< Earlier Kibitzing · PAGE 2 OF 2 ·
|Jul-09-15|| ||sfm: How is it now, is it a draw if fifty moves have been made without moving a piece?|
|Jul-09-15|| ||piltdown man: No, it's fifty moves without capturing a piece or moving a pawn.|
|Jul-09-15|| ||Penguincw: Last computer vs. computer GOTD: Arasan vs Critter, 2014 (Mar/29/2015). The last time before that (Stockfish vs Jonny, 2014, Nov/02/2014), used this pun.|
|Jul-09-15|| ||Zhbugnoimt: after 148.Na8!! Black IS forced to accept. If he does not, he loses. For example: 148...Be7 149.Bf6! Bf8 150.Nb6 Bb7 151.Nd7 1-0 150...Kb7 151.Nxc8 Kxc8 152.Kxa6 or 152.Kb6 1-0.|
|Jul-09-15|| ||Mudphudder: LOL...with computer vs computer matches it's almost expected that if there are going to be any double-exclam moves then it would have to be at least beyond the 100th move of the game.|
|Jul-09-15|| ||solskytz: <Zhbugnoimt> I'm not convinced. There's also 148...Kb7 when it seems that black can set a fort later with ...Kc7-Kc8. |
Besides, the concept is not less astonishing, remarkable and impressive even if accepting the sacrifice WAS forced.
|Jul-09-15|| ||AylerKupp: <ASchultz> remember when they'd fall for the Noah's Ark trap due to the horizon effect. Bit different here. Yup.>|
No, not really. Computers still fail due to the horizon effect. It's just that with improvements in hardware, search tree pruning, and extension calculation the horizon is a lot further away than it used to be. I don't think that any algorithm which is based on forward searching will ever be immune from the horizon effect, it's just a question of where that horizon is. I have egotistically coined what I call "Ayler Kupp's corollary to Murphy's Law" which you can find in my forum's header.
The most extreme example I came across is a position, which I dearly wish I had saved, when the engine evaluated the position as slightly favorable for White even though Black had a mate in one next move.
|Jul-09-15|| ||AylerKupp: <<dumbgai> Komodo seems to start aimlessly shuffling pieces around at move 26. Then just when it looks like neither side could make progress, it goes for a suicidal pawn grab and loses.>|
I call that effect "engine dithering", when an engine doesn't know what to do and moves its pieces back and forth without any apparent purpose. This typically happens in closed positions when the horizon effect strikes again and the engine can't search deep enough to reach a position where it finds an advantage for itself. And I can only guess that Komodo could not search deep enough to determine that the pawn grab was suicidal.
Another example of a suicidal engine is this game: Aylerkupp / Rybka vs Kutztown46 / Stockfish, 2011. It's initially interesting because Stockfish makes a positional pawn sacrifice with 11...c5 in order to get some counterplay (yes, engines can do that sometimes). After much engine dithering, Rybka retains the pawn advantage for almost 80 moves. But then it comes up against the 50-move rule. Since Rybka evaluated the position to its advantage ( [+0.65] ) after 90...Rc8, it finds a draw result unacceptable and plays 91.b4, sacrificing a pawn. Since after the pawn sacrifice it still evaluates the position as favorable ( [+0.16] ), it's satisfied with the result.
But the circumstances for this game were somewhat unusual and Rybka is very short of time. What transpires is then a comedy of errors when Rybka's search depth kept getting smaller and smaller until our opponent was gracious enough to offer a draw on principle. At the end Rybka still evaluated the position in its favor but, with the low search depth, such an evaluation was not reliable.
|Jul-09-15|| ||kevin86: My nomination for the most boring game of the year.
White finally sacs a piece to break open this dreadfully locked position (sorry Bob Marley), or else they still may be playing.
|Jul-09-15|| ||maxi: <solskytz> 148...Kb7 loses immediately to 149.Nb6 forcing the exchange of N for B (after which Black's a-Pawn falls just the same, -but without the Knight sac-).|
The effect of 145.Na8 is to place Black in Zugzwang, forcing him to eventually play e6 which robs Black (White squares) Bishop of his escape square e6. When 148.Na8 comes around for a second time, taking the Knight is forced.
|Jul-09-15|| ||maxi: Some of you are not giving the engines their due. This is a deep game.|
|Jul-09-15|| ||morfishine: I don't analyze 158 move games though I like the finish here|
|Jul-09-15|| ||SChesshevsky: <AylerKupp:... And I can only guess that Komodo could not search deep enough to determine that the pawn grab was suicidal.>|
I'm thinking "the suicide effect" could be a bit more common than one might expect from the perceived all-knowing computer.
Though I'm sure it's rare on a total game % basis, I'm wondering what it would be on a lost games %.
Which leads me to the questions.
Assuming not doing the pawn grab draws, what would be the protocol for a computer vs computer draw? Do they have to wait for the 50-move rule or 3-times rep or can a computer offer a draw before hand?
Assuming both computers know White's better around move 109...Bg7 due to the open a5 square and weak b7, better but maybe not that much better. Is it possible for Komodo to play weaker evaluation moves that likely still draw or is its mandate to play the move that may be stronger for a win but more risky?
Just looking at the position at 109..., without calculating, it looks like if Black just put the K on c8, B on e8 and just shuffled the DSB or N, I'm not sure White can make much progress. At least compared to the text, so why take the risk Komodo?
|Jul-09-15|| ||AylerKupp: <Schesshevsky> Any perception of an all-knowing computer is wrong for no other reason that an engine's evaluation of positions depends on the accuracy of its evaluation function, and evaluation functions represent a tradeoff between accuracy and speed of calculation. If the evaluation function takes too long to calculate it will hinder the engine's ability to search deeply in a given amount of time and its play will suffer relative to other engines whose developers have made a better tradeoff between accuracy and speed. So, if using a computer for analyses, <ALL> computer lines should be validated for correctness in the face of the potential impact of he horizon effect, over ambitious search tree pruning, and other considerations such as fortresses which engines currently seem to have problems evaluating correctly. This validation can be done either by a human (if he/she is sufficiently strong) or by other engines that have different evaluation functions.|
As far as attempting to answer your questions:
1. I have no idea what the % of suicide moves influence the % of lost games, other than to remark that in a human game a suicide move would referred to as a "blunder" and is understandable in many circumstances like time trouble or simple "humanness". So I wouldn't know how to characterize a computer's move as a "suicide move" or simply a "blunder" and calculate the %.
2. The protocol for a computer vs. computer draw or a computer resignation is typically not implemented in the engine but in the GUI. For example, when using the Arena GUI in engine vs. engine matches, the GUI can declare a loss for one side if the evaluation falls below a user-specified value for a specified number of consecutive moves (typically 3). The GUI can also adjudicate a draw using tablebases, when a specified number of moves are played with the evaluation at [0.00], or simply when the game reaches a specified number of moves. And other GUIs will likely have different criteria.
3. I think that all engines will play the move that they evaluate as best in any given position; they won't play a move that they consider inferior in an attempt to swindle an opponent. <Sally Simpson> and I have had several discussions along these lines and whether it would be worthwhile to have a "swindle mode" where, in a lost position, an engine might play a move evaluated to be inferior in an attempt to induce its opponent to make an inferior and possibly losing move in return. My opinion is that it is probably not worthwhile to implement and use such a "swindle mode" against a computer opponent unless the move results in an advantage that could not be perceived without a very deep search, and that the engine is able to search deeper than its opponent.
4. I don't know what version of Komodo was used in the 2015 WCCC but the latest version (9.1) evaluates the position after 109...Bg7 at [+0.48], d=35, so it assessed that White was somewhat better. Putting the Black king on c8 and the Black LSB on e8 results in an evaluation of [+0.49], d=36, essentially the same.
Komodo has a parameter called "Drawscore" (similar to "Contempt" in other engines) with a default value of -7 which is used to try to prevent draws against real or perceived weaker opponents, so Komodo will try for a win even if it evaluates its position (as Black) at [-0.07] or higher. The Komodo developers suggest that a larger (more negative value) like -15 be used against older or much weaker engines, and I suspect that they might have felt that Jonny (ranked #56/57 in the latest CCRL engine vs. engine tournament list) fell in this category. But Jonny was supposedly running on a 2400-core system so that would to a large extent compensate for many if not all software deficiencies so, if the developers used a <very> high value for Drawscore, I don't think that was justified. So perhaps that's why Komodo might have felt obligated to take a risk and go for a win in spite of White's advantage, but that's just a guess on my part. However, given their suggestions, I don't think that the developers would have used a sufficiently large value for Drawscore to force Komodo to go for a win in spite of the magnitude of the evaluated advantage for White after 109...Bg7. But, who knows?
|Jul-09-15|| ||solskytz: <Maxi> of course the pawn falls - but I wasn't sure that white will be able to win this position, if black just shuffles his K from c7 to c8 and vice versa after that.|
|Jul-09-15|| ||solskytz: The point was not so much the loss of the a-pawn, as the penetration of the king into the whole black baseline.|
|Jul-10-15|| ||SChesshevsky: <AylerKupp: Komodo has a parameter called "Drawscore" (similar to "Contempt" in other engines) with a default value of -7 which is used to try to prevent draws against real or perceived weaker opponents, so Komodo will try for a win even if it evaluates its position (as Black) at [-0.07] or higher.>|
Thanks for the terrific information. I don't have a chess computer program so I'm probably extremely uninformed.
I had no clue about this "anti-drawing" parameter. It may explain some of those rare occurrences when a computer forgoes a likely draw for more sharp play. It seems that parameter is very human. I'd guess when set to an extreme it might forsake a lot for the "pride" of not drawing a weaker opponent.
If that was the case for Komodo in this game, maybe another case of pride does goeth before a fall.
|Jul-10-15|| ||AylerKupp: <SChesshevsky> The idea behind the anti-drawing parameter is simple. Normally you're satisfied with a draw if you consider yourself equal or weaker than your opponent and are not satisfied with a draw if you consider yourself stronger than your opponent. Or maybe the tournament/match circumstances are such that a draw will not do you much good; you need a win at (almost) all cost. These various anti-drawing parameters modify the value assigned to a draw, normally [0.00], to the specified value in an attempt to avoid draws as much as the value represents.|
This comes into effect when the engine is trying to decide which move to make. Normally it will choose the move that results in the highest evaluation of the position. So if the engine evaluates its position to be at a slight disadvantage, it will select a move sequence that results in a draw by repetition if there is one. But if the anti-draw parameter is other than zero, the engine will think that a move that results in a draw by repetition has a lower evaluation than a move that results in an evaluation that is only slightly inferior (negative) and so reject the drawing move.
A large value for this anti-draw parameter effectively means that the engine has "contempt" for its opponent (and hence the often-used term Contempt for this anti-drawing parameter) and "thinks" that it can outplay it/him/her even though its evaluation of its position is inferior. Who says that hubris can't be programmed into a computer! :-)
|Jul-11-15|| ||RandomVisitor: 2015 tournament results:
Pos Name Gms W D L Score SBgr
1 Jonny 8 6 2 0 7.0 24.25
2 Komodo 8 6 1 1 6.5 20.00
3 Hiarcs 8 3 4 1 5.0 14.75
4 Protector 8 3 4 1 5.0 14.25
5 Shredder 8 3 3 2 4.5 12.00
6 Ginkgo 8 2 4 2 4.0 9.75
7 The Baron 8 2 2 4 3.0 5.50
8 Maverick 8 1 0 7 1.0 0.00
9 Fridolin 8 0 0 8 0.0 0.00
|Jul-11-15|| ||Penguincw: One thing I forgot to mention: this is currently the 2nd longest game in the <cg> database played in 2015. It is one move short of M Vachier-Lagrave vs Ponomariov, 2015.|
Also longest decisive game since Morozevich vs Tomashevsky, 2013.
|Jul-11-15|| ||QueensideCastler: According the bulletin - 129... Nxa4 was a severe mistake.|
|Oct-06-15|| ||MindCtrol9: Because these computers are super strong,I am not going to say that 15.Nh4 is a bad move or something like that.I like 15.Ba5 instead of the computer's choice,but this is just at a glance.|
|Oct-06-15|| ||MindCtrol9: The capture of the pawn 129....Na4 brought the penetration of the white King and winning the game which from my point of view,had to be a draw if it not for the 129 move by Black.|
|Nov-18-15|| ||AzingaBonzer: Komodo grabbed the pawn due to its contempt setting being too high.|
|Nov-18-15|| ||kontoleon: Where is the posibe continuation after 145... KXa8?
151...a5 the black open the game and finally lose|
< Earlier Kibitzing · PAGE 2 OF 2 ·