.

If I were the CM on the dev team, I would be praying my bosses and the press never found out what I actually did here to secure the match results shown.  There was really no reason to fake something that will probably be true in short order anyway...

Ironically, once machine learning surpasses the chess world's, GMs will quickly begin to realize that they cannot really learn much from engines anymore, because while those engines will play and understand chess far better than a human ever has, ultimately they will be completely incapable of explaining why in a way we can learn from wink.png...and we will eventually fall too far behind to dope it out ourselves.  That is, unless they develop true AI and the computer does become capable of "teaching" human beings in an interactive way.