Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / What's new in Rybka 3's Evaluation.
1 2 3 Previous Next  
- - By lkaufman (*****) Date 2008-07-20 04:48
     It's time now to talk about the many evaluation improvements in Rybka 3 over Rybka 2.3.2a. There are so many that I can only list the more important ones. A lot of the improvement consists of assigning better values to previously existing terms; I won't go into that. Here is a list of what's new, in five categories:

A) General terms
1. Better definition of mobility
2. New concept of "range"
3. New "contempt" - now when set Rybka tries to avoid various drawish factors, such as neutral trades, blocked pawn structures, and symmetrical pawn structures.
4. New definition of game phase
5. Revised basic piece values
6. Scoring tactical threats in the evaluation (previously handled only by search)
7. New table of how the value of minor pieces depends on which other ones are on the board
8. Safe squares for pieces
9. Relative placement of certain pieces and pawns

B) Opening Play
1. Try to avoid separating neighbor pawns (as in pawns on c2 and d4, knight on c3).
2. Value the right to castle.
3. Connect your rooks.

C) Endgame Play
1. Some rook endgame heuristics
2. Advanced passers are the key to queen endgames.
3. More knowledge about which endgames are drawish and which are not
4. New terms regarding good rook placement

D) Middlegame Play
1. Understanding good and bad bishops
2. Much more advanced understanding of king attacks
3. Importance of defending the king by pieces
4. "Space" (previously addressed only indirectly)

E) Pawn Structure
1. New definitions regarding unopposed pawns
2. Classification of types of backward pawns
3. Pawn Islands
4. Extra penalty for doubled isolated pawns
5. terms relating to how pawns are placed relative to neighboring pawns
Parent - - By Uly (Gold) [mx] Date 2008-07-20 05:50

> 2. Much more advanced understanding of king attacks


Ouch. If after leaving book Rybka becomes some kind of sniper trying to shoot the enemy king, the engine's playing style also looks very promising :)

This post giving clues to engine programmers about what to look for on their programs means that Rybka's most dangerous competition, is Rybka.
Parent - - By lkaufman (*****) Date 2008-07-20 21:34
We were severely deficient in king attack/king safety knowledge before, we're just catching up with everyone else in this area I think.
Parent - - By George Tsavdaris (****) Date 2008-07-20 22:38
By improving King safety i guess it will have a double effect right? I mean it would have a positive effect in both defensive positions(Rybka would defend better if it will be under attack and will see the attack sooner) but also in attacking positions too(that is where Rybka attacks, it will attack better, find the sacrifices sooner). Right?

In how many plies the new Rybka 3 will find 16.Bxh6! here?
2rr2k1/p1qbnpp1/1p2p2p/n1ppP2Q/P2P3N/2PB4/2P2PPP/R1B2RK1 w - - 0 16


Also would Rybka 3 avoid the previous bad moves for black(Rybka 2.3.2a at 40/120' time control) that brought him in that position?
And actually which moves where the bad one's that brought Rybka 2.3.2a in that bad position? How Rybka 3 would fare with black here? She would end up in the same disastrous position?

Also would Rybka 3 play the very good 19.e6!! move that achieves a miniature against Rybka 2.3.2a (a very rare phenomenon!)?

Here is the game:

[Event "CEGT 40/120 (AMD64 x2 4200+)"]
[Site "CEGT"]
[Date "2007.06.25"]
[Round "13.1"]
[White "HIARCS 11 2CPU"]
[Black "Rybka 2.3.2a 64 2CPU"]
[Result "1-0"]
[ECO "C19"]
[Annotator "0.17;0.26"]
[PlyCount "45"]
[EventDate "2007.??.??"]
[TimeControl "40/7200:20/3600:1800"]

{AMD Athlon(tm) 64 X2 Dual Core Processor 4200+ 2211 MHz W=16.0 ply; 593kN/s;
HS-8moves.ctg B=18.2 ply; 186kN/s; HS-8moves.ctg} 1. d4 {0} e6 {0} 2. e4 {0} d5
{0} 3. Nc3 {0} Bb4 {0} 4. e5 {0} Ne7 {0} 5. a3 {0} Bxc3+ {0} 6. bxc3 {0} c5 {0}
7. Nf3 {0} Qc7 {0} 8. a4 {0} b6 {0} 9. Bb5+ {0.17/16 299} Nbc6 {
(Sd7) 0.26/18 264} 10. Ba3 {(0-0) 0.28/16 271} O-O {0.25/18 125} 11. O-O {
0.28/15 170} Na5 {0.25/17 46} 12. Bd3 {0.25/16 272} Rd8 {(Te8) 0.27/17 284} 13.
Nh4 {0.42/15 155} Bd7 {(c4) 0.29/18 439} 14. Qh5 {0.71/16 207} h6 {0.28/17 236}
15. Bc1 {0.69/15 274} Rac8 {(Sf5) 0.31/17 193} 16. Bxh6 {(Dg4) 2.64/15 206}
gxh6 {0.98/17 47} 17. Qxh6 {3.58/15 155} Nf5 {1.03/18 62} 18. Bxf5 {
(Dg5+) 3.75/15 210} exf5 {0.95/16 0} 19. e6 {4.21/15 157} fxe6 {3.79/19 500}
20. Rfe1 {(Tae1) 4.20/14 160} Be8 {3.88/17 129} 21. Re3 {4.79/16 162} f4 {
4.89/17 326} 22. Re5 {6.19/15 165} Rd7 {5.12/17 410} 23. Qxe6+ {6.74/15 168}
1-0
Parent - - By lkaufman (*****) Date 2008-07-21 04:31
Yes, there should be a double benefit to improved king attack/safety knowledge. But don't confuse eval issues with search issues. The eval will help Rybka to have more chances to reach positions where she has some sacrificial breakthru (or to avoid positions where the opponent has them), but finding deep tactics is up to the search. Fortunately that has also been improved greatly. My computers are all tied up with tests until the absolute freeze date (which is imminent); after that I can answer questions about particular positions or games. I Can tell you that GM Perelshteyn showed me a brilliancy game he played that he was very proud of because Rybka 2.3.2a couldn't find the sacrifice even after a long time, but my Rybka 3 found it in 27".
Parent - - By cma6 (****) Date 2008-07-22 00:44
LK, in 27 ply or 27 minutes?
Parent - - By lkaufman (*****) Date 2008-07-22 01:13
27" means 27 seconds.
Parent - By cma6 (****) Date 2008-07-22 03:21
Amazing! Rybka 3.0 is the wonder of the world!
Parent - - By Vinvin (***) [be] Date 2008-07-22 08:53

> I Can tell you that GM Perelshteyn showed me a brilliancy game he played that he was very proud of because Rybka 2.3.2a couldn't find the sacrifice even after a long time


Please, give us this sac position.

Thx
Parent - - By lkaufman (*****) Date 2008-07-22 15:55
    Here is the game. The move in question is 28.Nh4, which only makes sense if you plan the sac on g6 next. So we're talking about finding 28.Nh4.

[Event "World Open"]
[Site "Philadelphia USA"]
[Date "2008.07.06"]
[Round "8"]
[White "Perelshteyn, E."]
[Black "Cheng, Bi"]
[Result "1-0"]
[ECO "D45"]
[WhiteElo "2555"]
[BlackElo "2333"]
[PlyCount "79"]
[EventDate "2008.07.02"]
[EventType "swiss"]
[EventRounds "9"]
[EventCountry "USA"]
[Source "Mark Crowther"]
[SourceDate "2008.07.08"]

1. d4 d5 2. c4 c6 3. Nc3 Nf6 4. e3 e6 5. Nf3 Nbd7 6. Qc2 Bd6 7. g4 dxc4 8. Bxc4
e5 9. g5 Nd5 10. Bd2 O-O 11. Ne4 Bc7 12. O-O-O Re8 13. Rhg1 exd4 14. exd4 N7b6
15. Bd3 Nf4 16. Bxf4 Bxf4+ 17. Kb1 Nd5 18. a3 g6 19. h4 Bf5 20. h5 Re6 21. Nc5
Bxd3 22. Nxd3 b5 23. hxg6 fxg6 24. Rge1 Qd6 25. Nc5 Rxe1 26. Rxe1 Rc8 27. Re6
Qf8 28. Nh4 Re8 29. Nxg6 Qf7 30. Nxf4 Nxf4 31. Rxe8+ Qxe8 32. Ne4 Nd5 33. f3
Kg7 34. Qh2 Qe6 35. Qb8 Qf7 36. Qe5+ Kg6 37. Qd6+ Kf5 38. Ng3+ Kxg5 39. Ne4+
Kh5 40. Qe5+ 1-0
Parent - - By Kappatoo (*****) [de] Date 2008-07-22 19:56
I barely dare to say it, but Nh4 is actually quite obvious from a human perspective (I'm not claiming that it's trivial to prove it's actually winning).
Parent - - By lkaufman (*****) Date 2008-07-22 20:04
It's an obvious candidate move, but it makes no sense unless you plan to sacrifice, and to see that the sacrifice is correct (not necessarily winning, just better than any other move besides Nh4) is extremely difficult.
Parent - By Kappatoo (*****) [de] Date 2008-07-22 20:18
Yes, what I wanted to say is that Nh4 with the idea of sacrificing on g6 is obvious. Even more since a human doesn't want the kingside attack to be stopped, and Nh4 also prevents Qf5, after which the queens' exchange is hardly avoidable.
I don't want to argue how difficult it is to find and actually play Nh4 in a game. But I think one can observe an interesting phenomenon here: Especially in online transmissions of human tournaments kibitzers often say things like 'How can he miss this? My Rybka saw it in 4 seconds.' Or 'Yes, trivial. This is also Rybka's choice at depth 17.' Even when some world class players has just executed a brilliant variation. I often find this development pretty sad.
(Here, we may have another aspect of this tendency: A move which would have maybe gone more or less unnoticed 10 years ago is now considered brilliant because Rybka takes a very long time to find it.)
Parent - - By Felix Kling (Gold) [de] Date 2008-07-22 20:32
The most obvious move is Rxg6, which is just a perpetual, then you see "hey, I can sac my knight" and Ne5/Nh4 are candidates. on Ne5 black can simply exchange the knight, so Nh4... and if Rxg6 is already sound, Nxg6 simply has to be good :)
Parent - - By Roland Rösler (****) [de] Date 2008-07-22 22:52
I´m sure Perelshteyn would have played Rxg6 if Anand was his opponent (or Nh4 with tremulous fingers). It´s always easier to sacrifice something when your opponent has -200 Elo (like in this case) than he has +200 Elo (like in the Anand case).
Parent - - By Felix Kling (Gold) [de] Date 2008-07-22 23:13
well, this position has to be good for white and if you want to play fighting chess and have nothing to lose like against anand, you simply want to play Nh4 :) of course some weak player with low time could play Rxg6,but ok...
Parent - - By Roland Rösler (****) [de] Date 2008-07-22 23:25
and have nothing to lose like against anand

A draw against Anand isn´t so bad. :-) But I see, against Anand you would never get this position.
Parent - By Felix Kling (Gold) [de] Date 2008-07-23 10:03
not sure about that :) maybe not that position, but you can win against any human, even if it's quite unlikely.
Parent - - By Uri Blass (*****) [il] Date 2008-07-22 19:34
It is slightly misleading to say that Rybka2.3.2a could not find it even after a long time

Rybka2.3.2a 32 bits single processor can find it in less than an hour.
Of course Rybka3 is significantly better here but I guess that 10 minutes(instead of 27 seconds) can be enough for rybka2.3.2a in the hardware that you used for Rybka3.

Hardware Rybka2.3.2 sp A3000(2.01 ghz) 512 mbytes hash

2r2qk1/p6p/2p1R1p1/1pNn2P1/3P1b2/P4N2/1PQ2P2/1K6 w - - 0 1


Analysis by Rybka 2.3.2a 32-bit :

28.Nf3-e5 Bf4xe5
  ²  (0.68)   Depth: 5   00:00:00
28.Nf3-e5 Bf4xe5 29.d4xe5 Qf8-f3
  ±  (0.77)   Depth: 6   00:00:00  7kN
28.Nf3-e5 Bf4xe5 29.d4xe5 Qf8-f3 30.Re6-d6
  ±  (0.83)   Depth: 7   00:00:00  12kN
28.Nf3-e5 Bf4xe5 29.d4xe5 Qf8-f3 30.Re6-d6 Qf3-h1+ 31.Kb1-a2
  ±  (0.75)   Depth: 8   00:00:00  20kN
28.Nf3-e5 Bf4xe5 29.d4xe5 Qf8-f3 30.Re6-d6 Qf3-h1+ 31.Kb1-a2 b5-b4
  ±  (0.89)   Depth: 9   00:00:00  41kN
28.Nf3-e5 Bf4xe5 29.d4xe5 b5-b4 30.a3xb4 Qf8-f7 31.Kb1-a1 Rc8-f8 32.b4-b5
  ²  (0.52)   Depth: 10   00:00:01  123kN
28.Nc5-d7 Qf8-f5 29.Qc2xf5 g6xf5 30.Nd7-e5 c6-c5 31.Re6-c6 Rc8xc6 32.Ne5xc6 c5xd4
  ²  (0.68)   Depth: 10   00:00:02  133kN
28.Nc5-d7 Qf8-f5 29.Qc2xf5 g6xf5 30.Nd7-e5 c6-c5 31.Re6-c6 Rc8xc6 32.Ne5xc6 c5xd4 33.Nc6xd4 Kg8-f7
  ±  (0.75)   Depth: 11   00:00:02  152kN
28.Nc5-d7 Qf8-f5 29.Qc2xf5 g6xf5 30.Nd7-e5 c6-c5 31.Re6-d6 Nd5-b6 32.Rd6-f6 Bf4xe5 33.d4xe5 Rc8-f8
  ±  (0.98)   Depth: 12   00:00:03  234kN
28.Nc5-d7 Qf8-f5 29.Qc2xf5 g6xf5 30.Nd7-e5 c6-c5 31.Re6-d6 Nd5-b6 32.Rd6-f6 Bf4xe5 33.d4xe5 Rc8-f8 34.Kb1-c2
  ±  (0.90)   Depth: 13   00:00:05  281kN
28.Nc5-d7 Qf8-f5 29.Qc2xf5 g6xf5 30.Nd7-e5 c6-c5 31.Re6-d6 Nd5-b6 32.Rd6-f6 Bf4xe5 33.d4xe5 Rc8-f8 34.Rf6-d6 Rf8-e8
  ±  (0.98)   Depth: 14   00:00:06  381kN
28.Nc5-d7 Qf8-f5 29.Qc2xf5 g6xf5 30.Nd7-e5 a7-a5 31.Re6xc6 Rc8xc6 32.Ne5xc6 Kg8-g7 33.Nc6xa5 Kg7-g6 34.Na5-c6 Bf4xg5
  ±  (0.98)   Depth: 15   00:00:11  856kN
28.Nc5-d7 Qf8-f5 29.Qc2xf5 g6xf5 30.Nd7-f6+ Kg8-g7 31.Nf6-h5+ Kg7-g8 32.Re6-h6 Rc8-c7 33.b2-b4 Nd5-c3+ 34.Kb1-c2 Nc3-d5
  ±  (1.03)   Depth: 16   00:00:18  1405kN
28.Nc5-d7 Qf8-f5 29.Qc2xf5 g6xf5 30.Nd7-e5 a7-a5 31.Re6-h6 a5-a4 32.Rh6-d6 Kg8-f8 33.Kb1-c2 Kf8-g8 34.Rd6-d7 h7-h5
  ±  (0.80)   Depth: 17   00:00:35  2794kN
28.Nf3-e5 Bf4xe5 29.d4xe5 b5-b4 30.Re6-d6 b4xa3 31.Nc5-d7 a3-a2+ 32.Kb1-a1 Qf8-f5 33.Qc2xf5 g6xf5 34.e5-e6 Kg8-g7
  ±  (0.88)   Depth: 17   00:01:42  6649kN
28.Nf3-e5 Bf4xe5 29.d4xe5 b5-b4 30.Re6-d6 b4xa3 31.Nc5-d7 Qf8-f5 32.Qc2xf5 g6xf5 33.b2xa3 Kg8-f7 34.Nd7-c5 Rc8-e8
  ±  (0.73)   Depth: 18   00:02:33  10400kN
28.Nc5-d7 Qf8-f5 29.Qc2xf5 g6xf5 30.Nd7-e5 a7-a5 31.Re6-h6 a5-a4 32.Rh6-d6 Kg8-f8 33.Kb1-c2 Kf8-g8 34.Rd6-d7 b5-b4
  ±  (0.77)   Depth: 18   00:02:43  11039kN
28.Nc5-d7 Qf8-f5 29.Qc2xf5 g6xf5 30.Nd7-e5 a7-a5 31.Re6-h6 a5-a4 32.Rh6-d6 b5-b4 33.Kb1-c2 Kg8-g7 34.Rd6-d7+ Kg7-g8
  ±  (0.75)   Depth: 19   00:03:25  14495kN
28.Nc5-d7 Qf8-f5 29.Qc2xf5 g6xf5 30.Nd7-e5 a7-a5 31.Re6-d6 a5-a4 32.Rd6-d7 b5-b4 33.a3xb4 Nd5xb4 34.g5-g6 h7-h5
  ±  (0.95)   Depth: 20   00:04:44  20641kN
28.Nc5-d7 Qf8-f5 29.Qc2xf5 g6xf5 30.Nd7-e5 a7-a5 31.Re6-d6 a5-a4 32.Rd6-d7 b5-b4 33.a3xb4 Nd5xb4 34.g5-g6 h7-h5
  ±  (0.95)   Depth: 21   00:05:51  25812kN
28.Nc5-d7 Qf8-f5 29.Qc2xf5 g6xf5 30.Nd7-e5 a7-a5 31.Re6-d6 a5-a4 32.Rd6-d7 b5-b4 33.a3xb4 Nd5xb4 34.g5-g6 h7-h5
  ±  (0.95)   Depth: 22   00:08:04  36092kN
28.Nf3-h4 Qf8-f7 29.Nh4xg6 h7xg6 30.Re6xg6+ Kg8-h8 31.Kb1-a1 Bf4xg5 32.Rg6xg5 Rc8-g8 33.Rg5xg8+ Kh8xg8 34.Nc5-d3 Qf7-g6
  ±  (1.27)   Depth: 22   00:49:33  188390kN

(so k,  22.07.2008)
Parent - - By lkaufman (*****) Date 2008-07-22 19:43
I guess GM Perelshteyn considered an hour to be too long to wait, assuming his machine was like yours.
Parent - - By hal9000 (**) [no] Date 2008-07-22 23:40
I ran this position on Rybka 2.3.2a (64 bit) on my hardware (2 x Intel Xeon X5450).

Nh4 was found in 4:40 mins when using 512MB hash. Still quite slower than Rybka 3 (you use a 3 Ghz octocore, if I'm not mistaken?).

Analysis by Rybka 2.3.2a mp x64:

28.Ne5 Bxe5 29.dxe5 Qf3 30.Rd6 b4 31.axb4 Re8 32.e6 Qh3 33.b5 Qf1+ 34.Qc1 Nc3+
  ²  (0.68)   Depth: 5   00:00:00
  ±  (0.89)   Depth: 17   00:00:08  4514kN
28.Nd7 Qf5 29.Qxf5 gxf5 30.Nde5 Kg7 31.Nxc6 Kf7 32.Rh6 Kg7 33.Nce5 Rc1+ 34.Ka2 Rc2
  ±  (0.93)   Depth: 17   00:00:11  5929kN
  ±  (0.98)   Depth: 22   00:00:58  32351kN
28.Nh4 Qf7 29.Nxg6 hxg6 30.Rxg6+ Kh8 31.Ka1 Bxg5 32.Rxg5 Rg8 33.Rxg8+ Kxg8 34.Nd3 Qg6
  ±  (1.29)   Depth: 22   00:04:40  155mN
Parent - By lkaufman (*****) Date 2008-07-23 03:31
Yes, 3 Ghz octocore.
Parent - By vroger007 (**) [be] Date 2008-08-07 19:35
make that "less than half an hour"  on an old and slow AMD Athlon 64, 1.99 GHz, Win XP 32 bit, single processor, single core,
recently upgraded from 512M to 3GB RAM: the secret (as always) is to set outlook (rybka, not MS) to ultra-optimistic

Analysis by Rybka 2.3.2a 32-bit OUO,TUC,NUF,PA,1CPU:

1.Ne5 Bxe5 2.dxe5 Nb6 3.Ne4 Nc4 4.Rf6 Qe7 5.Nd6
  ±  (0.96)   Depth: 9   00:00:00  0kN
1.Ne5 Bxe5 2.dxe5 Nb6 3.Ne4 Nc4 4.Rf6 Qe7 5.Nd6
  ±  (0.96)   Depth: 10   00:00:00  0kN
1.Ne5 Bxe5 2.dxe5 Nb6 3.Ne4 Nc4 4.Rf6 Qe7 5.Nd6
  ±  (0.87)   Depth: 11   00:00:00  11kN
1.Ne5 Bxe5 2.dxe5 Nb6 3.Ne4 Nc4 4.Rf6 Qe7 5.Nd6 Nxd6 6.exd6 Qd7
  ±  (0.88)   Depth: 12   00:00:02  114kN
1.Nd7 Qf5 2.Qxf5 gxf5 3.Nde5 c5 4.Rd6 Nb6 5.Rf6 cxd4 6.Nd3
  ±  (0.98)   Depth: 12   00:00:03  192kN
1.Nd7 Qf5 2.Qxf5 gxf5 3.Nde5 c5 4.Rd6 Nb6 5.Rf6 Bxe5 6.dxe5 Rf8 7.Re6
  ±  (0.95)   Depth: 13   00:00:05  255kN
1.Nd7 Qf5 2.Qxf5 gxf5 3.Nde5 c5 4.Rd6 Nb6 5.Rf6 Bxe5 6.dxe5 Rf8 7.Re6 Rd8
  ±  (0.98)   Depth: 14   00:00:08  488kN
1.Nd7 Qf5 2.Qxf5 gxf5 3.Nde5 c5 4.Rd6 Nb6 5.Rf6 Bxe5 6.dxe5 Rf8 7.Rd6 b4
  ±  (1.00)   Depth: 15   00:00:14  841kN
1.Nd7 Qf5 2.Qxf5 gxf5 3.Nde5 a6 4.Kc2 Bxe5 5.dxe5 Nf4 6.Rd6 Nh3 7.Rd2 Kf7
  ±  (0.94)   Depth: 16   00:00:28  1759kN
1.Nd7 Qf5 2.Qxf5 gxf5 3.Nde5 a5 4.Rxc6 Rxc6 5.Nxc6 a4 6.Kc2 Kg7 7.Nh4 Bxg5
  ±  (0.77)   Depth: 17   00:01:29  5673kN
1.Nd7 Qf5 2.Qxf5 gxf5 3.Nde5 a5 4.Rd6 a4 5.Rd7 b4 6.axb4 Nxb4 7.Rb7 Nd5
  ±  (0.78)   Depth: 18   00:05:08  18061kN
1.Nd7 Qf5 2.Qxf5 gxf5 3.Nde5 a5 4.Rd6 a4 5.Rd7 b4 6.axb4 Nxb4 7.Rb7 Nd5
  ±  (0.84)   Depth: 19   00:07:41  25174kN
1.Nd7 Qf5 2.Qxf5 gxf5 3.Nde5 a5 4.Rd6 a4 5.Rd7 b4 6.axb4 Nxb4 7.Rb7 Nd5
  ±  (0.96)   Depth: 20   00:10:51  37634kN
1.Nh4 Qf7 2.Nxg6 hxg6 3.Rxg6+ Kh8 4.Ka1 Bxg5 5.Rxg5 Rg8 6.Rxg8+ Kxg8 7.Nd3 Qg6
  ±  (1.25)   Depth: 20   00:29:40 89369kN
1.Nh4 Qf7 2.Nxg6 hxg6 3.Rxg6+ Kh8 4.Ka1 Bxg5 5.Rxg5 Rg8 6.Rxg8+ Kxg8 7.Nd3 Qg6
  ±  (1.23)   Depth: 21   00:39:24  113mN

(Vermeir,  07.08.2008)
Parent - - By BigBen (****) Date 2008-07-21 08:25
Hi, It will be interesting to see how quick R3 finds the Bh6 move ... R2.3.2a rook 2 minutes via the chessbase 9 UCI

2rr2k1/p1qbnpp1/1p2p2p/n1ppP2Q/P2P3N/2PB4/2P2PPP/R1B2RK1 w - - 0 1


Analysis by Rybka 2.3.2a mp :

16.Bxh6 gxh6 17.Qxh6 Nf5 18.Qg5+ Kf8 19.Nf3 Bc6 20.Bxf5 exf5 21.Qh6+ Ke7 22.Rae1 Bd7
+/-  (0.82)    Depth: 18   00:02:01  86446kN
Parent - By miles (**) Date 2008-08-08 13:58 Edited 2008-08-08 14:19
Shredder 8 on 2.7GHz sp using 512 hash ( 218Kn/s )found 16.Bxh6 immediately, then for several seconds examined 16.Re1 and 16. Be3 then returned to 16. Bxh6 and followed Rybka's moves exactly after 3 minutes.
Parent - - By Venator (Silver) [nl] Date 2008-07-21 17:17
This takes 3 seconds on my Q6600, with 1 GB hash. Funnily it is the first move new Rybka likes, then she changes very shortly to 16.Qg4, but after 3 seconds back to 16.Bxh6!

19.e6! takes 2 seconds.
Parent - - By sharpnova (***) [us] Date 2008-07-22 03:08
I don't understand why it would take your Rybka 3 seconds and other people's Rybka minutes. Do you use infinite analysis? How many lines of analysis do you have your infinite analysis showing? And I guess whether you're using 64-bit or not shouldn't really have anything to do with the minutes vs. 3 seconds thing. I also use a 1 gig hash (well 819 MB) and using infinite analysis (with one line showing) it took my Rybka 2.3.2a mp 32-bit on Deep Fritz 8 UI over two minutes to find this.
Parent - - By Uly (Gold) [mx] Date 2008-07-22 03:27
I think he was using Rybka 3 Beta
Parent - By sharpnova (***) [us] Date 2008-07-22 05:54
Awesome
Parent - By Venator (Silver) [nl] Date 2008-07-22 06:34
Yes, it is Rybka 3 beta.
Parent - By BigBen (****) Date 2008-07-22 08:22

>This takes 3 seconds on my Q6600


wow, def impressed by that .. The comp I was on was at 4.2Ghz

Regards
Parent - By hal9000 (**) [no] Date 2008-07-22 23:57
Impressive. On my octal (Xeon X5450) it took 51 seconds:

Analysis by Rybka 2.3.2a mp x64:

1.Qg4 Kh8 2.Re1 Nc4 3.h3 Bc6 4.Nf3 Rb8 5.Ra2 Qd7 6.Bf4 Nf5 7.a5
  ²  (0.28)   Depth: 4   00:00:00
  ²  (0.28)   Depth: 16   00:00:39  25779kN
1.Bxh6 gxh6 2.Qxh6 Nf5 3.Qg5+ Kf8 4.Nf3 Bc6 5.Bxf5 exf5 6.Qh6+ Ke7 7.Ng5 Rg8
  ±  (0.83)   Depth: 16   00:00:51  33511kN
Parent - - By Carl Bicknell (*****) [gb] Date 2008-07-21 12:10
why was king attack/defence a low priority? I would have thought it would have been the no1. selling item.
Or does it make less of a difference in elo than people imagine?
Parent - By lkaufman (*****) Date 2008-07-21 19:13
In general I think that Vas did not believe before last year that work on the eval would pay off nearly as well as work on the search, and king attack/defense is a fairly difficult issue. After I joined the team he came around to viewing search and eval work as more or less equals. Making Rybka find king attack tactics is a search issue; there the problem is that in general an improvement in this area causes a weakening of play in other areas. But it seems that Vas has found a way around this with his new search in Rybka 3.
Parent - - By Hetman (*****) Date 2008-07-20 06:53
It is very intersting.!

Is being A7 described somewhere in the literature or it is development for Rybka ?

Rgds
Hetman
Parent - - By lkaufman (*****) Date 2008-07-20 21:35
It is probably not in other programs, but it was inspired by a comment I read in a chess book, which immediately rang true to me.
Parent - - By Hetman (*****) Date 2008-07-21 18:12
Perhaps you remember the  title ? :-)

rgds
Hetman
Parent - - By lkaufman (*****) Date 2008-07-21 19:20
I don't want to help competitors too much!
Parent - - By Hetman (*****) Date 2008-07-22 19:00
I am not the competitor :-).

I know that Breyer has written sth about in 1920s. I have been curious if that was his book.

Rgds
Hetman
Parent - By lkaufman (*****) Date 2008-07-22 19:32
No, it's a 21st century book.
Parent - - By plicocf (***) [br] Date 2008-07-20 07:11
IM Larry, can you explain to us the diferences between Rybka 3 and
Rybka 3 humanized? Which are the knowledges that humanized have
and Rybka 3 don´t have?

Paulo Soares
Parent - - By lkaufman (*****) Date 2008-07-20 21:38
There will be more about this soon, but the basic difference is that the humanized version has more terms, and higher weights on terms, that are part of chess literature. For example, much stronger incentive to avoid Nc3 with pawns on c2 and d4. In general, more weight on static features vs. dynamic ones.
Parent - - By exigentsky (***) [us] Date 2008-07-22 03:41
How much weaker is the humanized version and will this version be included for those who buy R3? What if we buy from Convekta or ChessBase?
Parent - - By lkaufman (*****) Date 2008-07-22 04:27
It's about 20 Elo worse in hyper-bullet chess, hasn't yet been tested at real speeds. I hope that in serious chess the gap will be less or nonexistent. I am told that it will be included free for anyone who buys Rybka 3 from any source.
Parent - - By exigentsky (***) [us] Date 2008-07-22 06:17 Edited 2008-07-22 06:22
OK, sounds good. I wonder if at time controls of 2 hours for 40 moves the humanized version might be stronger. I'm not even sure what "hyper-bullet" chess is but results might not translate so clearly to real chess. I realize engines are not like people, but this would be an interesting test.
Parent - By lkaufman (*****) Date 2008-07-22 15:46
I mean speeds around game/1 second when I say "hyper-bullet" (my term). It is quite possible that you are correct, but probably the two versions are close enough in strength at 40/2 that it would require something like 1000 games to prove which is better. We just can't tie up our machines for tests at long time limits, but we hope that after the release testers and owners will run tests like this.
Parent - - By Labyrinth (*****) [us] Date 2008-07-20 10:32
5. Revised basic piece values

What are the new ones? What were the old ones?
Parent - By Felix Kling (Gold) [de] Date 2008-07-20 21:37
@1: better
@2: worse
:)
Parent - - By lkaufman (*****) Date 2008-07-20 21:46
Actually it's impossible to tell you exactly what the piece values are (aside from not wanting to help competitors), because the values are intertwined with other terms. It does no good to know that the value of a bishop is 3.5 for example if you don't know how we adjust this for mobility. If mobility is always a positive number the value would have to be low, if it can be positive or negative the value would be higher. I can say that I raised the relative value of minor pieces to pawns, because I think I reduced this ratio too far in the 2.32a version.
Parent - - By Roland Rösler (****) [de] Date 2008-07-20 23:49
Have you any difference in eval of the two white bishops (the black squared and the white squared vice versa with black) at start position? I only ask, because Bobby Fischer favors the white squared bishop with white.
Up Topic Rybka Support & Discussion / Rybka Discussion / What's new in Rybka 3's Evaluation.
1 2 3 Previous Next  

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill