Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Ehlvest vs. Rybka book odds match predictions
Poll Ehlvest vs. Rybka "book odds" score (Closed)
Ehlvest wins the match 4-2 or more 3 5%
Ehlvest wins the match 3 1/2-2 1/2 4 6%
Tie Match 3 5%
Rybka wins the match 3 1/2-2 1/2 4 6%
Rybka wins the match 4-2 25 38%
Rybka wins the match 4 1/2-1 1/2 15 23%
Rybka wins the match 5-1 6 9%
Rybka wins the match 5 1/2 - 1/2 2 3%
Rybka wins the match 6-0 3 5%
- - By lkaufman (*****) Date 2007-07-01 03:56 Edited 2007-07-03 04:12
The "book odds" match against GM Ehlvest (FIDE 2629) is now just four days away. I'd like to invite readers of this forum to make predictions of the outcome. As a reminder, it's six games, 90'+30" increment for Ehlvest, half time for Rybka, all White pieces for Ehlvest, no opening book past move three for Rybka, no endgame tablebases, limit of 512 MB for hash tables. Also, Rybka plays blindfold standing on her head with both arms tied behind her back (if this made any sense I would have thrown this in too)! All opinions welcome! 
Parent - - By Michael Hart (***) Date 2007-07-01 04:19
On a quadcore,  with half time and a handicapped book, and always playing black minus tablebases, I'd put Rybka's Elo somewhere about 300 above Ehlvests' FIDE 2643, give or take.  With plenty or preparation, I'd be willing to bet Ehlvest could knock this down to a 270 handicap.  From this I would expect Rybka to score around 83% or in other words 5 points to Ehlvests' 1 point.


Parent - By lkaufman (*****) Date 2007-07-01 14:09
     Your analysis looks about right mathematically, and indeed Rybka defeated Deep Fritz 10 by 17 1/2 to 12 1/2 in a test I ran last night simulating the match conditions (but 10' chess). Actually the handicap was even more, because instead of 2-1 time odds I ran equal time but with Rybka in SP mode. Rybka played all Black with the three move book (which I'm still working on), Deep Fritz used its own book.
     However Deep Fritz doesn't know about the three move book, and Ehlvest does. This is the big unknown factor in the match, and the main reason that I chose to get involved with organizing it. Also there is the factor that Rybka is generally happy to draw from poor positions, and she will almost certainly get six poor positions in this match.
     So while I would be very pleased with a 5-1 victory, I would not predict or expect it myself.
Parent - - By Hetman (*****) Date 2007-07-01 07:11
I would like to know if the permanent brain will be switched on, before I vote for the option. I think Rybka will win but it is important factor.
Parent - - By lkaufman (*****) Date 2007-07-01 13:59
Permanent brain will be on.
Parent - - By Hetman (*****) Date 2007-07-01 20:54
so I bet on 4:2 in favour of Rybka.
Parent - By Fulcrum2000 (****) Date 2007-07-02 18:44
Me too, 4 draws + 2 wins for Rybka.
Parent - - By diskamyl (**) Date 2007-07-01 07:30 Edited 2007-07-01 07:45

rybka2.3.2a cannot evaluate simple k+p endings correctly without tablebases. what if ehlvest plays to take advantage of this factor? (and I assume it's easy for a GM to exploit such obvious and important weaknesses).
Parent - By grolich (***) Date 2007-07-01 09:47
It's not that easy to force Rybka into an ending she does not understand.
It already performs much better in pawn endgames than previous versions. Plenty of mistakes, true, but you can never be sure if the specific ending you are aiming for will work or not. It's only a general weakness. Not a specific one. If it were possible to know in advance that Rybka will not handle a specific endgame correctly, a GM may be able to prepare for positions in which this endgame is a fallback plan. But even that is very hard.

Even if Ehlvest does manage to prepare that well, he still has survive the middlegame with a reasonable position in order to have any say in the matter.

It would be great if Ehlvest could show spectacular preparation that would knock Rybka out, but the odds are heavily against him.

I'm more concerned about the effects of the lack of a good opening book,
because the results of tests against other programs do not carry a lot of weight here,
as a human GM can prepare specifically for rybka's lousy opening play, and narrow the gap considerably.
The question is, by how much? is it enough?
Parent - - By Uri Blass (*****) Date 2007-07-01 12:56
I voted for 3-3.

My guess is that playing for a draw is easier than playing for a win and
I will not be surprised to see 6 draws.

I think that rating is clearly meaningless here and it is better to play in the right style.

I remember that Arnold hasidovsky (with fide rating near 2200) drew against 3 computers in the israeli league when GM's lost against computers at the same time so my opinion is that you can get no conclusion about result against computers based on rating.

Parent - - By lkaufman (*****) Date 2007-07-01 14:19
     You are certainly right that playing for a draw against a computer is much easier than playing for a win. I recall a match between GM Hubner and an engine (I forget which one) a few years ago which did indeed end with a score of all six games drawn.
     My guess is that Ehlvest will not exactly play for a draw, but will play for good positions that he would normally win against human opponents but will only be able to draw (if he is very careful) against Rybka. He is a very professional player, and would love to win the big prize, but is objective enough to know that a drawn match and Three Thousand dollars would still be an excellent result, so I expect him to play very carefully.
     I think that the terms of this match favor normal play over anti-computer play. If someone knows how to make draws by some special development scheme that avoids early interaction (for example), he probably doesn't need White and won't benefit from the tiny book (since he's not playing book himself). To exploit the advantages of White and the tiny book, you just need to be a very strong player with a good understanding of openings, which pretty much means a high rating. So while I might agree that a special anti-computer style would perform better in a draw odds match, I'm not inclined to believe that this is so for the terms of this match.
Parent - - By Gaмßito (****) Date 2007-07-01 17:46
That engine that played the match against GM Huebner was Fritz. And yes, both played a six games match and it ended in a draw. GM Robert Huebner played very solid and I think that he would have a win in the match.  
I suspect GM Ehlvest has learned enough of his first match and now he will play very carefully in the second match. In my opinion it seems to be logic that he could end with a best result now.  

Parent - By Roland Rösler (****) Date 2007-07-01 19:21
They have their match in Dortmund 2001 (six years ago). The games you can download here: . You can be absolutely sure, that Rybka (version of the next Ehlvest match) makes a short 6-0 against this old Fritz version (and Hardware).
Parent - - By Gaмßito (****) Date 2007-07-01 18:06
Of course. More rating does not means always a better play against the computer.

There are very strong players, without very high rating, that play much better against the computer than a lot of very strong players with higher ratings. Obviously the strongest player can learn how to play against a computer then and get better results in the future. But for only high ratings, it does not mean that a player without experience against programs could play much better against a computer.

Parent - - By FWCC (***) Date 2007-07-01 20:55
If Rybka won your test match by that score against Fritz then it would seem Ehlvest hardly has a chance.Ehlvest has the human factor(being able to adapt and learn)especially from his last match,so this match could end up being closer than thought,however I feel Rybka will win by 4.5 to 1.5  Will this match be broadcast over Chess assistant or on what other network?I assume 2.3.2a will be used?Thanks and hope we have a good match.
Parent - By lkaufman (*****) Date 2007-07-01 23:04
When we know the broadcast venue it will be posted. The version used will be our latest private version, which should be a few Elo points stronger than 2.3.2a, and may also have some special feature for this match.
Parent - - By Hetman (*****) Date 2007-07-02 19:45
that it is true. The players with positional style of playing can adapt but  I do not know if the 'tacticians' can adapt to (anti)computer play.
IMSO the players from the line Capablanca-Smyslov-Petrosian will be most succesfull in that area.
Parent - By Gaмßito (****) Date 2007-07-03 02:30 Edited 2007-07-03 02:40

Yes, strong positional players with enough experience in avoiding computer tactics can generally have much better results against the machines than aggressive tactical players. Although the most important thing is to know how to play against the machine, among them, trying to avoid all kinds of risks, at all costs. In my opinion, a strong tactical player can also improve against computers if he knows how to adapt its game. But it is not so easy, because it sometimes implies also to change the openings to other calmer ones, and that can go against the tactical player's nature.

The point here, is that it is not good, nor appropriate nor intelligent to enter into the enemy territory. Nowadays, chess programs have incredible tactical skills and are just stronger and superiors. The best human players cannot  compete with that tactical strength and of course it is necessary to always keep this in mind before play against the machines.


Parent - - By Torstein (*) Date 2007-07-03 14:17
My guess is that playing for a draw is easier than playing for a win and
I will not be surprised to see 6 draws.

I agree that we will see many draws, but I belive the "tactical factor" (read human tactical mistakes) will play a part as well. So 4 or 5 draws and 1 to 2 decided games, Score 4-2 or 3.5-2.5 in Rybkas favour.
Parent - By Roland Rösler (****) Date 2007-07-03 14:56
Your consideration is quite good, but I hope (and believe) that Jaan will make one win in a game. This one win is needed, because otherwise Vas and Larry would crack up :-).
Parent - - By Lukas Cimiotti (Bronze) Date 2007-07-01 20:58
This is very difficult to predict. The missing (whole) book should be one problem. Missing Tablebases might be even worse - i just tried KBN-K against Rybka - i had  K. Time control was 60 min for 50 moves. Computer was my (well known) oct. It was a draw. I also tried it against DS9, DF10 and DJ9, but with 5 min time for 50 moves. DF10 got it, DJ9 too (mate at move 49 - that was very close), DS9 claimed a win in 150, but 50 moves rule made it a draw. This is one thing, that gives Ehlvest a chance to draw in such a situation. And he might win some games because of to short opening book. But this is pure speculation. Games will be very interesting :)  I am eager to see them. When will they start?

btw. Vas sent me a mail about your quad
feel free to contact me, if there are any problems left
Parent - - By lkaufman (*****) Date 2007-07-01 23:17
     Schedule has been posted under "match offer"; broadcast venue should appear there once I know it. I must disagree with your statement that "missing tablebases might be even worse" (than missing book). Missing tablebases might affect the result of one game in fifty or a hundred; while the missing book (after move 3) will hurt in every single game! I don't expect Ehlvest to get direct wins from the opening, but I do expect him to get clear advantages in every game. You can play almost anything with White, but not with Black.
     My quad problems were speed-related. It seems totally reliable now at 3.47 Ghz (It was apparently unstable at the 3.6 speed it had been set for). That should be the running speed for the match. Thanks for your help.
Parent - - By Banned for Life (Gold) Date 2007-07-02 00:09
Will you be changing the book during the match? I think to a large degree, the result of the match will depend on how much preparation Ehlvest puts into the opening. If he can predict the position after Rybka's third move, he should be able to work with Rybka (before the match :-)) to find excellent continuations for himself. If you don't change the opening book between games, he may be able to exploit the same opening many times.
Parent - - By lkaufman (*****) Date 2007-07-02 04:00
There is no rule against changing the book. It will already have some variety, but naturally I would try to avoid repeating a bad opening. This might not require "changing" the book, simply changing which lines are designated as tournament lines. I don't think Ehlvest will be able to predict the openings -- certainly not before the first game, but even after the match starts there is enough variety to avoid duplication. Also the engine play itself may be varied by changing parameters if I see a reason to do so. I think Ehlvest will simply rely on his skill and knowledge of chess theory to get favorable openings; probably he will conclude as I did that guessing which openings Rybka will play and how she will follow them up is unlikely to work.
Parent - - By Banned for Life (Gold) Date 2007-07-03 15:02
As white, Ehlvest could probably choose moves to reduce the number of lines you could play in the first three moves without significantly degrading Rybka's position to less than 200. From there, he could count on Rybka playing like Rybka (I know you will tweak Rybka's personality for the match, but I suspect this will change less than 10% of the moves that Rybka would normally make with 2.3.2). If you were up against an engine with a good book extended for these 200 lines, I think this would be a serious problem for you and would expect maybe 2 book losses from Rybka wandering into sharp lines without supporting theory. I agree that it won't be as easy for Ehlvest, or any other GM, to take advantage of this.

Parent - By lkaufman (*****) Date 2007-07-03 15:17
Maybe someone with unlimited free time could attempt something like this, but Ehlvest plays nearly non-stop and could not possibly undertake such a time-consuming project with only limited prospects of being able to win the match that way. I'm glad this is the case, because although I want to see our opponent play his best, I would not like to see any games won because he anticipated the exact moves Rybka would choose -- that would rather defeat the point of the match.
Parent - - By Vasik Rajlich (Silver) Date 2007-07-03 19:04

this is very difficult. I'm sure you've played against a good handful of unattended Rybkas in the freestyle tournaments. Just how well can you predict what is going to happen during your analysis?

Parent - - By Banned for Life (Gold) Date 2007-07-03 20:52
This would be valid in a freestyle competition if you came up against an unassisted Rybka with a 3-move opening book (although the variable amount of ponder time by the centaur would add additional variability).

In the engine room playing from book against a Rybka not in book, there is a fair amount of consistency when playing against a Rybka with the same speed-time product which leads to playing the same 15-20 moves over and over and over again in some lines with only a small amount of branching. There are a lot of these type of lines that lead to book wins, say a Rybka eval of > 1 coming out of book. Of course in the engine room, people will use their books to avoid these lines, but the Rybka in this match can't do that beyond move 3. Of course the problem with this concept is that it would require Ehlvest to memorize a large number of lines (thousands), even if he attempted to play the same opening each time.

My point is really that the argument that a good book is worth only 65 Elo is seriously flawed. This argument would hold only if the book was not specialized for use against Rybka. Because most of the engines in the CB engine room are Rybka's, there is a strong tendency to pick up good anti-Rybka lines. Most of the better books in the CB engine room are specialized for a different speed-time product (because most people are playing 3-0), but this is only a technicality.

Parent - - By lkaufman (*****) Date 2007-07-03 23:42
You are probably right that a special anti-Rybka book would be worth (much) more than 65 Elo, but on top of the problem of generating and memorizing this book (for a human), there is the further problem that it would almost surely call for him to play openings he has no experience with. If Rybka did not follow his analysis (and remember, with program changes this is fairly likely) he would be left on his own in unfamiliar turf. So I think it would be a very dubious strategy for him.
Parent - - By Banned for Life (Gold) Date 2007-07-04 01:59
I'm not sure why you believe it would be difficult to make a book for Ehlvest using his own games. There are 463 readily accessible games on line at:

with Ehlvest playing white. These could be easily turned into a book. Additional black moves would be added at plies 2, 4, and 6 to cover all reasonable moves that you may conjure up. Ehlvest would then have to add in any missing moves from plies 3 and 5. This cycle of adding in white moves, then black moves is repeated until there is a reasonable "spanning" set that should cover any reasonable 3-move opening you come up with in response to an Ehlvest prefered opening.

After this, you just play engine matches with black playing with a settup equivalent to what you will be using against a randomized Rybka simulating Ehlvest. You do this for a whole bunch of games, and then Ehlvest reviews the results and picks out the lines that he likes that leave Rybka in a bad position.

This is totally impractical for Ehlvest given the low stakes match he will play, but I think it would be a sound approach if he was playing this match for the Kramnik stakes (i.e. pocket an additional half million dollars for winning the match).

Parent - By lkaufman (*****) Date 2007-07-04 03:30
You might be right, but anyway I'm sure Ehlvest will be confident that just playing normal chess he will get sizable advantages. Either we play dubious defenses right away that leave White with a healthy advantage, or we play mainlines and let him play a dozen moves of theory without thinking until Rybka plays a move known to be inferior. Either way, Rybka will get bad openings. I hope though, that they won't be too bad!
Parent - - By Lukas Cimiotti (Bronze) Date 2007-07-02 21:24
Today i made four tests to add something to the ongoing speculations. In each test 100 games 1+0 were played. I used RybkaII book and my octs.
1. no tablebases for one side:   +24 =59 -17  +24 Elo   Winner: engine with TBs
2. short book of 3 for one side: +36 =56   -8  +99 Elo   Winner: longer book
3. all games black for one side: +16 =61 -23   -24 Elo   Winner: Black - a surprise, maybe not enough games played ? or RybkaII book is better for black?
4. all 3 handicaps for one side:  +46 =50  -4 +155 Elo
So you are right, book seems to be the greatest problem, especially if you got black. Absence of TBs decided only ~5 games (I had a look at each of the 100 games). But in WCCC krpknp would have helped a lot in the game against Zappa.
Parent - - By lkaufman (*****) Date 2007-07-03 01:44
     I'm surprised that tablebases made a difference as often as one game in twenty -- Vas always says that the benefits are really tiny. The short book being worth a hundred Elo might be true the way you are testing, just cutting off the normal book there, but hopefully a three move book with lines chosen for this purpose (like mine) will not be quite so costly. As for Black coming out ahead, normally in computer chess White is worth nearly as much as in GM chess. I ran a few fixed depth tests to check your theory. With the Deep Fritz 10 book, (Rybka playing both sides), White won by a slightly above normal +37=75-18. With the RybkaII book in non-tournament mode, White won by an even large +91=123-46. But with the RybkaII book in tournament mode, I got a tie score in one hundred games. However there are a lot of "traps" to watch out for in doing such tests; for example if you use the Fritz interface there are settings for book learning that can influence results. It is possible though that Jeroen has done more work for Black than for White; if these results are truly representative it would tell him which color needs more work. But the samples are too small. 
     Finally you left out the 2-1 time handicap, and the reduced hash table size (which is effectively another modest speed handicap). So based on your data and allowing for these, the handicap is roughly a full class. That does seem about right; somehow I can't imagine even Kasparov at his best being favored against Ehlvest if we could simulate this handicap, let's say by forcing him to play a move not seen in GM practice by his third move while playing Black every time and with half time on the clock.
Parent - - By Lukas Cimiotti (Bronze) Date 2007-07-03 08:27
I think the number of games was really small - it was just done to add some stuff to the discussion. But i am now retrying this test with a larger number of games to verify it. On the other hand in 3 games Rybka played worse because of tablebases, like sacrificing a rook to get into a TB win of 26 moves. I think, the influence of tablebases depends somewhat on the hardware, as TB accesses cost time. So if you put them on a slow drive and permit frequent accesses, they can even hurt performance. In this case you can watch cpu usage going down rapidly. I made the test on the computer WCCC run on. It has TBs installed on a 32 GB flash solid state disk with an access time of 200 µs, so it is roughly 50 times faster than a normal 7200 rpm SATA drive. Nalimov usage is set to normal. Book learning is always off, i always use tournament mode. Permanent brain is always off, as this normally leads to random results. I made some tests about the influence of size of hash tables - but couldn´t find any significant difference. Of cause in longer games results might vary. I´ll retest that some time. Vas says in FAQ for Rybka 2.0 doubling of speed gives 70 rating points - so doubling of time should do this, too. I´ll make a test about this, too.
Parent - - By Sesse (****) Date 2007-07-03 10:02
How can you say a computer is playing worse by sacrificing to get into a tablebase win? It sees an obvious way to win, and goes for it. Even if it was possible to win with more material left on the board at the time of mate, I don't really think that's bad...

/* Steinar */
Parent - - By Lukas Cimiotti (Bronze) Date 2007-07-03 10:13
Without that sac it was a win in 7.
Parent - By Sesse (****) Date 2007-07-03 11:26
Still, it was a win, no? Chess, unfortunately, does not have points for elegance or good attempts. (Especially the latter would have influenced my rating, I guess :-) )

/* Steinar */
Parent - - By lkaufman (*****) Date 2007-07-03 14:32
Doubling of time might be worth 70 points with ponder off, but less with ponder on, perhaps 50 or so. Hash tables are supposed to be worth at least 5 Elo per doubling if there's enough time to utilize them -- you wouldn't be able to detect this easily in engine-engine testing. Book testing in tournament mode might run into a problem with a tendency to get repeat openings, especially if one side has no book. Your sample size might be smaller than it seems.
Parent - - By Lukas Cimiotti (Bronze) Date 2007-07-03 20:05
One question is - how much hash is used at all - i don´t know, but in earlier tests i found no difference between 16 MB and 256 MB on a slower computer in 3+0 games. Today i am running 1+0 on the new oct, one side gets 128 MB, the other gets only 1 MB - factor 2^7. After 218 games there is an advantage of 27 Elo points to the engine with 128 MB. So 5 Elo per doubling sounds o.k. assuming hash isn´t full in 1+0. This time i am using a very large book, tournament mode off.
Parent - - By Banned for Life (Gold) Date 2007-07-03 20:54
If you look at the non-master processes in task manager, you will see that Rybka has a minimum hash size. Its been a while since I looked but I seem to recall 8 MB...
Parent - - By Lukas Cimiotti (Bronze) Date 2007-07-03 21:02
what do you mean exactly? I don´t think that Rybka overrides gui settings - except for some special versions. I have got them, since my computer played in WCCC. But i am using the normal Rybka 2.3.2a for test only.
Parent - - By Banned for Life (Gold) Date 2007-07-03 21:30
Looking at UCI options from the command line, Rybka 2.3.1a says hash size settings can be between 2 MB and 4 GB with a default of 32 MB.
Parent - By Lukas Cimiotti (Bronze) Date 2007-07-04 05:07
According to the task manager, you are right. When set to 1 MB it seems to take 2. So the factor is only 2^6 and the Elo gain is 4.5 per doubling - that is very close to the theoretical 5 points stated above.
Parent - By lkaufman (*****) Date 2007-07-03 14:50
I ran a test last night simulating the match conditions a bit more closely than earlier tests. Deep Fritz 10 (all White, 4 proc.) vs. Rybka 2.3.2a in SP mode, 15'+5" increment, DF using its tournament book, Rybka using my three move book, double hash table size for Fritz. Score (on two quads) was +9=12-11 for a narrow Rybka victory. These two programs are rated 170 points apart (average of five ratings at different time controls on CCRL and CEGT). The SP mode overdid the speed handicap (I think it's about 2.7-1 instead of 2, and no ponder) but I didn't include EGTB in the test which offsets this. Also, if I had tested at the realtime limit of 90+30, I believe DF would have done much better, as winning advantages mean more with more time. So considering all of this, I stick to my estimate that between computers this handicap is nearly equal to a class (200 points). 
Parent - - By Marc Lacrosse (**) Date 2007-07-03 09:24
In a (relatively) large experiment of mine played at very fast time control, playing with a good quality generic book was worth 65 points when compared with playing without book  for four first class engines.
In the forecoming match the timing will be much slower and rybka will be in a much better situation than playing without book at all.


My experiment (1900 games) :
Parent - - By lkaufman (*****) Date 2007-07-03 14:27
Yes, 65 points feels about right for no book in blitz. It's about the difference between playing White and playing Black. The much slower time limit will favor the player with the book, in this case Ehlvest, I think, because the advantages one gets from knowing theory are much more likely to matter in slow chess than in blitz. Also, no book should hurt Black more than White. On the other hand, the three move book should counteract this. So perhaps these factors cancel out, and your 65 point figure is a fair estimate for the book part of the handicap. 
Parent - - By FWCC (***) Date 2007-07-04 01:51
Larry,with the match approaching and all of the anticipation,is the GM line getting long now?I know it's a bit early as THIS match has not started yet,but do you have the next opponent lined up for Rybka and about when will the NEXT match be(after the piggybank is re-loaded)?Is Joel Benjamin still a candidate?It seems these matches should attract big sponsors.This is also good for the coming match against Junior,the more publicity the better but everyone knows Rybka just won the championship and when she beats Junior this will be sealed.
Parent - By lkaufman (*****) Date 2007-07-04 03:24 Edited 2007-07-04 03:46
     Joel Benjamin is not just "a candidate"; he is definitely set to play the next match in the first half of August (you probably missed the announcement). We have agreed on a pawn handicap match like the first Ehlvest match, but with a more serious time limit (one hour + 30" increment). Of course, if Ehlvest makes a close match this time and the sponsors want to leave money on the table for another match with no material handicap, Joel would be agreeable as long as he gets a decent guarantee. With an extra pawn, many GMs believe they have good chances, but with even material most don't expect to win a match. After Joel, GM Emanuel Berg of Sweden is next in line, we just have to solve scheduling problems.
Parent - By Graham Banks (*****) Date 2007-07-04 02:56
Rybka 5-1   :)
Up Topic Rybka Support & Discussion / Rybka Discussion / Ehlvest vs. Rybka book odds match predictions

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill