Cursed win at TCEC

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

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
User avatar
hgm
Posts: 22334
Joined: Fri Mar 10, 2006 9:06 am
Location: Amsterdam
Full name: H G Muller
Contact:

Re: fortress_draw_rule

Post by hgm » Fri Nov 18, 2016 12:42 pm

The 50-move rule is a FIDE rule. Proposing not to apply it is changing the FIDE rules.

I don't buy the 'more interesting' argument. The games would still be adjudicated as soon as a tablebase position is reached. They would just receive a different score. What is interesting about that?

Unless you consider it interesting to see if engines would fall into the trap you set up for them by lying about the rules, or, by sheer coincidence would manage to avoid it. If that sort of suspense is the point, we could also before each game designate a randomly chosen piece other than King as the royal piece, and adjudicate an immediate win for whichever side captures that.

Note that this game would never have become a cursed win if Houdini had known that cursed wins would be counted as wins, rather than draws. Abolishing the 50-move rule would not have made any difference when the engines would have been aware of it, rather than being tricked. It still would have ended as a draw. So how would that have been more interesting?

Ralph Stoesser
Posts: 408
Joined: Sat Mar 06, 2010 8:28 am

Re: fortress_draw_rule

Post by Ralph Stoesser » Fri Nov 18, 2016 1:15 pm

The 50 move draw rule was not created with destroying inevitable checkmates in mind. Humans can't usually calculate that far, so a outright 50 move rule for all cases makes sense as long as humans are involved. But if only KI is involved and one side can prove that it can checkmate in any number of moves, we as humans should be happy and amazed about it. Remember, this game is about checkmate in the end, isn't it? :)

The TCEC rules are not clear enough in that case. To consider the TCEC decisions unfair or stupid for those cases makes therfore sense. I agree with that. Can you elaborate in detail how and why this game would have ended in case both engines would have used gaviota tablebases for example? I can't.

Ralph Stoesser
Posts: 408
Joined: Sat Mar 06, 2010 8:28 am

Re: fortress_draw_rule

Post by Ralph Stoesser » Fri Nov 18, 2016 1:30 pm

FIDE rules are made from humans for humans, but not for pure engine competitions. In that sense they stay untouched by this proposal.

User avatar
hgm
Posts: 22334
Joined: Fri Mar 10, 2006 9:06 am
Location: Amsterdam
Full name: H G Muller
Contact:

Re: fortress_draw_rule

Post by hgm » Fri Nov 18, 2016 2:08 pm

I thought I saw someone write in the threat that initially it was draw even without 50-move rule, and then Houdini started doing moves that would be sub-optimal in the chess variant without 50-move rule because it did not care, so that in the end it converted to a cursed win. I did not look at the game myself; that is more effort than I am willing to invest on this.

I think the TCEC rules are very clear in stating the 50-move rule applies. Whether it would make a better game when the 50-move rule was replaced by a more complex rule is another discussion. It is actually a hard problem. It is very unsatisfactory to introduce rules that are dependent on technological progress, in this case the availability of end-game tables. It is quite conceivable that engines will be able to generate the relevant slices of 9, 10 or 11-men EGT with many blocked Pawns on the fly. Suppose an engine that can do that claims it sees a mate-in-1000 with 900 moves to conversion in a 9-men position... What do you do now, after 50 reversible moves?

The 50-move rule is a real bitch. On the one hand you need something to force stubborn players (and none are more stubborn than computers...) into accepting a draw. OTOH, it is often difficult to define what progress is. This is especially problematic in games that forbid perpetual (= repetitive) checking, but easily allow 30 checks in a row before the possibilities to check without repeating are exhausted. Then the losing side can stall for 30 moves every time you do one move to make some progress. Which is especially painful in games where the eventual win comes by promoting a piece that can also move backwards. Because even if you would make only forward moves reset the ply counter, the piece could just be shuttled forward-backward to reset the counter without making any progress. So it becomes virtually impossible to define what progress is.

I guess the best solution would be this: a side that has a win can request a larger number of moves than 50 in order to achieve it. But if he does so, and doesn't manage to win (e.g. because the game converts to insufficient material, or runs into a repetition), he will be declared loser. That should be enough to discourage frivolous use, and limit extension requests for cases where the player indeed sees a certain win.

syzygy
Posts: 4258
Joined: Tue Feb 28, 2012 10:56 pm

Re: fortress_draw_rule

Post by syzygy » Fri Nov 18, 2016 9:26 pm

Ralph Stoesser wrote:To remove the stalemate rule would be a major change. Most chess studies would be worthless. The character of the game would change too much. Additionally it would result in a change of the FIDE rules, but nobody here proposes to change the FIDE rules. The proposal is to overrule the 50 move rule in case of

1) pure engine competition
2) a tablebase checkmate has been found in current position
First, how does that not change the FIDE rules?

Second, what means "a tablebase checkmate has been found in current position"?

In game 17, SF did not find a checkmate or a tablebase checkmate. Btw, what is the difference between "checkmate" and "tablebase checkmate"?

Perhaps you mean that the 50-move rule should be ignored in any position that somewhere in the world or in this universe by calculation has been determined to be a checkmate (for a chess-like game that is not FIDE chess) prior to the date that the TCEC game is played? Do you care at all whether the engines know about this calculation and have access to it? Do you care at all whether the TCEC adjudication mechanism has access to it?

Or do you mean that whatever the TCEC TB adjudication mechanism returns must be accepted as correct, whether the TBs it uses are faulty or not?

Third, I don't get why people are looking into changing a clear rule into something funny just to solve this problem when there is the very simple and correct solution of adjudicating to the correct result. It is trivial to look up the correct result of 5-piece positions. In addition, the engines are aware of the correct results and choose their moves having in mind the correct results.

Ralph Stoesser
Posts: 408
Joined: Sat Mar 06, 2010 8:28 am

Re: fortress_draw_rule

Post by Ralph Stoesser » Sat Nov 19, 2016 12:29 am

1) Why should that change the FIDE rules? They stay exactly the same. The question should rather be why a piece of software should try to 100% conform to a set of rules that were created for human entities under special tournament conditions? Does it make sense? In all cases?

2) Again: Look at the ICCF rules.

10.Tablebase Adjudication

a. ICCF acknowledges some tablebases as valid for claiming win/draw/loss in positions solvable with the following tablebase: Convekta Ltd, which solves all positions with maximum 6 men. Each certified tablebase will be available on the ICCF Webserver system.

b. In case the tablebase shows a win that supersedes the 50 moves rule, the win will be awarded

3) For engine competition an inevitable mate should also take priority over a rule that has been created for human competitions under special thinking time constraints. Is that so hard to accept? What do you think why ICCF has such an adjucation rule that superseeds the 50 move draw rule? What is the idea behind that horrible FIDE rules derision? :shock: Note that this rule is from January 2016. So they could have used "correct" tablebases. Your tablebases are not so universally important as you might think. They are usefull in all cases where following the FIDE rules is important: Study of FIDE games with computer assistance and perhaps in case an engine should ever again participate in an human FIDE tournament.

syzygy
Posts: 4258
Joined: Tue Feb 28, 2012 10:56 pm

Re: fortress_draw_rule

Post by syzygy » Sat Nov 19, 2016 1:08 am

Ralph Stoesser wrote:1) Why should that change the FIDE rules?
Do you not realise that the 50-move rule is a FIDE rule?
http://www.fide.com/fide/handbook.html? ... ew=article
Article 5.2 e)
The question should rather be why a piece of software should try to 100% conform to a set of rules that were created for human entities under special tournament conditions? Does it make sense? In all cases?
We've been through this 10 times now. Yes, it makes sense because without it engine-engine games last thousands of moves. And no, a rule like "ignore the 50-move rule when I like it" simply does not work.

And I will repeat this: applying the same rules to human-human, human-engine and engine-engine games poses no problems whatsoever. And why fix a problem that does not exist...
2) Again: Look at the ICCF rules.
At least they have put it in the rules.

But the rule is stupid.

First, why do they have it? Answer: because they thought there was no alternative. They thought TBs necessarily ignore the 50-move rule. That is wrong, but they did not know better. With proper TBs, there is no problem and there is no reason to change the game to fit the shortcomings of TBs.

Second, do you realise what they have done? They let Convekta decide the outcome of games. They "certified" the Convekta TBs... do you think they checked that those TBs are correct? Certainly they did not. They just placed blind trust in the correctness of what Convekta is offering.
3) For engine competition an inevitable mate should also take priority over a rule that has been created for human competitions under special thinking time constraints. Is that so hard to accept?
Do you know the meaning of what you are saying?

How do you know that a position with 20 pieces left on the board and which is approaching 50 moves has no "inevitable mate"? You do not know that. The engines may seem unable to make progress and they may actually be unable to make progress, but we have no way of knowing that there is no "inevitable mate". The concept you are proposing is broken. Unless you can prove a draw, which practically will never be possible, you will have to play out the games until electricity runs out.
Note that this rule is from January 2016.
Wrong again. At least try to get the facts rights...

The rule took effect on 1 January 2014. Long before that date I had added Syzygy50MoveRule to my SF TB-fork, which fully complies with their silly rule. My TBs happen to include cursed and real wins, so there is no problem whatsoever. But that does not stop the ICCF rule from being misguided.

And ICCF picked 6-men. Why 6? Why not 7? Now if an engine has 7-piece TBs you still can't trust the result.

Lyudmil Tsvetkov
Posts: 6033
Joined: Tue Jun 12, 2012 10:41 am

Re: fortress_draw_rule

Post by Lyudmil Tsvetkov » Sat Nov 19, 2016 8:33 am

hgm wrote:I thought I saw someone write in the threat that initially it was draw even without 50-move rule, and then Houdini started doing moves that would be sub-optimal in the chess variant without 50-move rule because it did not care, so that in the end it converted to a cursed win. I did not look at the game myself; that is more effort than I am willing to invest on this.

I think the TCEC rules are very clear in stating the 50-move rule applies. Whether it would make a better game when the 50-move rule was replaced by a more complex rule is another discussion. It is actually a hard problem. It is very unsatisfactory to introduce rules that are dependent on technological progress, in this case the availability of end-game tables. It is quite conceivable that engines will be able to generate the relevant slices of 9, 10 or 11-men EGT with many blocked Pawns on the fly. Suppose an engine that can do that claims it sees a mate-in-1000 with 900 moves to conversion in a 9-men position... What do you do now, after 50 reversible moves?

The 50-move rule is a real bitch. On the one hand you need something to force stubborn players (and none are more stubborn than computers...) into accepting a draw. OTOH, it is often difficult to define what progress is. This is especially problematic in games that forbid perpetual (= repetitive) checking, but easily allow 30 checks in a row before the possibilities to check without repeating are exhausted. Then the losing side can stall for 30 moves every time you do one move to make some progress. Which is especially painful in games where the eventual win comes by promoting a piece that can also move backwards. Because even if you would make only forward moves reset the ply counter, the piece could just be shuttled forward-backward to reset the counter without making any progress. So it becomes virtually impossible to define what progress is.

I guess the best solution would be this: a side that has a win can request a larger number of moves than 50 in order to achieve it. But if he does so, and doesn't manage to win (e.g. because the game converts to insufficient material, or runs into a repetition), he will be declared loser. That should be enough to discourage frivolous use, and limit extension requests for cases where the player indeed sees a certain win.
Rules should be changed, when they are unjust, which is precisely the current case.

Whe you have a law, that is unjust, and you have to decide based on that law, what you do: take a decision following the unjust law, or change the law to accomodate justice? Any decent person would prefer to change the law.

The game was already a white win: SF had been showing 105cps from the early mg, Houdini some 80cps. That is a win, however you would like to look at it. Why penalise SF because some tablebase rule artificially prunes the mating sequence, rendering it a draw instead? The claim that H went for the tb position because it saw it is 0.0, and otherwise black had better options, is simply wrong, as, most probably white had a win throughout. It was KQPP vs krbb, the3 pieces are much stronger than the queen, but apart from having 2 passers, white also has the additional advantages that they are 1) connected, 1) very advanced - 5th and 6th rank, 3) provide the white king a shelter, 4) the white king is very advanced, supporting its passers, 5) the black king lacks pawn shelter. Not sufficient? I guess the white position, barring a miracle, is simply won throughout, tbs or not, so why penalise the stronger engine and the won game?

Concerning your ridiculous claim that KP vs K with stalemate position also favours white, but is still a draw, how on earth you compare this absurd position with the tbs mate? In KP vs k and stalemate, the position is simply draw, as there is no mate, while in the referenced TCEC position, this is a win, there is mate. If you agree that DRAW=WIN, then you are perfectly rigth, otherwise, why you make such impossible claims? Is DRAW=WIN?

Lyudmil Tsvetkov
Posts: 6033
Joined: Tue Jun 12, 2012 10:41 am

Re: fortress_draw_rule

Post by Lyudmil Tsvetkov » Sat Nov 19, 2016 8:47 am

syzygy wrote: Third, I don't get why people are looking into changing a clear rule into something funny just to solve this problem when there is the very simple and correct solution of adjudicating to the correct result.
Many reasons:

1) the rule is unjust, it penalises the side having a win, you artificially transform a win into a draw.
2) what is the purpose of wdl, when, according to current rules, W=D?
3) tablebase mates in more than 50 moves are relatively frequent, there are at least 500 of those discovered as of now, certainly more will be added later; why take a blindfold stance, claiming those are all draws?
4) my artistic sense gets bitterly offended, when all side's efforts and an objectively won position are declared a draw; I guess everyone else's artistic sense gets offended too?
5) rule is simple to implement and without detriment to the overall game development; you split the rule into 2 portions, one concerning the general 50-moves rule for the standard game situation, not to make the game too long, and a second one providing for the specific case of pawnless endgames, where the 50-moves rule will be adjusted to 100-move rule or higher, to accomodate the won endgames. Very simple.
6) Rules change with progress. Until we knew there are tbs wins longer than 50 moves, the 50-move rule was adequate, now, when we know there are such longer wins, the 50-move rule becomes outdated and should be changed.

User avatar
hgm
Posts: 22334
Joined: Fri Mar 10, 2006 9:06 am
Location: Amsterdam
Full name: H G Muller
Contact:

Re: fortress_draw_rule

Post by hgm » Sat Nov 19, 2016 9:51 am

Lyudmil Tsvetkov wrote:Rules should be changed, when they are unjust, which is precisely the current case.
Rules that are changed as one sees fit whenever they have to be applied do not qualify as rules. One decides upon rules in advance, and makes all involved aware of those. In particular one does not penalize participants for obeying rules they were told to obey, because as an afterthought you think the agreed-upon rules are bad or sub-optimal. That seem pretty elementary properties of the rule concept.
Whe you have a law, that is unjust, and you have to decide based on that law, what you do: take a decision following the unjust law, or change the law to accomodate justice? Any decent person would prefer to change the law.
There is nothing 'unjust' about the 50-move rule, anymore than that it is unjust that a Bishop can only reach squares of one color, or that stalemate is a draw, or that in case of a 3-fold repetition the side with more material doesn't win.
The game was already a white win: SF had been showing 105cps from the early mg, Houdini some 80cps. That is a win, however you would like to look at it.
That is total bullshit. I have seen Stockfish and Houdini score positions as +1150cp, and still lose.
Why penalise SF because some tablebase rule artificially prunes the mating sequence, rendering it a draw instead?
Because that 'artificial' rule was the one adopted for the tourney seems a pretty good reason to enforce it. To me, at least...

Concerning your ridiculous claim that KP vs K with stalemate position also favours white, but is still a draw, how on earth you compare this absurd position with the tbs mate? In KP vs k and stalemate, the position is simply draw, as there is no mate, while in the referenced TCEC position, this is a win, there is mate. If you agree that DRAW=WIN, then you are perfectly rigth, otherwise, why you make such impossible claims? Is DRAW=WIN?
Not draw=win, but stalemate=win. Logically most stalemates are wins, because the stalemated side is forced by zugzwang to move his King into check, upon which it will be captured and he loses. It is only a draw because of the artificial FIDE rule that says it is a draw. But in fact it is just a position where you cannot avoid your King will be captured on the next turn. So this rule unfairly robs the opponent of the opportunity to capture the King, by terminating the game one half-move before it with an artificial draw result.
Last edited by hgm on Sat Nov 19, 2016 10:03 am, edited 3 times in total.

Post Reply