Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / Highest level of play: AlphaZero or Aleksandr Dronov?
Poll Who would win?
AlphaZero 7 47%
Dronov 8 53%
- - By Evil_Otto (*) Date 2019-05-26 07:27
I felt ambiguous whether to place this question into the Correspondence Chess subforum or the Computer Chess subforum.

In any case, I'm interested in looking at the chess games with the strongest moves. Of course I could always look at endgame tablebases and see perfection, but I want to see the highest quality of play throughout an entire chess game (to see strategies materialize with lots of pieces on the board).

AlphaZero's latest games against Stockfish were recently released: https://www.chess.com/news/view/updated-alphazero-crushes-stockfish-in-new-1-000-game-match

As mentioned in the article, the time controls were "three hours each game plus a 15-second increment per move."

Also, the games from the latest ICCF World Championship 30 were also released: https://www.iccf.com/ListGames.aspx?event=66745&status=finished

My question is as follows: If Aleksandr Dronov (arguably the world's current strongest correspondence chess player) would have played against AlphaZero, would he have likely won? Why (not)?

Assumptions
- assume that Dronov had access to the various chess engines in the versions that were available at the final stages of World Championship 30 (along with the hardware he had access to)
- assume that AlphaZero was trained according to the description that was released (see the above chess.com article)
- assume that, during play, AlphaZero has access to "four TPUs and 44 CPU cores"
- assume that Dronov has access to the time controls that he had during the games of the  ICCF World Championship 30 (10 moves in 40 days) and that AlphaZero has the time controls listed in the article: "three hours each game plus a 15-second increment per move."
Parent - By Dragon Mist (****) Date 2019-05-29 14:21

> - assume that Dronov has access to the time controls that he had during the games of the  ICCF World Championship 30 (10 moves in 40 days) and that AlphaZero has the time controls listed in the article: "three hours each game plus a 15-second increment per move."


Time control was 10 moves in 50 days, not 40 days.
Parent - - By rocket (****) Date 2019-05-31 11:49
Two weaker opponents teaming up does not beat god-like opposition. This is a no brainer.

If Dronov only played to draw as white, he could, though.
Parent - - By Evil_Otto (*) Date 2019-05-31 16:50
What makes AlphaZero=god-like opposition and not Dronov? AlphaZero also has components that aren't impressive by themselves (policy and value); its strength shows when they work together. Likewise, there is synergy when Dronov works with his engines. Arguably the biggest flaw of traditional engines is depth. But Dronov can constrain engines to look at promising lines based on his intuition/strategy - thus extending the depth (not as far as AlphaZero, but at least virtually all moves in the considered branch are evaluated). He also understands the biases in his engines and can choose one (set of) evaluation over another for a given position (depending on the characteristics).

It's not obvious to me why the AlphaZero synergy between policy and value is superior to that of Dronov's strategical insight & evaluation-function-meta-evaluation combined with his "weak engines"
Parent - - By rocket (****) Date 2019-05-31 19:05
The level AlphaZero plays at many times is beyond human comprehension, with or without Stockfish.
Parent - - By Evil_Otto (*) Date 2019-05-31 20:19
It doesn't seem that obvious to me and, according to the current poll results, to other visitors as well.  Statements "many times is beyond human comprehension" and "god-like" aren't convincing to me, however, without further elaboration.

I should mention though that the pushback I gave before was more in the spirit of the devil's advocate. I think AlphaZero would win because high depth of the "juiciest" few moves on each turn seems overall stronger than an initial strategy that's flawless at limited/shallow ply#. A hybrid approach will likely be extremely strong and if I had asked the question for the next  ICCF World Championship 31 and 32 instead of 30, I believe the answer would be in favor of Dronov (access to LC).
Parent - By rocket (****) Date 2019-06-01 11:42
Listen, Stockfish running at those cores is at least 600 elo stronger than any human. Stockfish is clueless against Alphazero the majority of the time with any type of extended TC. How do you figure a ten times more clueless human is going to be helped by Stockfish in such a match-up?

The human advantages these days are miniscule and not enough to affect any result of a game. A human is still superior at spotting blockades. Well, blockades ain't gonna win you any chess games.
Parent - - By rocket (****) Date 2019-06-01 11:43

>It doesn't seem that obvious to me and, according to the current poll results, to other visitors as well<


The vast majority in here are weak chess players around 13-1800. They can't form an educated opinion on a subject matter they don't understand passed a layman.
Parent - - By Labyrinth (*****) Date 2019-06-01 18:22
Your comments are kind of condescending/ignorant.

First there's an enormous time control difference per the 3rd clause, 10 moves in 50 days vs. a few hours with 15 sec increment and there are various techniques one can use in correspondence centaur type games. It's possible to simply look ahead at certain positions that are likely to arise and analyze them deeply to make the best of the engine's resources for example. Also avoiding blockades could in fact win games if it means avoiding a draw.

Second you're taking Elo way too seriously. It is simply a statistical way of predicting the outcome of games, it isn't some kind of fixed resource. You can also have two entities with very different strengths and weaknesses have the same Elo. Plus, we don't actually know the strength difference between humans and engines. It's clear that they are stronger, but how much stronger is not so clear. The number of games is simply abysmal.

Third, you don't know the Elo of the visitors here. Complete assumption on your part.

So it isn't really so cut and dry. I wish we could run the experiment but this is likely impossible.
Parent - By rocket (****) Date 2019-06-01 21:13 Edited 2019-06-01 21:37

>Also avoiding blockades could in fact win games if it means avoiding a draw.<


Not if the rest of ones attributes for winning are massively inferior, which in the case of man+Stockfish, they are.

>Second you're taking Elo way too seriously. It is simply a statistical way of predicting the outcome of games, it isn't some kind of fixed resource<


Elo ranking is an objective reflection of performances in playing chess across time. Large sample of games eliminate anomalies.

>You can also have two entities with very different strengths and weaknesses have the same Elo<


The strong all have one thing in common: superior pruning. They know which lines to devote time to and which to discard. This is the single greatest evolution you can make as a chess player. When you go over games lost and realize that lines you didn't play, which you viewed as dangerous, were in fact winning or equal. Improving your calculating abilities will aid this problem, which is why Nigel Short often speaks of how human chess is most often decided by tactics.

The best differ in how they go about pruning (some calculate their way there like Korchnoi and Topalov), but the principles remain the same.
Parent - - By Uly (Gold) Date 2019-06-07 09:12

> I wish we could run the experiment but this is likely impossible.


I have on good account that Leela Zero is now stronger than Alpha Zero (at least, the Alpha Zero of the last chess game released), in that either it finds the best moves of the Alpha Zero games, or it finds better moves.

So what stops Aleksandr Dronov from running Leela for a few hours per move and predict Alpha's move or something better and prepare something to defeat it?

It is correct that Alpha wouldn't stand a chance under these conditions.

A more interesting question is what would be the result with 10 moves in 50 days for both players. In this case I'm not sure Aleksandr Dronov is the best for the job, what you want is someone specialized in punishing engine moves (Alpha just has a different horizon, but I think defeating it on a corr game would hold similar principles.)
Parent - By Evil_Otto (*) Date 2019-06-09 03:35

> So what stops Aleksandr Dronov from running Leela for a few hours per move and predict Alpha's move or something better and prepare something to defeat it?


Since the case when Dronov has access to Lc0 makes the question trivial, I had posed it (the question) with an intended assumption of not having access to it. I.e. Dronov would only have access to traditional chess engines.

- assume that Dronov had access to the various chess engines in the versions that were available at the final stages of World Championship 30 (along with the hardware he had access to)

However, I just dug around and found out that Dronov is not even a participant and that the World Championship 30 is still ongoing.
- - By Evil_Otto (*) Date 2019-06-03 20:01 Edited 2019-06-03 20:04
I just searched around a little and found this gem from last year, where ChessBase interviews a strong correspondence player (emphasis mine ): 

"DN: What is your Centaur opinion over AlphaZero? Would you be able to beat it in a fair match?

WM: That whole spectacle seemed to be a spun tale with a stacked deck. If AlphaZero were to play in an actual engine tournament like the TCEC, I seriously doubt it would even make it to the semifinals. In terms of beating it though, I couldn't say if that would actually happen. It could play something like the Petroff and easily draw and that would look like some sort of small victory on its part. However, I will say it will NOT beat me in a Centaur CC game. That's just not going to happen.

DN: Did you look at the new AI called Leela Chess?

WM: I've heard it's based on the same concept as AlphaZero, and I understand it effectively builds a database based on games it plays against itself. It's up to 2500 FIDE level now from what I understand. I really don't know much about it, but I can say again that these types of engines won't be able to catch up with the knowledge built up from CC databases, unless, of course, those games are added as some sort of percentage-based play factor.


https://en.wikipedia.org/wiki/TCEC_Season_15 :lol:
Parent - By The Truth (**) Date 2019-06-06 23:51
And he was wrong! It only took a year for lc0 to go from scratch and patzer dumber than a child play to the level of stockfish now which has taken the team nearly a decade to make. Lc0 is most likely the endgame of computer chess. Highly likely that in 1-3 years with the latest gpu of that time that lc0 could match or even surpass the quality of whatever the top corr player analyzing with traditional engines in a day in only a few minutes.
Up Topic The Rybka Lounge / Computer Chess / Highest level of play: AlphaZero or Aleksandr Dronov?

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill