Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / New Match with FM John Meyer
1 2 Previous Next  
Poll What will be the score of the four game match at "Pawn and three move" odds? (Closed)
Meyer by 4-0 3 5%
Meyer by 3.5-0.5 2 4%
Meyer by 3-1 7 13%
Meyer by 2.5-1.5 8 14%
Tie Match (2-2) 14 25%
Rybka by 2.5-1.5 6 11%
Rybka by 3-1 10 18%
Rybka by 3.5-0.5 2 4%
Rybka by 4-0 4 7%
- - By lkaufman (*****) Date 2008-06-30 03:01
     By defeating Rybka convincingly (4-0) at knight odds, FM John Meyer (FIDE 2284) earned the right to challenge Rybka at a lesser handicap. According to chess tradition of the 18th and 19th century, the next smaller recognized handicap is "pawn and three moves", meaning White starts with two moves, plays the third move as White, and Black plays without the f7 pawn. So we have agreed to play another four game match this way, this time with no draw odds for either side. Match is this Sunday July 6 starting at noon Eastern Daylight time (6 p.m. central European time). I expect that coverage will again be on the Convekta "Chess Planet" site. Time limit remains the same at 30' plus 30" per move increment for each side. Because without restrictions White could open 1e3, 2Bd3, and 3Qh5 with mate in two more moves, we agreed that White will start each game with 1e4 and 2d4. This permits the game to begin with a definite initial position and normal chess rules of alternating color. I don't think anyone can argue that these two moves are suboptimal. Prize money is again $100 for match victory and $50 for a drawn match.
     Of special interest is the timing of the match. It corresponds roughly with the date Rybka 3 should be finalized, except perhaps for bug fixes and minor tweaks. So this should be the first public exhibition of a near-final Rybka 3. The match will use the same octal computer as used in the last match, again with a small handicap opening book. I plan to use some contempt factor, haven't decided how much yet.
     Some comments: According to Rybka herself, the pawn and three move handicap is around 80-85% as big a handicap as knight odds. Of course the material advantage is far smaller, but White has a big lead in development, a much safer king, good attacking chances, and Black is likely to experience serious difficulties in getting castled. On the other hand, White's strategy is not obvious as it is in the case of knight odds. Wholesale trading is likely to lead to an endgame that Rybka will probably not lose; White should try to make use of all of his advantages. Also, I believe that Rybka will play this handicap much better than she does knight odds, because her eval is tuned to normal chess where pawn down games happen all the time, whereas piece down games are just resigned. So despite the 4-0 embarrassment to Rybka at knight odds, both John Meyer and I consider this new match to be a fair one which either side might win. John does plan to prepare in the openings this time, which is a real threat to Rybka because there are far more forcing lines available at this handicap than at knight odds.
Parent - - By Kappatoo (*****) Date 2008-06-30 09:53
Maybe this is a naive question: Are you sure that it isn't possible to simply work out a forced win here?
Parent - - By lkaufman (*****) Date 2008-06-30 15:09
What is a "forced win"? I am sure that mate cannot be demonstrated, and almost sure that no win of a piece can be proven. It's clear that White will achieve a winning position by force, but not one as clearly winning as Knight odds. If the handicap is pawn and four moves (with no restriction other than staying in the first four ranks), a win of a second pawn can be proven, and at pawn and five a win of the Exchange plus a second pawn is forced. Perhaps that's why these handicaps do not appear in the historical records.
Parent - - By Uri Blass (*****) Date 2008-07-01 10:43
I believe that the distance to mate is smaller relative to games with knight odd.

I am not sure that the winning position white is going to get is  "not one as clearly winning as Knight odds"

It may be interesting if people give different programs many hours from the initial position to see if there is a significant jump in the score after a long time(like +2 after 1 minute-> +3.5 after 48 hours).

I am not sure if rybka is the best program to find a forced win here.

Parent - By lkaufman (*****) Date 2008-07-01 17:49
Your points are all good ones. I'll just add that even if some computer proves that White can achieve an advantage greater than knight odds, it does not mean that a human could remember all the analysis necessary to reach such a position against all plausible defenses. In this particular match, I expect that John will analyse the opening on his own and perhaps with his brother Eugene (a strong IM, and possibly a candidate for some future match) rather than with computers; he's just not a computer person.
Parent - By duncan (**) Date 2008-07-01 20:32
if not rybka , which program would be the best one to find a forced win ?

Parent - - By duncan (**) Date 2008-07-03 09:51
question to uri,

if rybka is not the best one to look for a forced win, which one would you say is ?
Parent - By Uri Blass (*****) Date 2008-07-03 10:09
I do not know.
I only said that it is not obvious that rybka is the best program to find a forced win because winning games and finding a forced win are different tasks.

Parent - By turbojuice1122 (Gold) Date 2008-07-03 12:30
Rybka Winfinder 2.2 perhaps
Parent - - By duncan (**) Date 2008-07-01 21:06
please could you explain a bit more how if the  handicap is pawn and four moves a win of a second pawn can be proven ?

thank you
Parent - - By lkaufman (*****) Date 2008-07-01 22:43
1e3, 2Bd3, 3Qg4, and 4Nc3. Check it out with any engine, you'll see that Black must shed the "e" pawn and allow a queen trade. Without the knight on c3 Black can avoid the loss of the second pawn by running his king out to c6. White would have a nice advantage, but perhaps not as much as by normal moves.
Parent - - By duncan (**) Date 2008-07-02 17:29
question to Mr kaufman

PM by andrew williams on a material-only search from from the starting position, gets 33 ply in just over 5.5 hours. This was about 3 years ago on a 500 mhz machine I believe.

About how long would winfinder (64 bit) take to find 33 ply on a q6600. ?
Parent - - By lkaufman (*****) Date 2008-07-02 19:38
I have no involvement with the old "winfinder" as I work on evaluation, which is minimized in winfinder. But I understand it still does a real evaluation, not just material, so the answer to your question must be "a very long time". It is the fact that all scores are equal until material is won that makes it possible for a material-only search to reach such depths. Add in even one piece of knowledge, say "mobility", and most of the depth gains will disappear.
Parent - - By duncan (**) Date 2008-07-02 20:16
thanks for your reply. Is there any possibility of  a pure  material only winfinder coming out? if yes with the new search alogorithms of rybka 3 how long do you estimate it would take to get to 33 ply from starting position. (q6660) if no is there any program you know that does a good material only search ?

or should I put this question to vas.?
Parent - - By lkaufman (*****) Date 2008-07-02 22:27
I'm sure that Vas could do this easily, but I don't think there is enough interest in it to justify such a version. Only Vas could answer your second question, but it has little meaning as a ply has different meanings in different programs, unless they are just full width and capture search like thirty years ago.
Parent - - By Uly (Gold) Date 2008-07-02 22:46

> I don't think there is enough interest in it

It's because the engine outputs useless garbage:

34=  +0.00   13:09:24.491898767447   1.a3 a6 2.a4 a5 3.b3 b6 4.c3 c6 5.b4 axb4
                                     6.cxb4 e6 7.Bb2 Nf6 8.b5 cxb5 9.axb5 Rxa1
                                     10.Bxa1 g6 11.Nc3 d6 12.Nf3 e5 13.e3 Ke7
Parent - - By lkaufman (*****) Date 2008-07-03 00:18
Another problem with the idea is that proving the win of a pawn (or even two) does not tell us much, as there may be full compensation.
Parent - - By Miroslav Kvíčala (**) Date 2008-07-06 10:48
From my point of view three moves + f7 pawn is too big advantage for FM level player...I tried it yesterday 3min. per game and all two games I drawn easily :)
Parent - - By lkaufman (*****) Date 2008-07-06 17:20
Judging from the first game today, it looks like you are right. Probably pawn and two moves would be much more balanced.
Parent - By SR (****) Date 2008-07-06 17:22
Why is there no link to the game?
Parent - - By duncan (**) Date 2008-07-03 23:49
regarding the differnce in ply I think vas said winfinder can be 4-10 times faster which I assume to be 3 extra ply.(2^3) = 8

I saw an estimate that rybka can do 30 ply in 17 hours, not sure what machine, guess a 6600 or better.;hl=ply
"1 minute*2^10=1024 minutes or 17 hours and 4 minutes"

as winfinder  has extra 3 ply  so winfinder is 33 ply in 17 hours

compare to post modernist

33 ply on 500 mhz machine took 5.5 hours

a core 2 duo 6600 has almost 5 times the mhz + 3 times as efficient * 3 times for the 4 processors

= 5 *3*3 = 45 which is more than 5 ply

so material only pm can do 38 ply in 5.5  hours or 40 ply in 22 hours

(5.5 *4)

which would give pm more than a 6 ply lead over winfinder (40-33)

Mr williams said his software proves that there is no material gain for white in the first 33 moves which means I guess it is a full width search.

so hypothetically a pure material only winfinder would get a gain more than 6 ply, but with vas's very efficent algorithms worth lets say 2 ply, it would be 8 ply.

any commennts on these estimates.?
Parent - By Roland Rösler (****) Date 2008-07-04 00:35
any commennts on these estimates.?

I think a Winfinder without material eval would be faster! Mate is the aim! Any comments? :-)
Parent - - By lkaufman (*****) Date 2008-07-04 02:29
Personally I would find it more interesting if the material-only search valued the bishop-pair as half a pawn (close enough to the truth for this purpose). I am reasonably certain that no win on a pawn by White can be proven in any reasonable depth (and probably there is no such win possible even with unlimited depth), but there is a much higher probability (though still small) that a win of half a pawn (with bishop pair so valued) might be provable someday, though even this would surely require at least 50 plies.
Parent - - By duncan (**) Date 2008-07-04 10:20
so if a hypothetical winfinder material only version can get 42 ply in a day on a 6600,  an octal nehalem
should be at least 4 times faster should get 44 ply in a day, to get to 50 ply will take 64 days. (2^6). Not impossible especially if you can do it in stages by saving the results of previous searches.

You say you would find it more interesting if the material-only search valued the bishop-pair as half a pawn, do you have any interest in designing such a version of winfinder. Is it just a matter of dumping eval code of winfinder, and tweaking material points or is it more complicated than that ?
Parent - By lkaufman (*****) Date 2008-07-04 17:58
Only Vas could answer for sure, but I guess it would be easy to make such a version, but time-consuming to optimize it. I could simply set all the values other than material to zero and the version would exist, but it would not search nearly as deeply as possible. 
Parent - - By duncan (**) Date 2008-07-06 01:16
I understand you to say even if it turn out there was a win for white, you doubt that the first pawn win would be in first 25 ? moves.

Is that because the first 25 moves have been well explored?

If you would want to create a computer generated opening book, without any human input, which would be of a superior standard to that which exists today, how many ply would rybka have to be set to. ? would 40 be enough ?

lets say you set a contest for rybka, You have say 6 games. If the human is not mated by 45 moves, then it is a win for human, (even if queen down), for rybka to gain a point it has to mate by move 44.

what level would the human have to be for you to be confident of winning ?

would you have to change rybka' settings for such a match to allow for changed human tactics in dragging out to move 45.?
Parent - - By lkaufman (*****) Date 2008-07-06 04:09
Yes, the openings have been researched well enough that a forced win of a pawn in the first 25 moves is unlikely. Ironically, the most popular defense to 1e4 in top play is to aim for the Marshall Gambit, in which White does win a pawn in ten moves, but Black usually gets enough compensation to hold the draw. Top players seem to prefer this to having a slightly inferior position with equal material. To create a superior book, 40 ply Rybka from the start might be enough, but of course that means with best evaluation, not just material. It would certainly not be superior in many openings, but perhaps overall it might be. To last 45 moves moves without being mated, you don't need to be a strong player, you just need to study what style and openings make for a long game and one which must be won in the endgame. It makes no sense as a handicap because Rybka doesn't understand the conditions; a radically different program would be needed to perform well at this handicap.
Parent - - By duncan (**) Date 2008-07-06 14:08
how do you combine winfinder with rybka,

what happens if rybka says best move is move x which will give you .5 pawn gain in 10 moves, winfinder says move y will give a  1 pawn gain in 11 moves.  but you are worried that maybe the positioin is so poor that you will lose more at move 12.

when do decide to go with rybka'a advice and when with winfinder ?
Parent - - By lkaufman (*****) Date 2008-07-06 17:25
You have hit on the reason that it is not a good idea to choose the best move by running both programs, because there is no good way to decide in a case like this. You can make up some rule such as the difference between the scores has to be a certain amount, but it's rather arbitrary. Certainly on one computer it is better to just let normal Rybka run on her own; if you have two computers there might be some basis for doing what you say.
Parent - - By duncan (**) Date 2008-07-06 17:46
I can understand why not to play both on one machine, as rybka does the best moves, so why weaken rybka with an 'inferior' product.

if rybka says best move is move x which will give you .5 pawn gain in 10 moves, winfinder says move y will give a  1 pawn gain in 11 moves and sticks with it for 12 moves, then perhaps you can have more confidence in winfinder, especially if rybka only values move y at zero rather than negative.

do you have any advice on how to combine rybka or winfinder on separate machines. has any testing been done on that sort of thing. ?
Parent - - By lkaufman (*****) Date 2008-07-06 18:36
No testing has been done on this, but my best suggestion is to go with Winfinder move if score exceeds normal version move by more than X amount (maybe one pawn, this would require experimenting to determine the best value).
Parent - - By duncan (**) Date 2008-07-06 19:00
would you say similarly

abandon rybka move if winfinder scores less than normal version move by  more than X amount 
Parent - - By lkaufman (*****) Date 2008-07-06 19:11
I don't understand. If winfinder score is lower than normal version, of course you keep the normal version move. Your question makes no sense, please explain.
Parent - - By duncan (**) Date 2008-07-06 19:44
rybka advises move x as the best and says after 10 moves it will gain you .5 point.

winfinder says move x will lose you a point at move 11. I assume you should abandon move x and do another move even though rybka considers it inferior.
Parent - By lkaufman (*****) Date 2008-07-06 20:16
Normally you only know the best move found. Setting programs to show 2 or more moves weakens them. But if you see the same move on both screens, and the score does drop by a pawn (if that is our threshold), and then another move improves the score by at least a pawn, then it might indeed make sense to go with the deeper searching winfinder. 
Parent - By Uri Blass (*****) Date 2008-07-06 11:10
33 plies of full width search from the opening is practically impossible.
I guess Mr williams used null move pruning and did not use full width search.

Parent - By Uri Blass (*****) Date 2008-07-03 12:55
I think that depth here is clearly misleading because of pruning.

Chess programs are  going to do more pruning if you only change the evaluation to only material and do not change the search.

I suspect that you may get something like the following:

10 plies of normal search(that has null move pruning) may be equivalent to 9 plies of brute force search if you use evaluation
that include positional factors.
10 plies of normal search may be equivalent only to 8 plies of brute force search if you use only material evaluation.

Parent - - By Vinvin (***) Date 2008-06-30 10:17
Here's the position and the pgn to get there :

rnbqkbnr/ppppp1pp/8/8/3PP3/8/PPP2PPP/RNBQKBNR w KQkq - 0 6

[Event "?'"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "?"]
[Black "?"]
[Result "*"]

1. e4 f6 2. d4 f5 3. Bg5 f4 4. Bxf4 Nf6 5.
Bc1 Ng8 *
Parent - - By George Tsavdaris (****) Date 2008-06-30 20:10
I'm not sure why you wasted one move :-)

[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "?"]
[Black "?"]
[Result "*"]
1. e4 f5 2. d4 f4 3. Bxf4 Nf6 4. Bc1 Ng8 *
Parent - - By turbojuice1122 (Gold) Date 2008-06-30 20:15
Because by the rules, pawns that don't exist are only allowed to move one space ahead on all of their moves :-D
Parent - - By George Tsavdaris (****) Date 2008-06-30 20:37
Nice rule. :-P

I'm not thrilled about that match, since this deviates from Chess a lot. Having a Knight less is a simple handicap. Having a Pawn less is again a simple handicap rule.
But defining the handicap as "white should move twice and then one more time, and black will miss the f-Pawn and white can't play 1.e3, Bd3, Qh5+" is a bit too much of confusion.

I would like to see simple Chess and not Chess variants as like this one. Yes i know the usual, that Rybka will win damn easy such games, but i would like to see it happen and not just hear it will happen. After i see Rybka win 2-3 such matches again some GM's or IM's or even FM's, then yes i would like to go to handicap games.....

BTW i won my first match against Bright 0.3a in this type of handicap. :-) 4'+2" was the time control.
Black has indeed serious problems. My 2 Pawns dominate the center and the lack of f-Pawn is serious enough for the Bg5 pin.

[Event "Blitz:4'+2""]
[Site "?"]
[Date "2008.06.30"]
[Round "?"]
[White "George"]
[Black "bright-0.3a"]
[Result "1-0"]
[ECO "A82"]
[Annotator "User,user"]
[PlyCount "118"]
[TimeControl "240+2"]

{256MB, No.ctg, USER} 1. e4 f5 2. d4 f4 3. Bxf4 Nf6 4. Bc1 Ng8 5. Bg5 {0}
d6 {2.38/13 13} 6. Nc3 {10} Nc6 {2.30/12 10} 7. Nf3 {7} h6 {2.27/12 5} 8. Bh4 {
3} g5 {2.11/13 10} 9. d5 {3} Bg7 {2.14/13 9} 10. Bg3 {4} Bxc3+ {2.22/12 8} 11.
bxc3 {1} Na5 {2.25/12 6} 12. Bd3 {11} Nf6 {1.96/11 7} 13. O-O {3} O-O {
1.91/10 6} 14. Nd4 {8} Kg7 {2.23/11 9} 15. f4 {2} g4 {2.20/11 8} 16. Qe2 {6} c5
{2.90/10 6} 17. dxc6 {8} Nh5 {3.42/11 7} 18. cxb7 {18} Nxb7 {3.50/13 5} 19. f5
{1} Nxg3 {3.49/12 5} 20. Qxg4+ {5} Kh8 {3.57/13 8} 21. Rf3 {29} Nxf5 {3.10/11 5
} 22. exf5 {1} d5 {3.56/11 5} 23. Rh3 {8} Rf6 {3.37/12 4} 24. Qg5 {2} Qf8 {
3.47/13 7} 25. Re1 {6} Nc5 {3.49/12 4} 26. Rxe7 {6} Bd7 {3.54/12 7} 27. Nf3 {14
} Re8 {3.08/11 4} 28. Rxe8 {1} Bxe8 {3.27/13 7} 29. Nd4 {2} Bd7 {3.30/12 7} 30.
g4 {3} Qg7 {3.35/12 7} 31. Qf4 {4} Kh7 {3.50/10 6} 32. Re3 {5} Qf7 {3.63/10 5}
33. Qc7 {5} Nxd3 {3.65/12 4} 34. cxd3 {2} Qg7 {3.93/12 6} 35. h3 {6} a6 {
4.25/11 3} 36. Ne6 {30} Bxe6 {3.82/14 3} 37. Qxg7+ {1} Kxg7 {3.71/12 0} 38.
Rxe6 {2} Rf7 {4.30/14 3} 39. c4 {5} dxc4 {3.95/15 5} 40. dxc4 {1} Rc7 {
4.09/16 6} 41. Re4 {5} Rc5 {3.66/13 4} 42. Kf2 {5} Ra5 {3.83/15 6} 43. Ke3 {2}
Rxa2 {3.83/14 3} 44. Kd4 {11} a5 {3.64/13 4} 45. Re3 {2} a4 {2.94/13 4} 46. Rc3
{3} a3 {2.97/15 5} 47. Kc5 {1} h5 {2.92/15 3} 48. Kb4 {2} hxg4 {3.05/15 4} 49.
hxg4 {1} Rg2 {3.16/16 3} 50. Kxa3 {2} Rxg4 {3.95/19 5} 51. c5 {0} Kf6 {
4.29/19 4} 52. c6 {1} Rg8 {5.64/19 4} 53. c7 {1} Rc8 {5.99/19 5} 54. Ka4 {2}
Kxf5 {5.90/18 5} 55. Kb5 {1} Ke5 {#18/1 0} 56. Kb6 {1} Kd4 {#17/1 0} 57. Kb7 {4
} Re8 {#16/1 0} 58. c8=Q {3} Rxc8 {#16/1 0} 59. Rxc8 {2} Ke3 {#14/1 0} *
Parent - - By lkaufman (*****) Date 2008-06-30 21:32
You seem to forget that Rybka decisively won two normal chess matches against GMs Ehlvest and Benjamin, despite disadvantages that did not change the nature of the game (half time + all Black pieces + 3 move only book in first case, all Black pieces + draw odds in second). I agree that this particular handicap is not very aesthetically pleasing, but it's hard to find a better option in between pawn odds and knight odds. Rook for knight odds is an option, but it may not be more than f7 handicap. The main alternative is probably 2 pawns handicap, but then we get into the issue of which pawns.
Parent - - By JohnL (***) Date 2008-06-30 21:59
I think R vs N odds is particularly interesting. Wouldn't this give Rybka much better chances than just 2 pawns or similar because the knight is stronger than the rook in the opening?
Parent - - By lkaufman (*****) Date 2008-06-30 23:00
Yes, it would. I am reluctant to use it in these matches precisely because I fear that people will say something like "It's not a real handicap, because the knight is stronger than the rook in the opening". I think it is quite definitely a handicap, but not a big enough one for the present match; maybe more suitable for a GM. On the other hand, computers seem to like rooks over minor pieces more than do humans, and I'm also a bit afraid that we might just see a repeat of the last Meyer match, where the human just heads for the endgame and Rybka does not see any way to avoid this. Still, I would like to try it sometime, if only to find out the truth about this handicap.
Parent - - By JohnL (***) Date 2008-07-02 20:42
I understand that you need to adapt to the opinion of the audience.

But I would guess Rybka beats GM:s in rapid with R vs N, if you just tuned it more to avoid simplifications.
To win you really would have to battle yourself all the way through to the endgame.
Parent - By lkaufman (*****) Date 2008-07-02 22:30
Actually Rybka already understands pretty well not to trade pieces when down the Exchange, because rooks shine in the endgame. The problem is that the human's strategy is very simple, just trade pieces. But I would like to see one such match. I think we tried one such informal game at a fast time limit with Benjamin, which I believe was drawn.
Parent - - By Roland Rösler (****) Date 2008-07-01 01:12 Edited 2008-07-01 01:34
I think, the matches "draw and white odds" aren´t played out. A match GM vs Rybka 3 sp (with 8-10 move opening book and 5 men tbs) would be fun. Time control for GM 90'+30" and Rybka mp on your 8-core 15'+5" (similar to Rybka sp 90'+30"). GM sees the moves, while Rybka is in opening book and when we have 5 pieces on the board, result of tbs are taken (similar to the Kramnik vs Fritz match). Two games per day, two hours rest between the games; a 60 move game lasts at most 2.3 hours.
Rybka 3 sp is 500 Elo better than the GM (~2550). The result of the match must be good for Rybka 3, but let us see ...

PS: I would like to see Roman in this odds match. I´m not so interested in the Nakamura match, because here the result is clear for me!
PS2: In the above poll I think Meyer is overrated because of his last success. I voted for 3.5.-0.5 for Rybka, although I´ll keep my fingers crossed for Meyer.
Parent - By lkaufman (*****) Date 2008-07-01 04:32
Giving all those advantages to Roman (or Benjamin or Ehlvest or Nakamura) might be too much. We were able to decisively defeat Benjamin at draw and white odds partly by choosing a not so standard defense (Philidor) that left him on his own fairly early. If he also had access to an opening book during the game, especially our book, he could probably have steered the game into a line where White plays for "two results" (draw or win) as they say. I'm confident that Rybka 3 could give White odds and six to one time odds and short opening book odds and access to opening book by the human GM odds and equal tablebase access odds to a strong GM and win a match, but adding draw odds to all that would probably be too much. At the very least, trying to make an opening book suitable for such a match would be quite a project.
Parent - - By Uri Blass (*****) Date 2008-07-02 15:35
Another option is pawn+castling handicap that means rybka without a pawn and without the right to castle.

Parent - By lkaufman (*****) Date 2008-07-02 16:00
Yes, that is also reasonable. I suspect that Rybka would not play that handicap quite as well as more traditional handicaps, because her play is optimized based on normal openings where both sides usually castle.
Up Topic Rybka Support & Discussion / Rybka Discussion / New Match with FM John Meyer
1 2 Previous Next  

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill