Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Biggest weaknesses / biggest potential to improve rybka
- - By Jitugo (**) Date 2007-03-22 08:53
Hi,
this thread is to collect your impressions and priorities of the areas where rybka should improve in its play. This is NOT a bug thread and should not be about specific positions (exepctions for illustrative purposed are allowed of course) and their tactical solutions. Heres my list - feel free to add anything you feel is missing. While i can backup all points with concrete games and positions - you don't have to, its more about your general impressions.

1) Rook Endgames
Rybka should give a higher priority to rook activity, putting rooks BEHIND pawns, knowing about not to move its pawn to a7 if its rook is - unfortunately, but most likely for rybka's current play - on a8.

2) Pawn endgames - rybka often trades down to pure pawn endings without having any clue about who is better - and therefore looses quite a lot of games which otherwise would have been drawn. Maybe it needs better heuristics or better search extensions, but both are probably required. Does rybka not know about "reserve tempi" when scoring a pawn ending? No reasonable player would trade down to a pawn ending, which he hasn't completely determined.

3) Bishop endgames with the bishop pair
Rybka really likes the bishop pair - maybe a tiny bit to much. It mostly passes good opportunities to transform the advantage of the two bishops into another advantageous minor piece ending. This mostly results in rybka later exchanging bishops when opposite color bishops result. these opposite color bishop situations need to be scored more accurate and as more drawish even if rybka is up 1-2 pawns.

4) Opening play - Rybka has a tendency to put its c-kniohts in front of its c-pawns when out of book. I reckon that most cooked book lines against rybka use this to put rybka out of book and in a worse position. A small - very small - penalty for putting its c-knight in front of its c-pawn would probably cure the problem and increase the likelyhood that rybka finds its way back into book.

5) Specific opening play:
Rybka with white does not understand the positions resulting from the f3 e4 expansion plan in the queens gambit declined: 1. d4 d5 2. c4 e6 3. Nc3 Nf6 4. cxd5 exd5 5. Bg5 Be7 6. e3 O-O 7. Bd3 Nbd7 8.Nge2 Re8 9. O-O c6 10. Qc2 Nf8 11. f3

This might be a difficult position for programs because it will be rather closed for a number of prepareratory moves. But even if you force rybka to put its a-rook on d1, playing Kh1, hiding the g5 bishop via h4, f2,g1 and just stubbornly will undo all your efforts playing, Bf2,g3 Kh1,g1 and never really considering e4.
If you then force it to play e4, it will afterwards (after lets saw dxe, fxe) again much to prematurely play d5 giving black the e5 square for a knight and the game. At least this should be avoided at all costs and i am quite amazed that rybka plays this without worries.

I am just about to give up playing these positions with rybka, as you cannot prepare all lines in a book. This is a pity as they are highly advantageous for white.

thats about it for now- lets see what impressions you other guys got.
Parent - - By grolich (***) Date 2007-03-22 10:45
I would say that king safety and attack is a less well understood area by Rybka.
Although I'm not sure whether this is caused by lack of positional understanding, or perhaps it will be improved tremendously without any specific understanding when Vas gets around to improving Rybka's tactics.

(i.e. I'm not certain if it is more the heuristics' fault, and Rybka does not understand it enough, or something else).

I would further say that weaknesses related to color complexes are even less well understood, but I have found no modern engine that can really understand color complex weaknesses if the tactical opportunities it offers are not immediately within its range.

Activity in rook endgames is definitely a problem I've noticed too.
Parent - - By stephenNJUKI (**) Date 2007-03-22 16:36
I think Rybka needs to increment the value her knights as a match progresses. This is coz tactical play tends to become more important than positional advantage as a match transits from the middle game to end game. Bishops are excellent in gaining positional advantage so if their value to a knight was 4:3 in opening, it could morph to 4:4.5 in endgame.
Parent - - By grolich (***) Date 2007-03-22 20:12
Not really.
Two bishops advantage in an endgame can be in and of itself be a winning advantage in many endgames.
The same thing cannot be said for Knights. Actually, the two bishops are considered an awesome force in the endgame.

As to a single bishop, there are too many endgames where a bishop is better than the knight. no less than endgames where it is the other way around, and many strong players I know claim that in reality, the bishop is better in the majority of endgames
(Not that it helps, as you have to evaluate a single position each time... Statistics won't help too much).

And about the openings, there are quite a few modern opening variations where knights battle bishops successfully, and in some of them
the two knights battle the two bishops successfully (which are pretty difficult for some engines, which makes me think they overvalue their bishops, and this is still the middlegame).

Messing around with either piece values cannot be done on any "general" opinion or idea.
If you have many examples of Rybka overestimate / underestimate a certain piece, Show them all on this forum.
That's one of the things it is for.

Then Vas may be able to draw more specific conclusions about certain types of positions which require evaluation changes.
I'm sure he'll appreciate that.

The other thing is that your observations are simply incorrect from the chess point of view:
1)Tactical play does not become more important than positional advantage as the game moves from the middlegame to the endgame.
2)Bishops have an advantage over knights even in many tactical positions. Actually, knights like either closed positions, which by their very nature are less tactical, or open positions where they have outposts / center / other compensating factors. In other words, positional advantages.

But this too is an oversimplification (sometimes closed positions would just give the bishop/s the time to consolidate the position and break it later, which may be very bad for the knights, and sometimes.......). Also the line between positional considerations and tactical ones is a very thin one and easily crossed. So I'll stop here.

As to the value of those pieces, the value ratio bishop/knight is in all probability set to a lot less than 4:3 (This is more a thing to let Vas respond to, but it seems way too excessive. As much as I like my bishops, they don't have a 4:3 advantage in the vs. a knight).

If you have examples of positions where knights were underrated, please, by all means, show them on the forum.
I, for one, would be very interested. And I'm sure I'm not the only one.
Parent - - By stephenNJUKI (**) Date 2007-03-23 04:39 Edited 2007-03-23 05:29
Grolich you are right.. a bishop pair can mate unlike knights! Am a bit surprised though that 1 bishop can be worth more than a single knight. Regarding king safety, my experience is most won games are decided by king attacks so I guess this is something that should also be put into consideration.
Parent - By grolich (***) Date 2007-03-23 16:00
I was also alluding to the fact that even a single bishop in an ending can still be enough to win against a knight.

Larry's reply was a bit more general, that statistically even a single bishop tends to just become better nearer the endgame.
I know of MANY endings where a bishop is a huge positional advantage, but also of many in which the opposite is true. That gave me some things to think about.

The thing is, after reading his (Larry's) post, I went over many GM games (annotated, It'll take me time to analyze so many games myself... too much time) in which a bishop vs. knight endgame arose, and it does seem that in most of them, the bishop is a big positional advantage, which in many cases offer good winning chances.
Most exceptions were when either only 2 or 3 pawns are on the same side of the board for each side (Which makes sense). A few other cases were when the bishop was just horribly bad and blocked.

But in a definite majority (and I don't mean a slight majority...) of them, it indeed seemed that the bishop is better.
I thought the Bishop would only have a small plus.

>my experience is most won games are decided by king attacks so I guess this is something that should also be put into consideration.


Perhaps, but for the evaluation to be any good it has to also take into account the many drawn positions that can arise.
Even an otherwise dead drawn position (to humans...) can be lost by Rybka or any other engine if it can't judge the possible resulting positions correctly. Not sure about the severity of that.

Also the advantage of the 2 bishops in just about any stage of the game does not stem from their ability to mate (knight and bishop can mate too, but they are usually inferior), it's because 2 bishops can simply control the board when used properly. Knights cannot (again, under normal circumstances). That control can either support (or translate to) a mating attack, or just to make the opponent's position passive and keeping him on the defensive, or for just trading to a better position when you need arise
(Think about this: It is easier (partly because of the bishop's range) to trade one of the two bishops for the knight, than it is to trade a knight for one of the bishops

The side with the two bishops has the option to trade (it is easier for him to do), or not, if he doesn't want to.
The side with the knight either has no such option, or that option is harder to achieve.

So they also just give more potential to the side that owns them.
Parent - - By lkaufman (*****) Date 2007-03-22 22:08
     Actually, this and related questions have been my main responsibility since joining the Rybka team. My view, as expressed in "Chess Life" back in 1999 (the article can be found on the net under Dan Heisman), is that the bishop pair is always valuable (nearly half a pawn), but a single bishop and a knight are virtually equal on average, with the balance swinging towards the knight with more pawns on the board and towards the bishop with less. This tends to mean that the (single) bishop is a bit better on average in the endgame, and the knight in the early stages, but whether that is true if we normalize for the number of pawns is debatable. Rybka's eval matches my view in some respects but not all. I plan to research this in the coming months.
Parent - - By grolich (***) Date 2007-03-22 23:03 Edited 2007-03-22 23:05
Interesting... You say that a single bishop tends to be stronger in an endgame against a single knight? hmm....
Interesting. I was of the opinion that it depended on how split (and far apart, but not only) the pawn positions are in the endgame... not necessarily the number of pawns...

Or do you just mean statistically better? If that is the case, then perhaps the reasoning I gave can be a reason as to why it is so:
the more pawns are off the board, the more positions with split pawn structures (and more positions are possible with pawns further apart, which tends to work against the knight) exist.

And from the point of view of the knights, the less pawns are on the board, the harder it is to reach a blocked position, and more positions arise that are open, which the knights do not like that much, which is the reason why they are statistically better with more pawns on the board?

What I mean to ask is whether or not your research is a statistical one, or does it go deeper and try to categorize and divide the problem/s into different types of positions, some of which can go against the "main line" of research? (so as to rate the bishop higher in endgames with only a few pawns ONLY in some of those positions, but leave it unchanged in others of the same type because of other characteristics).
Parent - - By lkaufman (*****) Date 2007-03-22 23:51
     Basically I mean these comments statistically. Of course there are many more factors than the number of pawns that influence knight/bishop value. Naturally factors such as you mention must be considered as well. But other things being equal, I maintain that more pawns favors the knight, probably because knights can jump over them, and bishops cannot. Database statistics clearly bear out this claim. What is not yet clear to me is whether it is just the presence of more pawns that favors the knight, or of more chessmen of all types as well.
Parent - - By Roland Rösler (****) Date 2007-03-23 01:36
I´m really shocked, when I read your answer. Can you explain shortly your real task in the Rybka-team, because I understand nothing yet?
Some other questions:
1. What is the database for your statistics (i.e. only human games or also engine games, quantity, elo-limit)? Decided only the result of the game for your statistics?
2. Give you a clearly defined table to an interface in the Rybka-engine and anything else will work fine?
3. What´s the aim of your work and are you sure that it works?

PS: All what I know is that qk is better than qb.
Parent - By lkaufman (*****) Date 2007-03-23 15:19
     Sorry, but I don't understand some of your questions. The database I worked with was only human games, average rating over 2300 Elo, about 600,000 games I think, only result considered. I use the results of my research only to get ideas for Rybka, we don't assume that they will help without actual test data (and lots of it!). Just for example, my database research proved that the bishop pair was worth close to half a pawn in all phases of the game, and Rybka has always valued the bishop pair very highly (higher than other programs) even before I came on board, and you know where Rybka stands in the ratings.
Parent - - By Jitugo (**) Date 2007-03-23 11:17 Edited 2007-03-23 11:38
Hi,

the following game is a good example of rybka not giving up the bishop pair. What would you play instantly in this position:

8/3k4/1Bnn3p/p5p1/8/PP3B2/7P/3K4 w - - 0 52


obviously taking on c6 would lead to a huge advantage 52. Bxc6+ Kxc6 53. Bxa5 Nb5 54. a4 Nd4 55. b4 Nf3 56. b5+ Kb7 57. Bc3 Nxh2 58. a5 Nf3 59. a6+ Ka7 60. Ke2 g4 61. Ke3 Ng5 62. Bd4+ +-
rybka 2.3.1 in the game played Bg4+ at depth 18 and the game continued for another 132 moves before being a draw. Rybka eventually finds the bxc6 at depth 21. Still it should prefer this move from the start. Is this a pruning or evalutation problem?

EDIT: After 53... Nf5 things are not that simple although rybka sees itself +1.10. But i do not see a winning line when analysing.... maybe rybkas Bg4 keeps more winning chances - interestingly enough in that ending i found a win for rybka 2.3.1 which it missed due to low time. So the issue remains .... complex

1. d4 {B/0 0} Nf6 {B/0 0} 2.
c4 {B/0 0} e6 {B/0 0} 3. Nc3 {B/0 0} Bb4 {B/0 0} 4. Qc2 {B/0 0} O-O {B/0 0} 5.
a3 {B/0 0} Bxc3+ {B/0 0} 6. Qxc3 {B/0 0} b6 {B/0 0} 7. Bg5 {B/0 0} Bb7 {B/0 0}
8. e3 {B/0 0} d6 {B/0 0} 9. f3 {B/0 0} Nbd7 {B/0 0} 10. Bd3 {B/0 0} Rc8 {B/0 0}
11. Ne2 {B/0 0} c5 {B/0 0} 12. Qb3 {B/0 0} d5 {B/0 0} 13. cxd5 {B/0 0} Bxd5 {
B/0 0} 14. Qd1 {B/0 0} h6 {-0.09/15 28} 15. Bh4 {(Lf4) B/0 0} cxd4 {B/0 0} 16.
Nxd4 {B/0 0} Ne5 {B/0 0} 17. Ba6 {B/0 0} Rc5 {B/0 0} 18. O-O {B/0 0} Ng6 {B/0 0
} 19. Be1 {0.09/17 116} e5 {-0.03/15 21} 20. Nb3 {0.16/19 48} Rc7 {0.15/18 0}
21. Bb4 {0.21/19 70} Ne7 {(Re8) 0.19/17 0} 22. Rf2 {0.33/16 56} Qa8 {
(Bxb3) 0.18/16 24} 23. Rd2 {(e4) 0.30/16 29} Bxb3 {(e4) 0.27/15 24} 24. Qxb3 {
0.27/14 0} e4 {(Rd8) 0.28/16 21} 25. Be2 {(Ld6) 0.35/16 41} Rd8 {0.27/16 39}
26. Rxd8+ {0.35/17 8} Qxd8 {0.31/17 0} 27. Rd1 {0.37/19 66} Rd7 {0.27/18 0} 28.
fxe4 {0.37/18 38} Rxd1+ {0.29/18 0} 29. Qxd1 {0.40/20 24} Qxd1+ {0.31/19 13}
30. Bxd1 {0.34/21 10} Nc6 {0.29/20 46} 31. Bd6 {0.48/22 17} Nxe4 {0.39/20 0}
32. Bc7 {0.50/20 24} f5 {(Nd2) 0.33/19 126} 33. g4 {0.53/18 18} Kf7 {0.27/19 0}
34. gxf5 {0.42/18 13} Kf6 {0.43/18 0} 35. Bg4 {0.48/20 31} Ne7 {(Na5) 0.37/19 0
} 36. Bd8 {0.46/17 7} Nd6 {0.38/18 7} 37. Kf2 {0.43/19 14} Kf7 {0.39/18 5} 38.
Bc7 {0.43/17 10} Nexf5 {0.35/17 0} 39. Bb8 {0.48/18 17} a6 {0.38/18 0} 40. Kf3
{0.46/19 15} g5 {0.38/18 11} 41. e4 {0.53/20 30} Nh4+ {0.41/20 0} 42. Ke2 {
0.55/20 18} Nxe4 {0.40/20 19} 43. Bc8 {0.53/20 9} Ke7 {(a5) 0.44/17 17} 44. Ba7
{(Lxa6) 0.54/19 18} a5 {0.40/18 14} 45. Bxb6 {0.60/20 1} Nd6 {0.48/19 17} 46.
Bg4 {0.60/20 0} Ng6 {0.47/18 31} 47. b3 {0.60/21 7} Nf4+ {0.44/17 4} 48. Ke1 {
0.59/21 17} Nd3+ {0.43/18 9} 49. Kd1 {0.58/19 7} Ne5 {0.41/18 18} 50. Bh5 {
(Le2) 0.50/20 22} Nc6 {(Nb7) 0.42/18 14} 51. Bf3 {0.55/18 10} Kd7 {0.41/18 6}
52. Bg4+ {0.60/18 4}  Ke7 {0.41/19 27} 53. Kc2 {0.55/20 4} ... and the game continued for 137 moves
Parent - By Vasik Rajlich (Silver) Date 2007-03-24 12:48
Actually, this is to my eyes a far-from-obvious decision. Maybe Bxc6+ is better, maybe Bg4+ (or some other moves). White stands well, but winning won't be trivial.

Vas
Parent - - By Vasik Rajlich (Silver) Date 2007-03-24 12:45
BTW, here are some of those papers Larry has written about these topics (at least that I know about):

http://mywebpages.comcast.net/danheisman/Articles/evaluation_of_material_imbalance.htm
http://mywebpages.comcast.net/danheisman/Articles/doubled_pawns.htm

Vas
Parent - - By Jitugo (**) Date 2007-03-27 09:21
Hi Vas,

yes i read them, and they are a good overall guideline but i suspect that nowadays a more concrete approach is needed taking into account specific pawn formations, king position etc.... which brings me to my pet idea of codifying the "progress" concept. When looking at endgame positions the material balance and structural components are only broad guidelines, when playing a good player looks if he can make "progress" in a given  positiion (most clearcut in endgame positions). Progress in it simplest form could lead to making the program prefer increasing evaluations when going through the interative deepening. This of course implies to find a away to score piece shuffling as a draw and identifying winning positions where the score reaches a local maxima with reasonable search depth. If i was ever going to write a chess programm, i would work backward from the endgame, trying to improve that first as those concepts seem to give current programs the most trouble (or at least its most obvious to me)  in that stage of the game. Maybe the idea to temporarily and internaly shortening the 50 move rule to lets say a 15 move rule when majority of pawns is locked might help to identify piece shuffling and fortress situations. Maybe other criteria than pawn moves and captures could be added to extend the move count. Also interesting is to construct tablebases on the fly for concrete positions in  endings where it is computationally feasable (many locked pawns  ) not using dtm but rather some other progress criteria.....

well this just some mumblings and broad ideas....

have fun,

Jitugo
Parent - - By Vasik Rajlich (Silver) Date 2007-03-29 12:09
Hi Jutigo,

of course, chess is a very concrete game. That's why we have search. In the evaluation, you just try to codify some general principles and hope they work more often than they fail.

Re. no-progress situations in the endgame, this is a massive research project. You're stepping outside the bounds of minimax (and therefore also outside the bounds of 30+ years of work by several generations of computer chess programmers). It would be interesting to work on this, but please don't hold your breath - time is very limited.

Vas
Parent - By Legendary (***) Date 2007-03-29 15:33
Vas,

It looks like to me you have plenty of helpers.
Parent - - By Torstein (*) Date 2007-03-26 09:54
Maybee the facts speaks for the Bishop in the endgame. Take a look at the following position from the game I found in John Watsons excelent book "Secrets of modern Chess Strategy". Salov - Waitzkin New York 1996: [D]
8/6pp/ppnk1p2/3B4/4PP1P/4K3/PP6/8 w - - 0 37


The game continued:

36... Kd6 37. Bg8 h6 38. h5 Kc5 39. Bf7 a5 40. a3 Ne7
41. Kd3 a4 42. e5 fxe5 43. fxe5 b5 44. Ke4 Kc6 45. Kd4 Nf5+ 46. Kd3 Kc5 47. Ke4
Ne7 48. Be8 Kc4 49. Bf7+ Kc5 50. Be6 Nc6 51. Bg4 Ne7 52. Be2 Kc6 53. Bf1 Kc5
54. Bd3 g6 55. hxg6 Nxg6 56. Kf5 Ne7+ 57. Ke6 Nc8 58. Be2 b4 59. axb4+ Kxb4
[D]
2n5/8/4K2p/4P3/pk6/8/1P2B3/8 w - - 0 60


and white can win outright with 60.Ba6! Nb6 61.Kd6 Ka5 62.Bd3 Nc8+ (?) 63.Kc7 Ne7 64.Be4 h5 65.Kd8 Ng8 66.Ke8 Nh6 67.e6  and you will find that white soon can push his pawn all the way to row 8. Instead play continued:

60.Kd7 Nb6+ 61. Kc6 Nc8 62. Bg4 Ne7+ 63. Kd6 h5 64. Bd7 Ng8 65. Be8 h4 66. Bf7 Nh6
67. Be6 h3 68. Bxh3 Nf7+ 69. Kd5 Nxe5 70. Kxe5 1/2-1/2
Parent - By Vasik Rajlich (Silver) Date 2007-03-26 13:14
Yes, this is a complex topic.

A number of very strong players have claimed that bishops are just better pieces than knights. I've read Watson's book (which is excellent, BTW, I don't read many chess books) and he seems to subscribe to this as well.

Anyway, Rybka doesn't have any strong general preference for a bishop at the moment. Everything will eventually be reviewed.

Vas
Parent - - By Felix Kling (Gold) Date 2007-03-22 23:33 Edited 2007-03-22 23:37
@4): Rybka should understand that you have to open files (soon or later) in the game --> the knight blocks one of the possibilities (e.g. after 1.d4 d5 2.Nc3 you only have e4 (and not c4 anymore) to open a file) --> it's not a good move...
Rybka should also understand then that after 1.e4 e5 the move 2.Nf3 is no mistake, since (an early) f4 would lead to king safety problems.

This would be much better than the solution with the penalty I think.
Parent - - By lkaufman (*****) Date 2007-03-22 23:53
     This is already high on my personal list of priorities of things to improve in Rybka.
Parent - - By stvs (***) Date 2007-03-23 00:59 Edited 2007-03-23 01:19
my opinion first the king safety next endgames.
the good think for this forum is that have many many guys with good knowledge about chess-programing-new ideas.
this forum is not only for fans but a great helper from smart guys-testers to vas. vas you are lucky(or smart) :)
after next rybka release put some gredits for the helps from many here :)
Parent - By FWCC (***) Date 2007-03-23 01:56
Yes,I agree king safety should be a high priority.It seems Fritz loves to mass it's pieces around the enemy king which I have seen in Rybka vs Fritz games.Rybka kind of ignores this until it is too late,she is busy on the Q-side amassing an attack there or pre-occcupied with something else.I can see the attack coming easily with a sac in the air and then it is too late.Of course this does not happen often ,but IT DOES and I am shocked she didn't see it coming when it is relatively easy to see,just sac after the build up when conditions are correct.Also of course this problem does not just occur when playing Fritz only that was just a situation.So a higher value must be placed on king safety as I am shocked when this happens especially from a 3000 player such as Rybka.

FWCC
Parent - By Roland Rösler (****) Date 2007-03-23 03:22
I´m very glad, that you bring the points in descending order of priority. The first two points concern to the endgame and you can be absoultely sure that Rybka fails here (like many other engines) while in the other points it isn´t so clear. But point 1 (move the a pawn from a2-a4-a5-a6 to a7 where the rook is on a8) is a shame for an 3000-elo-engine. But unfortunately this doesn´t cost enough elo-points (because Rybka wins in the middlegame). To point 2 I would say you are right, but often the engines don´t go immediately in the pawn-endgame (just with sacrifice) although it is a clear win (i. e.  
8/pp4pp/4k3/3rPp2/1Pr4P/2B1KPP1/1P6/4R3 b - - 0 30
   Lombardy vs. Fischer, New York, 19.12.1960, round 2). Instead they do nothing. And I don´t recognize it (I´m not Bobby Fischer)!
Now we come to my point, where I also recognize (often) nothing! ZUGZWANG!! And Rybka is no help. She see less. Because of this fault, you can´t say she is bad in endgames. It´s quite worse. You can´t trust her!
Therfore, when you are analysing endgames, you need a second (or an other) engine. That´s annoying.
In my eyes, the best an interresting part of chess is the endgame. And in this part of chess the engines do much worse. It´s like Steve Ovett (British 1500m champ) said: "The decathlon is nine Mickey Mouse events and the 1500 metres". And what he think: they are absolte failures in the most important discipline.
Parent - - By MightyMouse (**) Date 2007-03-23 05:06
I haven't read the other replies in this thread, nor indeed yours! :) I find myself too busy and will try again later, but have an opinion on this topic.

IMHO the biggest weakness is in semi-closed positions where Rybka doesn't consider pawn breaks, resulting in a dull draw. She shuffles her pieces around in meaningless fashion and waits for the 50 move rule to tell her to do something real!? Surely the pawn structure, including pawn breaks, should be a factor in determining the best move in a position.
Parent - By FWCC (***) Date 2007-03-23 23:55
Hello Mighty Mouse,I think in your situation to avoid a closed position or semi-closed position,raising the contempt value may cure this.She will try to avoid draws at all cost the higher the value.I think that is what the anti-human program tries to avoid anyway-closed positions that result in dull draws.
Parent - - By Vasik Rajlich (Silver) Date 2007-03-24 12:38
Endgames - the endgames need some specialized work. We also have to find a way to calibrate winning chances in endgames with winning chances in middlegames. Larry has been working on this second issue, I think 2.3.1 is better than 2.2. there, but the specialized knowledge is still missing.

Bishop pair - needs better heuristics. I don't think that Rybka overvalues the bishop pair in general, but in some positions, she does. (Or rather she overvalues the side with the bishop pair.)

Opening play - we have Jeroen for that :) Seriously, opening theory is really high-level, the issue of Nc3 vs c2-c4 (ie. activity vs space) is highly complex. Some authors apparently tune to opening theory but this seems just the wrong way to do things.

King attacks - practically, this is probably the biggest issue on this list. Dynamic vs structured play has to still be formulated somehow. This issue impacts both evaluation and search. However, style is a big part of this. If you write an engine which will only attack the king, with no other priorities, and this engine plays Rybka, guess how all Rybka losses to this engine will look. Rybka just has a slightly different approach than other engines on this topic.

Vas
Parent - - By Jitugo (**) Date 2007-03-26 07:48
Hi,

i disagree leaving the opening play to the bookmaker - because people will always find ways to throw a program out of book. My intend was not to tune rybka for every opening imaginable, but just to make it avoid blocking its c-pawns by its b-nights early in the game. I think this fundamental positional paradigm holds true for about 90% of openings. And it should be said that most openings where playing the b-night in front of the c-pawn is actually good are either second rate (e.g. chigorin, blackmar diemer) or the move has a tactical justification and would be found by rybka through search. As this "rule" is ónly needed until the b-night acutually moved once - it shouldnt even disturb middlegame play - at least not in a way where it would outweight its advantages.

On king safety,  i am not so sure whether rybka is doing so bad, it might be a matter of our perception: Granted, if rybka looses due to neglecting its own king safety: its ugly! And if you only look at rybkas losses it might really seem to be a big weakness. But i literally hundreds of times saw rybka fending off an good looking attack and winning the game - so i think there are some accurate statistics needed. But if one can improve rybkas attacking and defending skills without harming this strengh i am all for it!

Jitugo
Parent - - By Vasik Rajlich (Silver) Date 2007-03-26 13:18
This is a tricky issue. Some programs just penalize a knight-in-front-of-c-pawn setup. I don't like this, it's not fundamental. Of course, it's just one little rule, but this is how you start making a mess.

Re. king safety, I think there are two issues: true knowledge, and style. Rybka should have more knowledge about king attacks, I look for chances to do that. Stylistically, I think she's ok - I believe in balance.

Vas
Parent - - By lkaufman (*****) Date 2007-03-26 14:48
     My own view is that ANY piece directly in front of ANY pawn is bad, because the pawn cannot move. However if the pawn is part of the king's defense (or the "f" or "g" pawn at the start, since kingside castling occurs far more often than queenside), this principle becomes unimportant. The knight on c3 in front of the pawn on c2 is simply the most common example of this principle. If the pawn that cannot move is more than one rank behind a neighbor, then this becomes serious; hence all the books tell you that knight on c3, pawn on c2, and pawn on d4 is bad. But this would apply just as much to bishop on d3, pawn on d2, pawn on e4, for example.
Parent - By Gaмßito (****) Date 2007-03-26 21:42

I agree.

Central pawns are very important in the opening. These cannot be blocked by other pieces. They should not be blocked.
It is a fundamental principle of the chess game. You cannot see that, in the games of the best positional players of all the times.  It is very difficult to see that in Capablanca great games with 1.d4, or in Karpov games, they simply cannot break this principle. And if they played it at some time, (in simultan games) it was only to investigate or to prove something different.

But with 1.e4, of course, this things change a lot. Nc3 is very common in some openings.

Regars,
Gambitto.
Parent - - By Dragon Mist (****) Date 2007-03-28 12:29
I believe Philidor was the first to point that pawns should not be blocked by light pieces (thus Philidor's defence 1. e4 e5 2. Nf3 d6). However, I also think that modern chess seriously relativised this (and some other) chess postulats. It is much like classic Physics versus Einstein's relativity. This is all so much more emphasized with recent super-strong chess engines; things just aren't the same anymore! We should be open to new ways of playing perfect (or near-to-perfect) chess.

Dragon Mist
Parent - - By lkaufman (*****) Date 2007-03-28 13:56
     All chess principles are just guidelines, which often contradict each other. Even rules like "rooks are better than knights" or "don't bring out your king in the opening" have exceptions. The point is that a rule like "don't block pawns with your pieces" is a valid one that a program should know, which would apply if everything else is equal. For example, if a white knight is on d1, and his pawns are on e2 and c4, white should prefer Nc3 over Ne3 unless he can see far enough ahead to "see" a definite advantage to Ne3.
     The best example of this principle in action is the Breyer, in which Black voluntarily plays 9...Nb8 to unblock the "c" pawn.
Parent - - By Jitugo (**) Date 2007-03-28 15:37
Hi,
this is what i meant. Rybka should prefer not to block its pawn unless search sees a necessity to it. At the moment the evals for the blocking move and non blocking moves differ only by 0.01 or 0.02 - unfortunatly in favour of the blocking move. I would be contend for this to be only corrected in the opening, or hey maybe only until the b-night moved - if it would break something elsewhere.  I can see Vas Point in having rybka finding everything through search and not through rules which sometimes (or often) have sideeffects or actually attract rybka to strongly to certain positions preventing any progress from there. But in this case i can hardly see a drawback - a rule like larrys which penalizes  every piece in front of pawn might even do wonders for rybkas endgame play (e.g. Rooks in front of pawns... ).  I hope Vas comes to his senses and the penalty has to be really small maybe 0.01 or less. 
bye
Jitugo
Parent - - By lkaufman (*****) Date 2007-03-28 18:50
     I can't really speak for Vas, but I believe his objection was only to special case rules like N on c3 in front of pawn on c2, not to a general principle like the one I stated.
Parent - By Vasik Rajlich (Silver) Date 2007-03-29 12:18
Yes, indeed, this formulation sounds a lot better.

We can discuss it further by email. It's similar to the doubled pawn issues - the question is if pawn mobility should be rewarded in general, or only in some special cases.

Re. using general principles in the evaluation - sure, general principles can always be wrong, but there aren't many alternatives. Any general rule which works more often than it fails is worth having. The one thing I might object to is a rule which works only because of the configuration of pieces in the starting position. The risk there is that such a rule could look better than it really is.

Vas
Parent - - By Jitugo (**) Date 2007-03-27 08:38 Edited 2007-03-27 08:47
To put some fuel to the fire: start the game with 1. b3 and let rybka 2.3.1 search it then comes up with 1.... Nc6 and if you then play 2. d4 it likes 2. ... d5 and voila rybka has a suboptimal structure. It is simple to say "the bookmaker should have 1. b3 in his book", but then look at these positions where rybka 2.3.1 with black would  happily play Nc6. Many book-cookers throw rybka off by ludicrous moves like c3 or h3 knowing rybka will then play nc6 giving rybka a bad structure. No bookmaker can account for all possibilities and it really should be the other way around. If rybka leaves book it should still make sensible  moves (not blocking c pawn with b-knight) giving it the opportunity to get back into book and not play anti-positional which gives the book cookers an easy time. Nc6 should be the exception justified through search and not the rule. And it is really a harmless little rule, justified through hundred years of experience and there is no computer analysis which has convinced me of the opposite - and you can turn it of once the b-knight has moved. It shouldnt mess anything up, really:

rn1qkbnr/ppp1pppp/8/3p1b2/8/2P2N1P/PP1PPPP1/RNBQKB1R b KQkq - 0 3


rnbq1rk1/ppp1bppp/5n2/3p4/3PpB2/2P1P2P/PP1NNPP1/R2QKB1R b KQ - 0 7


rnb1k2r/ppp2ppp/3q1n2/3p4/3Pp3/2P1P2P/PP2NPP1/RN1QKB1R b KQkq - 0 7


r1bqkb1r/ppp1pppp/2n2n2/3p4/3P1B2/5N1P/PPP1PPP1/RN1QKB1R b KQkq - 0 4


rnbqkbnr/pppppppp/8/8/8/7P/PPPPPPP1/RNBQKBNR b KQkq - 0 1

r1bqkb1r/1pp2ppp/p1n2n2/3p4/3PpB2/2P1P2P/PP2NPP1/RN1QKB1R w KQkq - 0 7


rn1qkb1r/ppp2ppp/4pn2/3p1b2/3P1B2/2N1PN2/PPP2PPP/R2QKB1R b KQkq - 0 5


rnbqkb1r/ppp1pppp/5n2/3p4/3P4/2N2N2/PPP1PPPP/R1BQKB1R b KQkq d3 0 3
Parent - - By Legendary (***) Date 2007-03-30 15:49
I think Rybka's biggest weaknesses are it's lack of understanding certain King Attacks, Tactics, Piece Assembly, Strategical Plans.
Parent - By Juergen Faas (**) Date 2007-04-01 07:34

>>I think Rybka's biggest weaknesses are it's lack of understanding certain King Attacks, Tactics, Piece Assembly, Strategical Plans. <<


I think I have to disagree here. That´s too general to be helpful (especially "strategial plans"), and "piece assembly" actually is one of her strong points. In general, that is. Of course there are exceptions, like she underestimates the danger for single pieces locked-up and  locked-out (maybe you meant that?).

Also, I think I understand Vas´ comments on king safety now. :) General change of evals (in favor of king safety or attacks) would win some games, but lose more games. Specific knowledge is the key, i.e. the pawn structure f7, g6, h7 is not always problematic, but with no black-squared black bishop, but a safe white bishop on f6 or so, it IS mostly dangerous (I´ve seen this motive several times). This will take time, I guess.

Nice Sunday everyone
Up Topic Rybka Support & Discussion / Rybka Discussion / Biggest weaknesses / biggest potential to improve rybka

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill