Mate in 3: 7 minutes for SF 17

[Ras]

Risks that are worth it when you have an infinte number of games to recover from the potentially ensuing mishap, might not be worth it in the decisive game against your main competitor in a round robin.

Engines don't have an understanding of anything other than the position they're given. They don't know about a tournament situation because we don't have general AI. Also, the only reason why humans play so few games (in a single tournament) as to be statistically meaningless is because of human limitations that engines don't have.

Plus, one point you didn't address so far: how does that even matter in this position, given that it is won also without seeing the mate? In this instance, it's not even about taking a risk to lose a game vs. the chance to win others, it's taking the "risk" of a longer conversion vs. the chance to win other games. Game length isn't a strength metric in chess anyway.

[chrisw]

the engines are often not picking moves that lead to a win in the shortest possible number of moves.

Neither do humans. If given the choice between a highly tactical, but shorter win, or just trading down into a trivially won endgame, many humans will do the latter because it means less room for error. Since game length isn't a metric in chess, only win/draw/loss is, both ways are equally strong.

Strong is one of the most abused words in the language. And sloppiest.
Engineering and materials science is way more nuanced. Tensile strength, compression strength, brittle, ductile, hard etc etc.
What do you mean by strong? I use it to mean the limiting force/action required to break something, and the break can occur by many modes. You’re just lazily substituting it for high Elo, but as these conjured up puzzles show, there are many ways a chess engine can “break”, some pretty dumb. Also dumb is metrification of everything down to one number.

Btw, the loose way that chess players use “strong” is superior to 1D machine learning players. A player who wins five and loses five would (I think) be considered a stronger player than one who draws ten. A player who wins five faster and with style would be considered stronger than one who wins five slowly.
ML players have one metric only so they stop there.

[towforce]

...metrification of everything down to one number [is dumb].

Actually, in most optimisation models (e.g. linear programming), this is exactly what you have to do. It enables comparison.

A player who wins five and loses five would (I think) be considered a stronger player than one who draws ten.

A case for moving to football points: 3 for a win, 1 for a draw, 0 for a loss. In the quoted example, the first player would get 15 points, and the second player would only get 10 points.

[Ras]

A player who wins five and loses five would (I think) be considered a stronger player than one who draws ten.

Depending on the tournament rules, that may be a tie-breaker of last resort, but even then, it's very minuscule. However, engine pruning is not optimised to have a lower draw rate while keeping the score equal, but at increasing the score. Hence my example of WDL 9-0-1 vs. 6-4-0.

A player who wins five faster and with style would be considered stronger than one who wins five slowly.

Style is entirely subjective, and fast wins don't necessarily indicate better play. Could also be an attacking style vs. positional grind, and ofc also the choice of openings. Fast wins are easier to pull off with open than with closed openings.

However, the point stands that more Elo comes from scoring better in games, and scoring better in games means playing better (not: "prettier") chess.

A case for moving to football points: 3 for a win, 1 for a draw, 0 for a loss.

That would break quite a lot in chess and is just another example of "chess has become boring, so I want to play another game - but for some odd reason, I still want to keep calling it chess".

[towforce]

A case for moving to football points: 3 for a win, 1 for a draw, 0 for a loss.

That would break quite a lot in chess and is just another example of "chess has become boring, so I want to play another game - but for some odd reason, I still want to keep calling it chess".

Eh? Football didn't break when the EFL (English Football League) changed a win from 2 points to 3 points in 1981, so why would chess break?

https://en.wikipedia.org/wiki/Three_poi ... n_football

[Joerg Oster]

Normally a mate in 3 would require 7 ply (seen when the King gets captured). And it does not use excessive reductions. So it should see it quickly, or not at all. In this case I had expected the 'not at all', because the main line contains a zugzwang, and Fairy-Max does not do null-move verification.

Hi,
It’s quite clear that null moves and various combinations of extensions and reductions are fantastic winning features for normal gameplay, but they are absolutely detrimental in some checkmate problems and certain "artificial" positions, created specifically for entertainment and to challenge human minds, but totally irrelevant from a practical standpoint. Most of today’s strongest engines are probably (though I would remove this adverb) not the best tool to solve them. As an example, I took my modification of the old Fruit 2.1, made purely brute force, and it finds this checkmate faster than any other program tested here, in just 15 ms:

Code: Select all

id name Fruit 2.1 bruteforce AB
id author Fabien Letouzey
...
uciok
setoption name Hash value 1024
position fen 8/8/5k2/p6p/P1r2n1K/8/8/7R b - - 0 1
go movetime 1000
info depth 1 seldepth 1 score cp 541 time 0 nodes 2 pv c4a4
info depth 2 seldepth 3 score cp 511 time 0 nodes 54 pv c4a4 h1e1
info depth 3 seldepth 5 score cp 552 time 0 nodes 290 pv c4a4 h1d1 a4c4
info depth 3 seldepth 5 score cp 555 time 15 nodes 823 pv f4e2 h4h3 c4a4
info depth 4 seldepth 6 score cp 522 time 15 nodes 1326 pv f4e2 h4h3 c4a4 h1d1
info depth 4 seldepth 6 score cp 539 time 15 nodes 1615 pv c4a4 h1d1 a4e4 h4g3
info depth 5 seldepth 8 score cp 565 time 15 nodes 4043 pv c4a4 h1d1 a4c4 d1d6 f6e5
info depth 5 seldepth 8 score mate 3 time 15 nodes 7138 pv c4c3 h1h2 f6g6 h2h1 f4g2

I therefore think that countless other positions will gradually be found where the newest engine takes longer to find the solution, or doesn’t find it at all, but this should no longer come as a surprise to anyone.

Alex

It is just that the gap between solving a trivial mate in 3 and solely optimizing for game-play
has become too large, imho! Obviously, searching 1288 nodes is considered too costly nowadays.

Code: Select all

position fen 8/8/5k2/p6p/P1r2n1K/8/8/7R b - - 0 1
go mate 3
info string Starting Alpha-Beta Search ...
info time 1 multipv 1 depth 1 seldepth 1 nodes 1 nps 1000 tbhits 0 score cp 0 pv f4e6
info time 1 multipv 1 depth 1 seldepth 1 nodes 7 nps 7000 tbhits 0 score cp 0 pv f4e6
info time 3 multipv 1 depth 3 seldepth 5 nodes 654 nps 218000 tbhits 0 score cp 0 pv f4e6
info string Success! Mate in 3 found!
info time 4 multipv 1 depth 5 seldepth 5 nodes 1288 nps 322000 tbhits 0 score mate 3 pv c4c3 h1h2 f6g6 h2g2 f4g2
bestmove c4c3 ponder h1h2

[chrisw]

All machine learning engineers and students and no chess has had the effect to optimise Stockfish into a bean counter.

[Ras]

Eh? Football didn't break when the EFL (English Football League) changed a win from 2 points to 3 points in 1981, so why would chess break?

Chess is not football. Chess is a game with complete information to both players, subject only to logic. Tons of chess combinations that involve a draw of any sort would be massively devalued if two draws didn't equal one win.

I don't understand why people who don't like chess anymore can't just quit chess and start a game that they like better.

[Ras]

All machine learning engineers and students and no chess has had the effect to optimise Stockfish into a bean counter.

Pretty chess doesn't win the games. Winning games does (d'uh). And winning games is how we determine the stronger player.

Reminds me of the same kind of thought when computers became stronger players than humans. Big disappointment because while that had been anticipated, the expectation had been that this would only happen when and because computers would understand chess in terms of human intelligence. That's why chess had been the poster child of the early AI research. Only that it turned out that human intelligence and understanding not only isn't required for chess, it's actually weaker than crunching numbers.

[chrisw]

All machine learning engineers and students and no chess has had the effect to optimise Stockfish into a bean counter.

Pretty chess doesn't win the games. Winning games does (d'uh). And winning games is how we determine the stronger player.

Reminds me of the same kind of thought when computers became stronger players than humans. Big disappointment because while that had been anticipated, the expectation had been that this would only happen when and because computers would understand chess in terms of human intelligence. That's why chess had been the poster child of the early AI research. Only that it turned out that human intelligence and understanding not only isn't required for chess, it's actually weaker than crunching numbers.

You conflate the limited and one-dimensional ML engineer/student with creative human thought. AB with material pruning is a horrible algorithm whose drawbacks are easily seen with these null move and fortress examples. If the engineers could figure out a way to "materialise" or generalise-compute the manifestations of these examples then they'ld have a better algorithm, only they can't and instead deny the problem. Spare us please the cyclic arguments about statistical winning/strength.