Rybka Chess Community Forum
1. Nf3 Nf6 2. Ng5 Ng8 3. Nxh7 Nf6 4. Ng5 Ng8 5. Nh3 Nf6 6. Ng1 Ng8 ~1 (no h7 pawn) 7. d4 Nf6 8. Nf3 d5 9. c4 c6 10. cxd5 cxd5 11. Nc3 Nc6 12. Bf4 Qb6 13. Qb3 Qxb3 14. axb3 Bf5 15. e3 a6 16. Ne5 Nxe5 17. Bxe5 e6 18. Bxf6 gxf6 19. Kd2 Bd6 20. Bd3 Bg6 21. Bxg6 fxg6 22. h4 Ke7 23. g3 Rh5 24. g4 Rh7 25. Rh3 f5 26. gxf5 gxf5 27. Rg1 Kf7 28. f4 Rah8 29. Rgh1 Rh5 30. Ne2 Be7 31. Ng3 Rxh4 32. Rxh4 Rxh4 33. Rxh4 Bxh4 34. Ne2 Be7 35. Nc1
(35... Bb4+ 36. Kc2 a5 37. Na2 Kg6 38. Kd3 Kh5 39. Ke2 Bd6 40. Kf3 Kg6 41. Ke2 b6 42. Nc3 Rybka v2.3.2i20 (0:00.42)-0.11|d28
28.h5 maybe could have won?
No. Perhaps it could prolong a little bit the game.
After 28. h5 Rah8 29. Rah1 that pawn does not go any place. Now white rooks cannot leave the H column because then h pawn will fall, with a dead draw.
Roman plays very safe in the last game today. He simplified too much and conformed to draw easily.
Congratulations to GM Dzindzichasvili! If a pawn is worth about 200 elo this is a great performance! What would happen at longer timecontrols? I don't really believe in a chessprogram 250 elo stronger than Kramnik somebody was talking about. Not yet :) And also not within two years I think unless on 16 core machines or better.
A pawn has no constant value and I believe that at GM level a pawn worth more than 200 elo.
I also think that pawn+move has bigger value than a pawn.
I expect GM Dzindzichasvili to beat every human in similiar conditions.
True, pawn and move should be more than 200 elo Uri, Prof. Elo himself gave 50 +/- 3 elo for having the white pieces every time in master play, so that is about equal to half the value of the Bishop pair. About the value of a pawn not being constant, well, every different pawn was used so I think there is a constant average value for a pawn, maybe a bit above 200 at Grandmaster level but not all that much. I think that would depend on playing tempo again. As long as you don't go looking for analyzed wins with this opening advantage, the value of about 250 elo for having a pawn up and playing with the white pieces but not going for deeply analyzed lines seems a reasonable estimate to me.
Based on my own research pawn + move is easily more than 200 ELO. Let's do some back-of-the-cocktail-napkin analysis here. Uri, easy on the methodological critique, I'm on your side.
Consider that in both engine and human databases white wins slightly more than 54% of the time, all other things being equal.
Then there is the pawn differential. This one is trickier, but for the sake of discussion we can start with the assumption that a pawn is equal to 1.00. (This isn't how Rybka evaluates but let's concede that a pawn is worth something close to that.) Now, what is the success rate with a +1.00 evaluation? My calcs suggest about 78%.
Combine the two advantages and you have a likely 82% success rate for the white player where the two opponents were equally skilled.
Now convert that to an ELO advantage. Let me do this in shorthand, inaccurately but close enough: I think these conditions reduce Rybka's chances by roughly 512 ELO. Which means that conceding pawn + move puts a quad-core Rybka into the 2600 range.
I agree with your method, though I think that going from 78% to 82% by having white isn't quite generous enough because things change when you're near a tipping point where such and such evaluation "should" be worth a theoretical win every time--but that's a very difficult thing to quantify, so that might have to do for the time being.
However, 82% is nowhere near a 512 elo difference--it is actually very close to half of that: 263 points.
It all depends on your point of view. It is a 250+ handicap, but if you combine the handicap with the advantage the result is as I say when comparing performance ELOs.
But that's exactly what I'm saying: it already has been combined--in order to get a 500-elo advantage, the scoring percentage must be about 95%, which cannot happen.
Since we're dealing with Gaussian distributions, I think we can do as follows: take your 78% scoring percentage, find the elo difference required for that, and then add the elo difference that we know exists between being white and black--that should be independent of handicaps (though for reasons stated above, I don't think that it actually is--but it's the best that we can do): 78% is a 220 elo advantage, and having white pieces is a 60 elo advantage, giving a net elo advantage of 280 points. Now going in reverse, we see that we should expect equally-matched players to have white scoring 83.4%.
Convert 82% to an elo advantage and you get 263 points, not 512 points. Sorry but I don'get the shorthand method. At this range the elo calculation is not very accurate maybe let alone the numerical approximations (See the graph by Odd Gunnar Malin) but a score of about 80% is still usual enough.
Thanks. I don't dispute your chart as I don't dispute mathematics. My point is purely conceptual. If one player is given an advantage, is not the other player given a disadvantage? Would this not be reflected is comparative performance ELOs? My 500+ estimate is precisely a doubling of the 263 you cite (my inexactitude with respect to 500+ is due to minor inaccuracies in the method I outlined). I would argue that this number is the deduction you would make from Rybka's rating-list ELO (with the hardware it was allowed in the match) to arrive at its effective ELO under the match conditions. In this context the match being drawn makes a lot of sense.
I see what you're saying--but the Elo system simply takes into account differences in ratings: one player is given an advantage, and due to the fact that this is the same as the other player being given a disadvantage, there is such and such a rating point difference: you can think of it as either the advantaged player being given an elo bonus of X, or the disadvantaged player being given an elo penalty of X, but not both.
All right then, I get 1/2 point credit for coming up with a reasonable method to calculate the difference and we'll call it a day.
Watching the outcome of the match at pawn and move, Rybka, may have a punctuation elo between 2500-2600?
"Punctuation elo" seria una medida útil para calcular la puntuación inglesa con precisión exacta!
¿Qué es "Punctuaction elo"?
Gracias, como no suelo frecuentar foros de motores de ajedrez en español me confundí y creí que era algo diferente al "ELO system."
Gentlemen, I was being facetious. Punctuation consists of periods, commas, apostrophes, colons, semi-colons and such. I was suggesting that a "Punctuation ELO" might be a useful measurement of proficiency in English writing. What I believe he intended to say was "Performance ELO".
I think in this type of match and Pawn Move, punctuation elo, to give you a better idea of the difference between two opponents likely to 8 games. I am not talking about elo performance, which is when it comes to a player with different players with different elo.
Rybka, having approximately 500 elo a difference with respect to Dzindi without handicap, the probability of winning would be P = 94%.
We're having a translation problem, as punctuation elo in english means something like "Ortografía Elo" in spanish (referencing the punctuation orthography.)
I could say that Rybka performed worse than expected (even Vas was surprised), it's either a flaw on the system, that Dzindzi performs much stronger against a computer, that Rybka is weaker than we think against humans, that our concepts of handicaps aren't as good as we think (because we can't measure the value of the handicap exactly) or some other factor.
For me the problem is that chess engines play chess and only chess. they can't adapt to chess variants accordingly. firstly they can't realise the concepts of an opening with the absence of a pawn. secondly when in difficult position try again to find the best as they think position not to oppose practical problems. those problems are far from minor importanse and easy solution. although i think Rybka has about 550 elo difference from Dzindzi and i insist he is in no way stronger than his current elo those drawbacks reduse the difference about the half of that difference compare to a human player at the level of Rybka.So the final outcome seems fair both in chess board and to theory.
PS: I vote for 5-3 for Rybka estimate her performance about 100 elo above from what actually did but i was suprised finally from how short-eyed she manipulate the opening face lossing the opportunity to achieve reasonable complex positions forcing Roman to innacurancies and blunders.
PS: If i was in Larry's shoes i would never had begin such a match without a sufficient opening book in at least the most crucial handicaps because is well known that without O.B. Rybka plays pathetic.
P.S: Once more Rybka perform in a match well above expectation due to bad preperation of the team. I can remember a single occasion when Rybka suprise as possitive with her performance in a match.
P.S: I would be interest to see what Jeroen's genuine can do for those handicap books.
> PS: I vote for 5-3 for Rybka estimate her performance about 100 elo above from what actually did but i was suprised finally from how short-eyed she manipulate the opening face lossing the opportunity to achieve reasonable complex positions forcing Roman to innacurancies and blunders.
I voted 5-3 for Roman but admittedly that was after the result of first games.
> although i think Rybka has about 550 elo difference from Dzindzi and i insist he is in no way stronger than his current elo those drawbacks reduse the difference about the half of that difference compare to a human player at the level of Rybka. So the final outcome seems fair both in chess board and to theory.
I think that in a 'normal' match, that is without pawn handicaps but also without opening books, or maybe a very short neutral book for both, Dzindzichasvili could still excel against computers, and would do better than the 550 elo difference there theoretically is.
> P.S: I would be interest to see what Jeroen's genuine can do for those handicap books.
Well, there I would take an opposite view: please no big handicap books! I think Larry and Vas would agree, these are okay for avoiding total predictability of the computer moves in the opening but beyond that such a book would defeat the purpose of finding out how well the computer plays on its own. I think it is high time that computers learn to play the opening better on their own, as you also suggest Asos, that this is possible Rybka has shown by becoming clear first in the CCRL Chess960 list. Maybe it is an idea for Vas to ask Ray to let the present or next beta of Rybka 3 play some FRC matches to see how she performs relative to Rybka 2.3.2a in the CCRL list. I think it would establish a useful benchmark apart from other testing.
[url=]I think that in a 'normal' match, that is without pawn handicaps but also without opening books, or maybe a very short neutral book for both, Dzindzichasvili could still excel against computers, and would do better than the 550 elo difference there theoretically is.[/url]
the opening books are inevitable if you want not to repeat always the same opening lines. maybe again a human can perform a little bit better than his elo simply because elo formula is not so accurate between players with elo difference so high. only a draw after 5-6 games with white and all the black games defeats is enough to close that gap. i believe that against a 2800 GM Rybka can achieve around 80%.
[url=]I think it is high time that computers learn to play the opening better on their own[/url]
i also believe that but is if not impossible at least very difficult to do so having the same time and evaluation you use for midlegame also in the opening. opening has principles much different than midlegame and both of them different than the endgame. maybe you need a completly new code for each part of the game not only some parametrs which change their values accordingly. But finaly it's very very dificult to perform as good as 150 years of development and practise after only a few minutes of search.
I made a test some time ago with several engines playing without opening book, including Rybka, and Rybka performed some 200 ELO points (!) weaker than expected! The error bars were high, but it's expected that Rybka would perform even worse if her opponents had a book!
This isn't shown in Chess960 because Rybka faced engines that were good at playing from the opening position as their programmers added special code for that, but this code isn't specially helpful at other opening positions besides the starting one.
Anyway, I just think that the not tuned opening book of Rybka hasn't been considered in the handicap, and that it could be a bigger factor than they think.
I think, if the Contempt of Rybka, It weakens chess and not playing with full force, My question, Do we know that performance has Rybka with elo Contempt?
I have established that a small contempt setting does no measurable harm and will generally help against weaker opponents (in the case of Rybka 3, that should mean all opponents). We might even make such a setting the default. The setting of 50 as used in this match does hurt performance against closely rated opposition (maybe something like 10-15 Elo I think) but should help against weaker opponents by avoiding draws (or in handicap games by avoiding trades).
Certainly part of the handicap is that the computer has hardly any opening book in these pawn odds games. In theory I could have created a serious book for each of the eight handicaps, but aside from the fact that I am just too busy to do this, my main goals were just to make Rybka unpredictable (in case Roman had prepared openings against Rybka with no book) and to prevent getting a position out of the opening that was much worse than being a pawn down. For this second goal I did most of my work on the f7 handicap, as in this case it is quite easy for Black to emerge with an awful position after natural moves, and in fact we stayed in book in the f7 game all the way to the queen exchange. I should have devoted more time to the g7 opening, as this one is also very hard to play as Black, but I don't really know how Black should play that handicap. Other than the g7 game I don't think Rybka emerged from the opening much worse than being just a pawn down.
i can't wait for the possible rubber matches!
I think he could have gone for more but did not want to risk. Stll, a drawn Rybka match is a very impressive result. According to the supposed value of the pawn handicap, he performed on the level of Kasparov. While he did play well (even with his bad day), this suggests to me that engine ratings are quite a bit inflated.
I think the problem lies in the attempt to express odds (of any kind) in elo differences. they are really not that easy to match. at the top level, I believe a great majority of pawn odds games would end with wins, so the elo calculation would become almost insignificant. I don't think Rybka's elo is inflated if we take into consideration that there's no grandmaster who ever lived on earth that can (or could at their best) be competitive against Rybka in a championship-like match (at least 10 games). if you look at it from that point of view, maybe it should even be higher.
> I don't think Rybka's elo is inflated if we take into consideration that there's no grandmaster who ever lived on earth that can (or could at their best) be competitive against Rybka in a championship-like match (at least 10 games).
Maybe Paul Morphy, as I recall that he was playing Rybka's moves, or better (!).
The next step could be to leave play top chess programs in the big tournaments GM?
Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill