Cursed win at TCEC

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Harvey Williamson, bob

Lyudmil Tsvetkov
Posts: 6031
Joined: Tue Jun 12, 2012 10:41 am

Re: fortress_draw_rule

Post by Lyudmil Tsvetkov » Mon Nov 28, 2016 7:43 am

Uri Blass wrote:
Lyudmil Tsvetkov wrote:
Evert wrote:
Evert wrote: I'll ask you again: how is it more unjust to be deprived of a win than to be deprived of a draw?
Let me rephrase this differently.

Once upon a time, it was realised that there are 5-men positions that could be won, except for the 50-move rule kicking in (I remember it came up during a Timman-Yusupov game). In current parlance, they are cursed wins. This was deemed unfair, and the rule was changed so it was 50 moves, except in certain endings where it was 75.
Just a few years later, more cursed wins, with longer move chains, were discovered. It was realised that the rule as it was just made arbitrary exceptions, and was not fair. So a choice had to be made:
1. Keep the unfair rule as it was.
2. Revise the rule every time a new cursed win was found.
3. Abolish the rule entirely.
4. Revert to the plain 50-move rule.

Obviously, 4 was the choice made. The other options all have more undesirable features (arbitrary, not scalable, open for abuse).

Now, you claim that the existence of cursed wins makes the game less interesting, at least for computer chess, and the cursed wins should be treated as normal wins. In effect, this is option (2) above, but we might as well consider (3) because as technology improves, that's what it'd end up coming down to anyway. Does this make for a better game? I'll argue why I think it does not. You're free, of course, to push for your "Tsvetkov's Chess" that has different rules.

First of all, many of these extremely long wins feature move sequences that are utterly incomprehensible. Is it interesting as a spectator game? Not really.
More importantly, the game-theoretic value of the opening position of FIDE Chess is (almost certainly) a draw. It's possible that this is entirely due to the 50-move rule. We don't know the outcome of Chess without this rule. If it's still a draw, then it doesn't really matter much in the end. On the other hand, perhaps the opening position is really a cursed win. If that is the case, abandoning the 50-move rule makes the game a win for white.

So, what is more interesting: a game where Black's job is to defend the draw and keep White from winning by playing accurately, or a game where Black's job is to hope White makes a mistake, because there's nothing he/she can actually do themselves to affect the outcome of the game?
no need to revise the rule so often: a new position involving longer mate is found once in a while, half a decade or so. not very big burden on programmers, is it?

besides, we do not need to necessarily include all such longer mates, but simply include as many as possible with a reasonable approach, so that the damage is minimised. If you include 500 positions and leave out some 40 or 50, the damage will be minimised, while a shorter draw rule could be enforced. but I guess variable length rule is about the best approach.

the real problem is that outdated rule greatly impacts on engine strength. If SF hits on average 30 000 000 tbs positions and 300 000 of them are cursed wins, imagine the impact on SF's playing strength and choice of moves, at each and every ply. that is the real problem, engine strength is adversely affected, why allow that?

no one knows if the starting position is a win for white or a draw, at least at the current point in time. Many presume it should be a draw, but white wins at TCEC and other high-end competitions only go on increasing in numbers... Top engine scores for most opening positions also only tend to increase... So please wait a while, before coming to a conclusion. that is actually one of the aims of extending the rule: to allow strong engines to tell us what the theoretical game outcome is, and they will never do that, if restricted by any meaningless rules.
The fact that white win so much in TCEC is only because engines do not use books but often start from positions when white is clearly better.

I expect more than 90% draws if stockfish and Houdini play without books
or use 2 move book and start a match of 100 games from the 50 most popular positions short lines of not more than 4 plies.
evidence contradicts your claim, as book has been more or less always favouring white, while white wins with each new edition only grow. migth want to look at earlier editions, where black scored quite often.

also, you can not explain rising SF score with newer versions for the starting position, as well as most main lines.

on the contrary, play without books or very short book with equal bias to a book with longer lines will only bring more wins, as engines have more opportunities to go wrong.

Lyudmil Tsvetkov
Posts: 6031
Joined: Tue Jun 12, 2012 10:41 am

Re: fortress_draw_rule

Post by Lyudmil Tsvetkov » Mon Nov 28, 2016 7:48 am

btw., do you know that I am able to create a suit of 50 opening lines, that could be used in TCEC, knowing the style of the 2 finalists, a suit that will carry a very small bias with it, or no bias at all, highest scores in the range of less than 40cps for both engines, that will produce, if not 100, then at least 96 wins?

very easy to do, when you know the style of the engines.

but what purpose will that serve? Quite often a long, well-fougth draw, with lots of turnarounds, is worth much more to the spectator and as a quality of play than a monotonous, biassed-based win. I definitely prefer to see 2 such draws than 10 wins of the other kind.

Lyudmil Tsvetkov
Posts: 6031
Joined: Tue Jun 12, 2012 10:41 am

Re: Cursed win at TCEC

Post by Lyudmil Tsvetkov » Mon Nov 28, 2016 7:54 am

as predicted, first attack at an unfair rule might not quite achieve its aim.

second or 3rd attacks though will be decisive, as engine programmers and the community as a whole will, with increasing strength of the engines, become aware that rules that inhibit engine development are basically wrong and unwelcome.

User avatar
hgm
Posts: 22083
Joined: Fri Mar 10, 2006 9:06 am
Location: Amsterdam
Contact:

Re: Cursed win at TCEC

Post by hgm » Mon Nov 28, 2016 9:53 am

Actually it is more like a rear-guard skirmish by some soldiers that got isolated in the jungle and do not know the war is already fought and lost. You are running a few decades behind the facts. This very same issue was then discussed by people about 10,000 times wiser than you, and they came to the conclusion that, for a variety of reasons, it will just become an unclear and unworkable mess if you make exceptions to the 50-move rule of any kind. So they restored the original rule.

APassionForCriminalJustic
Posts: 415
Joined: Sat May 24, 2014 7:16 am

Re: fortress_draw_rule

Post by APassionForCriminalJustic » Mon Nov 28, 2016 12:37 pm

Lyudmil Tsvetkov wrote:btw., do you know that I am able to create a suit of 50 opening lines, that could be used in TCEC, knowing the style of the 2 finalists, a suit that will carry a very small bias with it, or no bias at all, highest scores in the range of less than 40cps for both engines, that will produce, if not 100, then at least 96 wins?

very easy to do, when you know the style of the engines.

but what purpose will that serve? Quite often a long, well-fougth draw, with lots of turnarounds, is worth much more to the spectator and as a quality of play than a monotonous, biassed-based win. I definitely prefer to see 2 such draws than 10 wins of the other kind.
You're all talk.

APassionForCriminalJustic
Posts: 415
Joined: Sat May 24, 2014 7:16 am

Re: fortress_draw_rule

Post by APassionForCriminalJustic » Mon Nov 28, 2016 12:43 pm

APassionForCriminalJustic wrote:
Lyudmil Tsvetkov wrote:btw., do you know that I am able to create a suit of 50 opening lines, that could be used in TCEC, knowing the style of the 2 finalists, a suit that will carry a very small bias with it, or no bias at all, highest scores in the range of less than 40cps for both engines, that will produce, if not 100, then at least 96 wins?

very easy to do, when you know the style of the engines.

but what purpose will that serve? Quite often a long, well-fougth draw, with lots of turnarounds, is worth much more to the spectator and as a quality of play than a monotonous, biassed-based win. I definitely prefer to see 2 such draws than 10 wins of the other kind.
You're all talk. The fact that you think that you can predict or influence 100 percent how many games the engines will win exemplifies your arrogance. There is absolutely no way to "force" these engines to win games (increase probability?! Yes). There have been countless games saved because the engines at this level are extremely resourceful.

Lyudmil Tsvetkov
Posts: 6031
Joined: Tue Jun 12, 2012 10:41 am

Re: fortress_draw_rule

Post by Lyudmil Tsvetkov » Tue Nov 29, 2016 10:31 am

APassionForCriminalJustic wrote:
APassionForCriminalJustic wrote:
Lyudmil Tsvetkov wrote:btw., do you know that I am able to create a suit of 50 opening lines, that could be used in TCEC, knowing the style of the 2 finalists, a suit that will carry a very small bias with it, or no bias at all, highest scores in the range of less than 40cps for both engines, that will produce, if not 100, then at least 96 wins?

very easy to do, when you know the style of the engines.

but what purpose will that serve? Quite often a long, well-fougth draw, with lots of turnarounds, is worth much more to the spectator and as a quality of play than a monotonous, biassed-based win. I definitely prefer to see 2 such draws than 10 wins of the other kind.
You're all talk. The fact that you think that you can predict or influence 100 percent how many games the engines will win exemplifies your arrogance. There is absolutely no way to "force" these engines to win games (increase probability?! Yes). There have been countless games saved because the engines at this level are extremely resourceful.
you are almost as a colleague of yours, who asserted my claim we will see more than 20 wins in current TCEC was ridiculous and 'mathematically impossible'. Still, we have just witnessed the 20th TCEC win, and the competition is still not finished. some people never learn.

of course, it is very easy to achieve 100 wins. I'll get 1000 opening positions, play 10 games each at 1' game, and, in the end, there will be at least 50 positions where, in spite of being non-biassed, both engines will score convincingly. not difficult to do, I do not know what bothers you.

Lyudmil Tsvetkov
Posts: 6031
Joined: Tue Jun 12, 2012 10:41 am

Re: Cursed win at TCEC

Post by Lyudmil Tsvetkov » Tue Nov 29, 2016 10:37 am

hgm wrote:Actually it is more like a rear-guard skirmish by some soldiers that got isolated in the jungle and do not know the war is already fought and lost.
you refer to the FARC?
hgm wrote: This very same issue was then discussed by people about 10,000 times wiser than you...
should be very wise, those guys...

User avatar
Guenther
Posts: 2226
Joined: Wed Oct 01, 2008 4:33 am
Location: Regensburg, Germany
Contact:

Re: fortress_draw_rule

Post by Guenther » Tue Nov 29, 2016 11:40 am

Lyudmil Tsvetkov wrote:
APassionForCriminalJustic wrote:
APassionForCriminalJustic wrote:
Lyudmil Tsvetkov wrote:btw., do you know that I am able to create a suit of 50 opening lines, that could be used in TCEC, knowing the style of the 2 finalists, a suit that will carry a very small bias with it, or no bias at all, highest scores in the range of less than 40cps for both engines, that will produce, if not 100, then at least 96 wins?

very easy to do, when you know the style of the engines.

but what purpose will that serve? Quite often a long, well-fougth draw, with lots of turnarounds, is worth much more to the spectator and as a quality of play than a monotonous, biassed-based win. I definitely prefer to see 2 such draws than 10 wins of the other kind.
You're all talk. The fact that you think that you can predict or influence 100 percent how many games the engines will win exemplifies your arrogance. There is absolutely no way to "force" these engines to win games (increase probability?! Yes). There have been countless games saved because the engines at this level are extremely resourceful.
you are almost as a colleague of yours, who asserted my claim we will see more than 20 wins in current TCEC was ridiculous and 'mathematically impossible'. Still, we have just witnessed the 20th TCEC win, and the competition is still not finished. some people never learn.

of course, it is very easy to achieve 100 wins. I'll get 1000 opening positions, play 10 games each at 1' game, and, in the end, there will be at least 50 positions where, in spite of being non-biassed, both engines will score convincingly. not difficult to do, I do not know what bothers you.
You said 20 wins for SF or more, unless I missunderstood you.
Now we have 12:7 (corrected standing).

(and even that 12:7 only because of a completely different opening choice system than before - the draw rate seems to be diminished by more than 10% because of that - which was still unknown at the point of my guess)

http://talkchess.com/forum/viewtopic.ph ... 12&t=62063

Psyck
Posts: 4
Joined: Tue Sep 17, 2013 6:17 pm

Re: Cursed win at TCEC

Post by Psyck » Mon Dec 05, 2016 1:52 pm

syzygy wrote:
basil00 wrote:Syzygy tablebase adjudication has been merged into cutechess: https://github.com/cutechess/cutechess/pull/183

This allows for up to 6 piece adjudication and proper handling of the 50-move rule.
Problem solved :D
This hasn't been done for the TCEC fork, so problem not yet solved according to Martin's comment on ChatWing.

On the bright side, the official ruling on game 17 has been announced via the final score published on Chessdom:

"Stockfish 8 won the 100 games Superfinal match against Houdini 5 with a total score of 54,5 – 45,5."
http://www.chessdom.com/stockfish-is-th ... -champion/

Hence this thread can finally be woven to sleep.
There is a chasm; between carbon and silicon; that the software cannot bridge.

Post Reply