I myself have a lot of trouble beating Rybka at knight odds on my quad in blitz, but I never thought it would be any challenge to do so at game in one hour. As I said, Joel Benjamin had little trouble at 3'+1", and he is rated only 55 points above the GM you mention. Was he using a quad, and what did he do to create variety in the openings? Perhaps I'll experiment myself with longer games. It is really hard to believe this!
In order to find out for myself how tough it is to beat Rybka at knight odds, I have started a knight odds match under fairly similar conditions, but with a faster time control, 20'+10" (the time limit used in the FIDE World Rapid Championship recently). This is the latest version of Rybka as used against Benjamin, running on the same 3.2 GHz quad, with a small knight odds book for variety (no repeat openings). Of course Rybka is White every game, and a draw counts as victory for Rybka, so no need to set contempt. So far I won the first game rather easily, by shedding a pawn to simplify to an easily won ending, and didn't even need to use half my time. It's a bit hard for me to imagine losing this match, but of course I am slightly handicapped by perhaps being ambivalent about the result, like a teacher playing his student. Still, I promise to do my best to win each game.
![[gb]](/mwf/v12/flags/gb.png "United Kingdom")
Sorry, lets be real. I am convinced that there are 1000s of players around the world (me included) that for sure can win with an extra knight even in blitz games. A knight is simply to much! But maybe the anti human version can put up more resistance?
In my view black with out pawn f7 has also an impossible task. I played a few such game (5 min + 3sek) and though I managed to lose a few games in the beginning (In the first game I was stupid enough to play 1.d4 that loose some of the advantage), and as soon as I had found a good and solid approach it was mainly just a matter of technique to drive home the advantage.
(Well, maybe I would change my mind if I were to play Rybka on a faster machine - mine is a single processor 3.6 GHz).
In my view black with out pawn f7 has also an impossible task. I played a few such game (5 min + 3sek) and though I managed to lose a few games in the beginning (In the first game I was stupid enough to play 1.d4 that loose some of the advantage), and as soon as I had found a good and solid approach it was mainly just a matter of technique to drive home the advantage.
(Well, maybe I would change my mind if I were to play Rybka on a faster machine - mine is a single processor 3.6 GHz).
![[us]](/mwf/v12/flags/us.png "United States")
Ahh...but the "good and solid approach" you found worked against a particular opening or type of opening...unless you had a special opening book for this, too? I think that will make a bit of a difference here. After all, it's a lot easier for Kramnik to draw against Fritz when he knows what's coming than when he doesn't...
Having a book for f7 handicap does make a big difference, as does the hardware, but I'm inclined to agree with you that it may just be too big a handicap to give to a grandmaster at FIDE time control and expect to win a match. If you want to check this out, play more of these games but rotate Black's first move among 1...d6, 1...e6, 1...Nc6, and 1...Nh6, adding a second move where needed to avoid the line you have already refuted. If you do this please let us know what happens. By the way, do you have a blitz rating on ICC or Playchess?
We discussed takeback handicap before and I think it is an interesting idea.
But as you said, Larry, it wouldn't scale very good. If the skill difference is too large the number of takebacks would be ridiculously high...
Still, I think it is an interesting challenge. How many takebacks do you need to draw/win against Rybka at a certain timecontrol?
Suggested rules:
- Each takeback is counted as one point. There could be that you notice your mistake a couple of moves after and have to do 5 takebacks, this will count as 5 points.
- The time is still running, so it is always possible to lose on time.
- Takebacks are not allowed to be taken back :-)
Maybe this is not a good way to practice... but I will try myself.
But as you said, Larry, it wouldn't scale very good. If the skill difference is too large the number of takebacks would be ridiculously high...
Still, I think it is an interesting challenge. How many takebacks do you need to draw/win against Rybka at a certain timecontrol?
Suggested rules:
- Each takeback is counted as one point. There could be that you notice your mistake a couple of moves after and have to do 5 takebacks, this will count as 5 points.
- The time is still running, so it is always possible to lose on time.
- Takebacks are not allowed to be taken back :-)
Maybe this is not a good way to practice... but I will try myself.
Tried a 2 6 game vs. Rybka 2.3.2a without contempt or opening book.
Almost gave up, but I finally managed to lock the position positionally for 50 move draw after 76 takebacks :-)
[Event "Computer chess game"]
[Site "D2"]
[Date "2008.01.09"]
[Round "?"]
[White "John"]
[Black "Rybkav2.3.2a.w32"]
[Result "1/2-1/2"]
[Time "20:31:30"]
[WhiteElo "2200"]
[TimeControl "120+6"]
[Termination "normal"]
[PlyCount "166"]
[WhiteType "human"]
[BlackType "human"]
1. d4 e6 2. Nf3 Nf6 3. Bg5 Nc6 4. e4 h6 5. Bxf6 Qxf6 6. c3 Qg6 7. Nbd2 d5 8. Qe2 Be7 9. g3 O-O
10. Bg2 Bd7 11. O-O Rfe8 12. e5 b6 13. b4 Qc2 14. Rfc1 Qa4 15. Qd1 Qxd1 16.
Rxd1 a5 17. b5 Na7 18. a4 c6 19.bxc6 Nxc6 20. Rdb1 Rab8 21. Bf1 g5 22. Bb5 g4
23. Ne1 f5 24. f4 h5 25. Ng2 Kf7 26. Nh4 Rec8 27. Rb3 Rc7 28. Kg2 Bd8 29. Rab1 Be8
30. R1b2 Bxh4 31.gxh4 Ne7 32. Bxe8 Rxe8 33. Nb1 Rc4 ...with 50 more moves
Almost gave up, but I finally managed to lock the position positionally for 50 move draw after 76 takebacks :-)
[Event "Computer chess game"]
[Site "D2"]
[Date "2008.01.09"]
[Round "?"]
[White "John"]
[Black "Rybkav2.3.2a.w32"]
[Result "1/2-1/2"]
[Time "20:31:30"]
[WhiteElo "2200"]
[TimeControl "120+6"]
[Termination "normal"]
[PlyCount "166"]
[WhiteType "human"]
[BlackType "human"]
1. d4 e6 2. Nf3 Nf6 3. Bg5 Nc6 4. e4 h6 5. Bxf6 Qxf6 6. c3 Qg6 7. Nbd2 d5 8. Qe2 Be7 9. g3 O-O
10. Bg2 Bd7 11. O-O Rfe8 12. e5 b6 13. b4 Qc2 14. Rfc1 Qa4 15. Qd1 Qxd1 16.
Rxd1 a5 17. b5 Na7 18. a4 c6 19.bxc6 Nxc6 20. Rdb1 Rab8 21. Bf1 g5 22. Bb5 g4
23. Ne1 f5 24. f4 h5 25. Ng2 Kf7 26. Nh4 Rec8 27. Rb3 Rc7 28. Kg2 Bd8 29. Rab1 Be8
30. R1b2 Bxh4 31.gxh4 Ne7 32. Bxe8 Rxe8 33. Nb1 Rc4 ...with 50 more moves
Hmm, 76 takebacks to get a draw. How many to get a win? Please don't feel obligated to spend the rest of your life finding out!
I won't, instead I assert the numerological hypothesis that one takeback equals 10 ELO ;-)
(My blitz skill = 2300) + 76*10 + (black disadvantage = 40) = (3100 = Rybkas ELO)
(My blitz skill = 2300) + 76*10 + (black disadvantage = 40) = (3100 = Rybkas ELO)
Based on this sample of 1 game, you would be correct. But I'm inclined to believe that the value of a takeback should be more than 10 Elo, perhaps in the 20-30 range.
![[hu]](/mwf/v12/flags/hu.png "Hungary")
Each takeback is probably worth less than the previous one. Do we really think that a 1000 player with 200 takebacks could beat Rybka?
Vas
Vas
Yes, you must be right. My 20-30 estimate was meant for the first few takebacks.
![[is]](/mwf/v12/flags/is.png "Iceland")
Do you have a clue what kind of a match will be organized next?
I think it would be interesting to have a chess match where the player has computer assistance. Perhaps a program that has elo similar to himself. At least that could lead to great chess.
Also the latest match, with draw odds was really interresting, and if a new and fresh player would like to play that could be interesting.
I would think some computer assistance should be used because it ruins the game if a bad blunder happens by the human.
I think it would be interesting to have a chess match where the player has computer assistance. Perhaps a program that has elo similar to himself. At least that could lead to great chess.
Also the latest match, with draw odds was really interresting, and if a new and fresh player would like to play that could be interesting.
I would think some computer assistance should be used because it ruins the game if a bad blunder happens by the human.
This idea has been suggested by others (including Joel Benjamin) and could certainly be tried if someone offers sponsorship for it. It should lead to interesting chess, but it does not provide any information about the strength of Rybka as the combined strength of a GM + engine is pretty much unknown. I myself am more interested in some sort of explicit handicap match, but I would be willing to host such a match if someone else sponsors it. My concern is that if the engine is rated at or below the level of the GM Rybka would still win easily, whereas if the engine is significantly stronger than the GM it might be more of an engine vs. engine game with occasional human involvement.
I'd like to see another draw odds match, maybe with a higher rated GM, this time with a win or draw worth 1.5 or two points.
I think that making a draw worth more than one point would really defeat the purpose of a game and also really put some strange ideas in people's heads. If one is really interested in such a thing, then it could simply be said, "if he gets more than two or three draws, he gets the full prize money" or something like that. I'm more in favor of simply the same thing as what we just had but against a higher-rated GM.
I completely agree. The only reason for having a draw worth more than a half point is so those rooting for the human have a slim chance of being able to say "Yeah, we won!" :)
The number of points awarded for a draw (or win) makes no difference to anyone but the player. It doesn't change the player's strategy or his chances of actually getting the win or draw; it only affects his chances of collecting whatever bonus we offer for winning the match. Awarding more points for a win than for a draw does change the strategy of both sides and might make some sense, but it's pretty clear now that the chances of anyone actually winning a game against Rybka (assuming a modest contempt factor) are negligible. The main benefit of awarding more points for a win would be to force us to be a bit more modest in setting the contempt factor. The most natural rule would be to award triple points for a win, since then the scoring of a draw would be exactly half way between a win and a loss, as they are in normal chess. In effect, this would be the same as normal scoring but with 25% being deemed a drawn match.
As for a stronger player, it's pretty clear that we would need a much stronger player for such a match to have any suspense. Nakamura is the only such player to have indicated a willingness in the past to play for the level of prizes we have been able to offer. If anyone knows of another player over 2650 FIDE who would be willing to play such a match for the modest sums we can offer, please have him get in touch with me.
As for a stronger player, it's pretty clear that we would need a much stronger player for such a match to have any suspense. Nakamura is the only such player to have indicated a willingness in the past to play for the level of prizes we have been able to offer. If anyone knows of another player over 2650 FIDE who would be willing to play such a match for the modest sums we can offer, please have him get in touch with me.
> it's pretty clear now that the chances of anyone actually winning a game against Rybka (assuming a modest contempt factor) are negligible
This statement really shows just how good these programs are getting. When was the last time a human beat Rybka under tournament time controls?
Rybka has played 32 official games with human opponents (all Grandmasters) since I joined the team at the end of 2005. Half were at pawn odds (with varying time limits and colors), and half were at White odds (with varying additional handicaps and time limits). Rybka only lost 1 of the pawn odds games (excluding a forfeit replayed to a draw) and did not lose any of the White odds + games (11 wins and five draws). Except for the sole pawn handicap loss, I am not aware of any official game that Rybka has lost to a human opponent over the board. Perhaps someone else has the statistics for the year 2005. Of course I'm sure that some of the forum readers have won games from Rybka, but generally there is no way to verify the conditions of play.
Nakamura would be great to get into such a match. In spite of my having made fun of him on this forum for his infamous statement regarding Rybka, I still have great respect for him as a chess player and would have quite a lot more (and I'm sure others would, too) if he was to agree to such a thing. I would see it as also having a bit more potential for Rybka publicity than with many of the other players. He's playing at 2700+ strength now, so you'd better snatch him up (if you can) before he officially cross that "barrier" on the FIDE list! :-)
I'll find out if he still has an interest.
Matches with computer assistance could also be interesting, but they are not "pure". You have to start setting limits on the grandmaster's hardware and software, and these limits are a little bit arbitrary.
It's probably worth trying, though.
Vas
It's probably worth trying, though.
Vas
One thing that makes a judgement about an engine's positional evaluation difficult is, that the eval it shows isn't for the current game position on the board but for another one, which may be quite different, at the end of the pv something like 8 or 10 full moves ahead. Sometimes I have to remind myself about that, if an eval seems mysterious. Maybe the impressions would be better if those positions where the eval comes from would be reviewed. Or alternatively, the program "Der Bringer" displays the eval for the current board position before search, in addition to the usual search infos.
At least some programs in the chessbase family can display the final position being evaluated. Perhaps Rybka 3 will have that capability, I don't know as I do not work on GUI issues.
Thanks... Due to my screen resolution and preferred layout, I never use the variation display in Fritz, which does that for the (1st) pv. With Rybka, I think it would only make sense though if "Display PV tips" is activated.
Btw. on this occasion, I want to mention that in the german CSS computer chess message board, the interest in the recent match was quite big. Especially considerations about what might come next, which conditions make sense etc. resulted in a big discussion with more than 40 postings in 2 days.
My conclusion is, that the major novelty in another match should be a stronger GM, maybe 2650+ (being aware that it is difficult to get one).
Btw. on this occasion, I want to mention that in the german CSS computer chess message board, the interest in the recent match was quite big. Especially considerations about what might come next, which conditions make sense etc. resulted in a big discussion with more than 40 postings in 2 days.
My conclusion is, that the major novelty in another match should be a stronger GM, maybe 2650+ (being aware that it is difficult to get one).
Note that Rybka's PV doesn't always lead to the position from which the score originates, even with "Display PV tips" activated. This is because Rybka takes the PV from the hash table, which sometimes contains a better move than the minimax move.
Vas
Vas
![[au]](/mwf/v12/flags/au.png "Australia")
I wonder about that larry (i.e., that humans have better evaluation). Could it be that we can learn something from Rybka, namely, that some of these static features are "overrated." Maybe rybka holds the truth and humans are struggling to catch up?
This question is actually at the top of my priority list to answer. So far evidence is mixed.
![[il]](/mwf/v12/flags/il.png "Israel")
I think that the evaluation of a GM cannot be compared to the evaluation of a chess program.
Disadvantage of humans:
GM's do not evaluate all factors of position to have a specific number and say that they prefer position A that is 0.34 pawns for white and not position B that is 0.33 pawns for white.
Disadvantage of computers:
Computers do not learn to change their evaluation in the middle of the search.
Computers may not have specific knowledge.
Uri
Disadvantage of humans:
GM's do not evaluate all factors of position to have a specific number and say that they prefer position A that is 0.34 pawns for white and not position B that is 0.33 pawns for white.
Disadvantage of computers:
Computers do not learn to change their evaluation in the middle of the search.
Computers may not have specific knowledge.
Uri
I think that people rely on evaluation aided by limited search, while computers rely on search aided by limited evaluation. This makes sense given the many order of magnitude difference in search speed.
Larry has indicated that he believes Rybka's static evaluation is getting pretty good, which I find pretty surprising.
Regards,
Alan
Larry has indicated that he believes Rybka's static evaluation is getting pretty good, which I find pretty surprising.
Regards,
Alan
Well, "good" is relative. Rybka's static evaluation is vastly improved over the last year, but not terribly high by human standards, perhaps FM level. As Uri points out though, such comparisons are difficult to make. Rybka is much better than any human at adding up the pluses and minuses in any one position, but human masters are much better at judging which factors are important.
Are the terms in the eval function independent from the depth of the search? That is, if you held matches between engines using only eval and Qsearch and improved the eval function based on these results, would it hold for deeper searches? If would certainly be a lot easier to test and optimize eval coefficients if search and eval could be decoupled in this way. Conversely, if the ideal coefficient weights are coupled to the search depth things would get a lot more complicated if you wanted to maintain the best eval function, both from an implementation and from a testing perspective.
By the way, FM strength eval seems like a huge improvement. At the beginning, I asked Vas how well the eval plus Qsearch would work and he estimated about 1400. Even if his guess was off by a few hundred Elo points, this would be incredible progress (or maybe Rybka's Qsearch has become more extensive?).
Regards,
Alan
By the way, FM strength eval seems like a huge improvement. At the beginning, I asked Vas how well the eval plus Qsearch would work and he estimated about 1400. Even if his guess was off by a few hundred Elo points, this would be incredible progress (or maybe Rybka's Qsearch has become more extensive?).
Regards,
Alan
This is a topic of discussion between Vas and I. It is clear that the best evaluation function at short depths is highly correlated with the best eval at deep searches, but it is also clear that the difference is non-trivial. We have mostly been optimizing for short searches, hoping that the difference at long searches is not too great, but we may revisit this. As for the second topic, there is a huge difference between asking what the rating of Qsearch+eval would be and estimating how well computers evaluate positions. When I speak of FM level eval, I mean that given a quiescent position, Rybka's static eval should be of comparable quality to that of an FM who does not consider any variations. Of course if an FM actually looks at a position, he will start to analyze, and will of course evaluate at a much higher level than Rybka without search.
Still, the quality of the Rybka eval has improved by nearly a class in the past year, not considering time considerations.
Still, the quality of the Rybka eval has improved by nearly a class in the past year, not considering time considerations.
Two questions and a comment:
1) How long, Larry, would you estimate it will be until Rybka has a GM-level static evaluation? For me, that would seem to be a huge leap in ability for a program, something that in fact programmers for years have been aiming for and saying would be a sign of incipient "intelligence" rather than just a brute force search. I understand that it would in fact NOT constitute intelligence, but it would be nevertheless fascinating to me that a fixed set of rules could, in a sense, quantify GM-level static evaluation.
2) I understand you and Vas are using the matches with GMs in part to show how strong Rybka is. But, for a 1900 rated patzer like myself, and someone who just likes to follow computer chess for its own sake (and has for 25 years, since "Mephisto II" and "Computer Chess Reports"!), playing odds against Rybka is really the only way I can play Rybka "for real" (i.e. full-strength), so I found these matches fascinating in their own sake. As a few posters have asked for, in Rykba 3, in addition to the more sophisticated Contempt factor that you developed for these matches, could you put a place in the menu and opening book for having Rybka play in 1)"Draw-Odds" Mode, 2) "Knight-Odds" Mode, 3) "Pawn odds" mode, 4)"Queen-Odds"(!) Mode, or combinations thereof? Of course it doesn't mean that much to "power" users, but there is a real thrill in playing against, in effect, a Kasparov-level player who is giving odds, and watching how skilled they are at equalizing. I think more people would find the feature exciting than you would imagine. I am a coach of a high-school chess club, and I know my students would love to have Rybka play with Queen or Knight Odds against them, for example.
3) Congratualtions on all the victories this past year. And on this forum-- it is great to see you and Vas being so open about so much.
Tony
1) How long, Larry, would you estimate it will be until Rybka has a GM-level static evaluation? For me, that would seem to be a huge leap in ability for a program, something that in fact programmers for years have been aiming for and saying would be a sign of incipient "intelligence" rather than just a brute force search. I understand that it would in fact NOT constitute intelligence, but it would be nevertheless fascinating to me that a fixed set of rules could, in a sense, quantify GM-level static evaluation.
2) I understand you and Vas are using the matches with GMs in part to show how strong Rybka is. But, for a 1900 rated patzer like myself, and someone who just likes to follow computer chess for its own sake (and has for 25 years, since "Mephisto II" and "Computer Chess Reports"!), playing odds against Rybka is really the only way I can play Rybka "for real" (i.e. full-strength), so I found these matches fascinating in their own sake. As a few posters have asked for, in Rykba 3, in addition to the more sophisticated Contempt factor that you developed for these matches, could you put a place in the menu and opening book for having Rybka play in 1)"Draw-Odds" Mode, 2) "Knight-Odds" Mode, 3) "Pawn odds" mode, 4)"Queen-Odds"(!) Mode, or combinations thereof? Of course it doesn't mean that much to "power" users, but there is a real thrill in playing against, in effect, a Kasparov-level player who is giving odds, and watching how skilled they are at equalizing. I think more people would find the feature exciting than you would imagine. I am a coach of a high-school chess club, and I know my students would love to have Rybka play with Queen or Knight Odds against them, for example.
3) Congratualtions on all the victories this past year. And on this forum-- it is great to see you and Vas being so open about so much.
Tony
GM level static evaluation would probaby be a few years away, but before that happens I think some version of the Monte Carlo method will be used that will in effect simulate GM level evaluation. As for Rybka 3 features, I don't work on the GUI or commercial book, but these would be nice extra features. I would be willing to contribute a small handicap book if requested, and the requested options would be simple to add. But except for the book question, it is easy enough to play handicap games with Rybka, just enter setup mode and remove the handicap piece. For draw odds, just set contempt high.
Actually, I wonder about how qualitively comparable human and computer "evaluations" really are.
Here are some (non-expert...) thoughts:
Rather than a concrete mathematical calculation of relatively few exact general factors, human (positional) eval is based of a big database of learned structures or "blocks". A relatively few of these are fuzzily matched and intuitively "summed".
There is some famous experiment where players of different strengths had to remember the pieces in a number of positions. Some were random and some where from actual games. There was a strong correlation between chess skill and number of pieces for the actual games, while the correlation was quite weak for the random positions, which may be explained by the "database of fuzzily matched blocks" - theory.
And the human doesn't evaluate black and white separately but rather compares each factor across the players. Which combination of structures for resp. side that "beat" the other.
Who has the advantage on the queenside, on the open line etc. Then the human perhaps uses some high level logical formulas for how to combine factors, which may have dependencies that is missed in a simple linear summing.
And surely, there are quite many other heuristic meta-techniques involved as well, depending on the intelligence level :-)
Also, the distinction between search and eval for humans seems quite problematic. Human evals include some kind of non-exact "guess"-searches for potential, are they search or eval?
So I guess it is a big strategic question if you really should try to mimic human thought and evaluation,
or if you should let the computer itself empirically self-improve with some hands-on help. I mean, the AI-issue is probably a hard one :-)
Here are some (non-expert...) thoughts:
Rather than a concrete mathematical calculation of relatively few exact general factors, human (positional) eval is based of a big database of learned structures or "blocks". A relatively few of these are fuzzily matched and intuitively "summed".
There is some famous experiment where players of different strengths had to remember the pieces in a number of positions. Some were random and some where from actual games. There was a strong correlation between chess skill and number of pieces for the actual games, while the correlation was quite weak for the random positions, which may be explained by the "database of fuzzily matched blocks" - theory.
And the human doesn't evaluate black and white separately but rather compares each factor across the players. Which combination of structures for resp. side that "beat" the other.
Who has the advantage on the queenside, on the open line etc. Then the human perhaps uses some high level logical formulas for how to combine factors, which may have dependencies that is missed in a simple linear summing.
And surely, there are quite many other heuristic meta-techniques involved as well, depending on the intelligence level :-)
Also, the distinction between search and eval for humans seems quite problematic. Human evals include some kind of non-exact "guess"-searches for potential, are they search or eval?
So I guess it is a big strategic question if you really should try to mimic human thought and evaluation,
or if you should let the computer itself empirically self-improve with some hands-on help. I mean, the AI-issue is probably a hard one :-)
Well, I guess we do a little of both.
I don't think that the difference between human eval and computer eval is really huge, but there are some differences.
Humans are efficient in pattern recognition. A computer could also search for patterns such as those that humans use, but it would take a lot of CPU cycles, so programmers tend to avoid it.
Computers are efficient at arithmetic - for example, counting the exact number of squares controlled by a bishop, and performing some sort of calculation on each of those squares. A human could also use this exact same approach, but it would take a lot of "neural cycles", so humans tend to avoid it.
Vas
Humans are efficient in pattern recognition. A computer could also search for patterns such as those that humans use, but it would take a lot of CPU cycles, so programmers tend to avoid it.
Computers are efficient at arithmetic - for example, counting the exact number of squares controlled by a bishop, and performing some sort of calculation on each of those squares. A human could also use this exact same approach, but it would take a lot of "neural cycles", so humans tend to avoid it.
Vas
simulating evaluation that is at least GM level evaluation is very easy.
You only need to change your evaluation function to return result of search to depth that is big enough.
If you say that it is not evaluation then evaluating mobility is also not evaluation because it is possible to claim that you practically search to see squares that the piece can go and you do not evaluate.
Uri
You only need to change your evaluation function to return result of search to depth that is big enough.
If you say that it is not evaluation then evaluating mobility is also not evaluation because it is possible to claim that you practically search to see squares that the piece can go and you do not evaluate.
Uri
You make a good point. The distinction between search and eval is sometimes rather fuzzy.
Indeed. Some eval terms have the sole purpose of helping with the search. And some eval terms require a form of 'searching'.
Vas
Vas
We can certainly distribute this book, if you're interested in providing it.
My WAG for 2500-level static chess eval: 8 years from today.
Vas
My WAG for 2500-level static chess eval: 8 years from today.
Vas
Sure, a small handicap book for variety won't take a lot of time. But with the current Fritz/Shredder GUIs, if you use setup to remove the handicap it will not recognize the book; only a move sequence to remove the piece will allow book recognition.
![[nl]](/mwf/v12/flags/nl.png "Netherlands")
Imo 1.d4 is the only appropriate opening choice in such a match.
I think there was good preparation for that, though--with good opening choices by black, I think it's more difficult for white to close things up than with 1.e4; Joel lost both games when he used this. This is quite contrary to much of the practice we've seen, but that practice has evolved. In particular, I think that black can eventually typically think about ...c5 pushes and, if necessary, ...e5.
Vas & LK,
In terms of Rybka's new understanding of contempt, would you consider the outcome of this match to be a success? I only ask as Benjamin scored on par with engines rated well above him, granted, with draw odds. Personally, I did not expect him to score more than a single point. What more can be said of Rybka's understanding of contempt?
In terms of Rybka's new understanding of contempt, would you consider the outcome of this match to be a success? I only ask as Benjamin scored on par with engines rated well above him, granted, with draw odds. Personally, I did not expect him to score more than a single point. What more can be said of Rybka's understanding of contempt?
I am reasonably happy with how contempt worked. Of course it is far from perfect, but the new terms we added seemed to help at least marginally, and now I can try to optimize their values before Rybka 3 release. I note that our 21 move victory can be directly credited to contempt, as the decision of which way to recapture on e5 was determined by the contempt value. At 50 or less Rybka would have recaptured with the knight, and with the resultant exchange of knights the game might well have ended in a draw. At 60 or more (it was set to 75 for this game) the recapture with the pawn is chosen to avoid trades, and in this instance that was the key to a quick win.
lobby to get nakamura to play. up the prize fund if u have to. let naka get access to an opening database for the first 15 moves or so if that's what it takes to lure him in. and give him double the time as rybka gets.
i'd pay good money to see his ass struggle for draws, "my brain is better than rybka 6 days out of the week" and all...
i'd pay good money to see his ass struggle for draws, "my brain is better than rybka 6 days out of the week" and all...
Handicap games are not interesting to me. Gms should refuse to play under those conditions...except for the color advantage. I too would pay to see any of the top ten GMs play Rybka on even terms.
Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill