Perfect play

[Lyudmil Tsvetkov]

Strong engines playing each other draw more often than weak engines playing each other.

Now, if we draw a graph (elo/draw%), find a best fit curve and extrapolate it, we should be able to estimate the strength of engines playing perfect chess (=100% draws).

Does this seem reasonable? Has it been done?

In the 1920s Capablanca claimed chess has been solved, the same claim was repeated by Fischer half a century later, but stronger human players and engines still come regularly along.

I bet perfect play is way above 6000 elo, current top engines' eval could be reasonably expanded at least 5-fold, and each new feature will bring some higher winning percentage. Same goes true for search.

I am not certain that perfect chess means drawing all games, rather than not losing a single game. The stronger the engines, the more games they win with white. How do you reconcile the 2 contradictory paradigms: the stronger the engines, the more games they are drawing, and, the stronger the engines, the more games they are winning with white?

For me, there are still too many unknowns in the equation for us to be able to draw definitive conclusions. But thinking that perfect play is maybe some 500 elo away is very funny to me; 2000-3000 elo is a more reasonable estimate, but I am not certain this is the upper bound too.

[Henk]

One can always switch over to Capablanca chess variant. But if no one does then it's over. Are there still engines created that play checkers.

[Laskos]

Thanks to Norm Pollock utilities I managed to better filter the data.
Now I have the draw rate of all engines when facing opposition within 20 elo.

The results are similar, although the linear extrapolation is a bit nearer, at 4400 elo.

Very interesting, thank you. If you observe, the better fit with parabola is due mostly to the results of 3000+ ELO engines. I suspect most of these engines are mildly related, and their draw rate is inflated. I would go for linear fit, and my last fit with Komodo, summing up diminishing returns, gave IIRC 4450 CCRL ELO points for perfect engine, close to your linear 4400. It's interesting that from different approaches we agree that the perfect engine is between 3800 and 5000 CCRL ELO points.

[Norm Pollock]

Another filter that should be applied is the elimination of intra-family games. That is games between different versions of the same engine, games between derivatives of the same engine, etc.

Such games by similar engines are likely to result in a high rate of draws because those engines often "see" the same things in each position.

[mjlef]

So with a 50-100 elo gain per year, I still have 10-20 more years of work to do!

[Adam Hair]

Yes, I have done some study on CCRL data, by filtering only programs which have an average opposition within 20 elo of their rating.

A better study would be to filter on a game by game basis instead of player by player, but I've had no luck with database programs (CCRL database seems too big for SCID to filter).

Below the results.

SCID chokes on how the CCRL database makes use of the Round tag in the headers.

http://talkchess.com/forum/viewtopic.ph ... 32&t=53803

If you know a little about regular expressions and use a text editor like Notepad ++, then you can edit the Round tags. This will allow you to process the database with SCID.

Alternatively, you can make use of some of Norm's great tools.

[Uri Blass]

Strong engines playing each other draw more often than weak engines playing each other.

Now, if we draw a graph (elo/draw%), find a best fit curve and extrapolate it, we should be able to estimate the strength of engines playing perfect chess (=100% draws).

Does this seem reasonable? Has it been done?

In the 1920s Capablanca claimed chess has been solved, the same claim was repeated by Fischer half a century later, but stronger human players and engines still come regularly along.

I bet perfect play is way above 6000 elo, current top engines' eval could be reasonably expanded at least 5-fold, and each new feature will bring some higher winning percentage. Same goes true for search.

I am not certain that perfect chess means drawing all games, rather than not losing a single game. The stronger the engines, the more games they win with white. How do you reconcile the 2 contradictory paradigms: the stronger the engines, the more games they are drawing, and, the stronger the engines, the more games they are winning with white?

For me, there are still too many unknowns in the equation for us to be able to draw definitive conclusions. But thinking that perfect play is maybe some 500 elo away is very funny to me; 2000-3000 elo is a more reasonable estimate, but I am not certain this is the upper bound too.

1)I think that elo is dependent on the pool of players so no elo for the perfect player without knowing the opponents and the opponents of the opponents.

2)It is not clear what is the definition of perfect play.
If it is not doing moves that changes the theoretical result of the game then a perfect player can get usually draws against stockfish because the player will not play moves that cause stockfish to blunder.

The perfect player may make weak moves from human point of view that do not lose the game only to find later the only way to draw and practically will claim 0.00 score by the 32 piece tablebases all the game.

[jefk]

well i claimed a few times already that chess is a draw,
but got criticized because there's no theoretical proof.

If we look at the red line in that Elo graph, then my
previous estimate of the maximum rating around 4000
looks quite reasonable.

From games by Brainfish against this other book which
is neutralizing it, we get a glimpse of the drawish results;
ofcourse you can try to go for sharper lines, sometimes
even gambits, but it's only a matter of time before
the other bookmaker will equalize; in the meantime
computers get faster, engines slightly better, and
you can argue for a millennium if the play perfectly,
but if all there games are drawn, then its good enough
for me; oh and it would be an emperical 'solution'
of the game of chess ofcourse; with a draw as result.

PS some people like such predictions because it's not
good for chess; well human chess is a whole different
game, and i'm quite sure it will continue, because
change in the rules always go very slowly; especially
in the Fide for human chess; which is how it should be;
except for the Fide itself maybe but that's another story

[corres]

[quote="Uri Blass"]
1)I think that elo is dependent on the pool of players so no elo for the perfect player without knowing the opponents and the opponents of the opponents.

2)It is not clear what is the definition of perfect play.
If it is not doing moves that changes the theoretical result of the game then a perfect player can get usually draws against stockfish because the player will not play moves that cause stockfish to blunder.

The perfect player may make weak moves from human point of view that do not lose the game only to find later the only way to draw and practically will claim 0.00 score by the 32 piece tablebases all the game.[/quote]

I agree with you totally.
I think this problem can be solved perfectly in the far future only.

[Laskos]

Strong engines playing each other draw more often than weak engines playing each other.

Now, if we draw a graph (elo/draw%), find a best fit curve and extrapolate it, we should be able to estimate the strength of engines playing perfect chess (=100% draws).

Does this seem reasonable? Has it been done?

In the 1920s Capablanca claimed chess has been solved, the same claim was repeated by Fischer half a century later, but stronger human players and engines still come regularly along.

I bet perfect play is way above 6000 elo, current top engines' eval could be reasonably expanded at least 5-fold, and each new feature will bring some higher winning percentage. Same goes true for search.

I am not certain that perfect chess means drawing all games, rather than not losing a single game. The stronger the engines, the more games they win with white. How do you reconcile the 2 contradictory paradigms: the stronger the engines, the more games they are drawing, and, the stronger the engines, the more games they are winning with white?

For me, there are still too many unknowns in the equation for us to be able to draw definitive conclusions. But thinking that perfect play is maybe some 500 elo away is very funny to me; 2000-3000 elo is a more reasonable estimate, but I am not certain this is the upper bound too.

1)I think that elo is dependent on the pool of players so no elo for the perfect player without knowing the opponents and the opponents of the opponents.

This is not what Arpad Elo had hoped to achieve with his ELO differences.

2)It is not clear what is the definition of perfect play.
If it is not doing moves that changes the theoretical result of the game then a perfect player can get usually draws against stockfish because the player will not play moves that cause stockfish to blunder.

The perfect player may make weak moves from human point of view that do not lose the game only to find later the only way to draw and practically will claim 0.00 score by the 32 piece tablebases all the game.

These things certainly can be refined in definition. Suppose we have the first ever perfect player: 32 men bases. It plays against the present day paradigm engines of say 3200, 3400, 3600, ... , etc ELO level. 32 men bases are not used smartly to fool around, say they are DTZ50 and just that. Then the ELO of the perfect engine will be probably close to what Maurizio shows.