Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Cluster Toga - Rybka
- - By Vasik Rajlich (Silver) [hu] Date 2007-12-30 10:34
I've been analyzing this game. One move 30, the following position arose:

1r3r1k/3qb2p/p2p1pp1/1p2p1Pn/4B2P/P1P1BP2/1P1R1Q2/1K1R4 b - -


This position looks roughly equal after something like 30. .. Qc7. Unfortunately, black can't get in .. a5 due to Bb6 & Bxa5 (prefaced by Bc5 if the queen is on d7 and/or Ba7 if the rook is on b8). 30. .. f5 also isn't very attractive due to 31. Bc2 followed by f3-f4.

At first glance, 30. .. Qh3 would seem to be the source of black's problems. It would seem like a classic case of a position where the score very slowly rises for white as his queenside pawns move forward. A lot could be said about this topic and how to deal with it, ie. how to compare rising scores for one move with stable scores for another move. There is no easy solution.

However, 30. .. Qh3 itself is not that commital (or bad). Black can continue 30. .. Qh3 31. Ba7 Rbc8! 32. Qb6 Qd7! 33. Qxa6 (only try for an advantage) fxg5 34. hxg5 (34. Rxd6 Bxd6 35. Rxd6 Qe7 36. Re6 Qd7 37. Rxe5 should also be ok) Ra8 35. Bb7 Nf4 with equality.

Another playable alternative is 30. .. Qh3 31. Ba7 Rbc8 32. Qb6 fxg5 33. Qb7 (33. hxg5 Qd7 transposes to 32.. Qd7 33. Qxa6 fxg5 34. hxg5) Rce8 34. hxg5 Ng3 35. Bc6 Qc8, since here 36. Qb6 is impossible due to 36. .. Bd8.

Even after 30. .. Qh3 31. Ba7 Rbd8?! 32. Qb6, black can still try 32. .. Qd7, although white now has 33. Bc6 and this is already a little bit unpleasant.

The really critical mistake seems to be 32. .. Qxh4. After this, there is no going back.

3r1r1k/B3b2p/pQ1p1pp1/1p2p1Pn/4B2q/P1P2P2/1P1R4/1K1R4 w - -


I'm curious about two questions:

1) How quickly does the score rise for white with a really deep normal search? After an hour or so on a core duo laptop, Rybka gets to +0.26 (after I force 33. Qxa6).
2) How does the randomizer behave here? Is it depth-dependent or threshold-dependent?

This type of positions seems tailor-made for a randomizer-type approach.

If somebody has a few days of CPU time to throw at these two questions, I'd be quite curious to see the result.

Vas
Parent - - By Felix Kling (Gold) [de] Date 2007-12-30 11:48 Edited 2007-12-30 12:54
i think it's a simple (or maybe not that sim ple for engines) strategical mistake, white has those bishops "looking" at the queenside and supporting the attack on the queenside and much more active pieces- so it's a mistake to open the position and allowing white to move his pawns, while black has difficulties to get his pawns going. Of course the randomizer will understand that, but I guess it should be possible to tell that the "normal" Rybka.
Some ideas what can be formulated as rule:
a) if one side has a passed pawn on the a- or h-file and the bishop pair, give a bonus (I think the bishop pair is quite good in supporting those pawns at the side of the board)
b) if the opponent has more active pieces in a middle/endgame, avoid the creation of passed pawns, exchanging pawns like this doesn't lead to a more drawish position in a number of cases.
c) if the bishops "look" at one side of the board, i.e. have more squares either at the king or queenside, give a bonus to attack there. other way around: if one side plays at e.g. the kingside, it's good to place the bishops where they can look at the kingside.

I'm not sure if those rules are really true, but maybe they are worth a try. I guess you implemented at least c) in some way already.
Parent - - By Vasik Rajlich (Silver) [hu] Date 2007-12-30 12:39
Of course, the ideal is to get this via evaluation. If the evaluation is good enough, you don't really need search.

While Larry is working on that :), I think it also makes sense to look for ways to search more effectively in positions like this.

Re. your heuristics, 'a' is interesting, I wrote it down. Rybka does something like 'b', and in fact she does partially understand this position. 'c' isn't really an evaluation heuristic, if I understand it correctly.

Vas
Parent - By Felix Kling (Gold) [de] Date 2007-12-30 12:52
mmh, c is a bit difficult to formulate, i guess you give a higher bonus to pieces which are on the right side for the attack, I don't know how you do that, but you could do this with the bishops in a similar way...
Parent - - By Nick (****) [gb] Date 2007-12-30 15:55
On question 2 the randomizer question, yes it seems Qh4 was the lemon, after almost 900 7ply games of a normal Rybka vs a randomize=10 Rybka on a Q6600@3.4 we see:
Parent - - By Eelco de Groot (***) Date 2007-12-30 17:26 Edited 2007-12-30 18:19
Thanks for all the interesting posts about this game, I have not really looked closely at the game with my own eyes but my old slow computer has been looking at some of these positions and now I'm not sure if the critical mistake is not actually a bit later stil than 32... Qxh4. Maybe it is that Glaurung 2.0.1 is just not looking deep enough, but here after 22 plies the score for Black is still positive on move 35 so three moves later than Qxh4.

I don't consider Glaurung as giving much weight -yet- to positional characteristics, Tord is I think more relying on deep searches mostly at the moment, - I sure wish I had the parallel computer to do Glaurung's search justice :). As Black is here in this position still two pawns ahead, Glaurung maybe in this case then overestimates the material advantage? Haven't checked other programs yet. It does not consider Rybka's 35... f5 anymore though among the two best moves after a 17 plies search, so does Rybka maybe change its mind about this move on deeper searching?

This is at 22 [Edit: now 23] plies in two best search:

3r1r1k/4b2p/Q1Bp1pp1/1p2p1q1/8/P1P2Pn1/1P1R1B2/1K1R4 b - -


Engine: Glaurung 2.0.1 default (256 MB, Athlon 2009 MHz)
by Tord Romstad

22     104:28 +0.09    35...Rg8 36.Bd5 Rg7 37.Qxb5 Nh5
                       38.Bb6 Rf8 39.Qc6 Nf4 40.Bc7 Nxd5
                       41.Qxd5 h5 42.Qe6 Qg3 43.Bxd6 Bxd6
                       44.Qxd6 Rgf7 45.Qc6 Kg7 46.Rd7 Rxd7
                       47.Rxd7+ Kh6 (4.302.732.055) 686

22     104:28 +0.01    35...Nf5 36.Qxb5 Ng7 37.Bd5 Qf5+
                       38.Qd3 Qxd3+ 39.Rxd3 Nh5 40.Bb6 Rb8
                       41.Ba7 Rbe8 42.Be3 Kg7 43.b4 Ng3
                       44.Kc2 Nf5 45.c4 Nxe3+ 46.Rxe3 h5
                       47.a4 Rb8 48.Kb3 h4 (4.302.732.055) 686
__________________________________________________

23     201:04 +0.07    35...Rg8 36.Bd5 Rg7 37.Qxb5 Nh5
                       38.Bb6 Rf8 39.Qc6 Ng3 40.b3 Nf5
                       41.Re2 h5 42.Kb2 h4 43.c4 h3 44.Qd7 Qh5
                       45.Rdd2 Kh7 46.Qb7 Re8 (8.267.409.581) 685

23     201:04 0.00     35...Nf5 36.Qxb5 Ng7 37.Qb7 Nf5
                       38.Qb5 Ng7 (8.267.409.581) 685

A few moves later in the game Glaurung and ahem, Belka, saw a clear advantage for White so judging from that could 35... f5 have been the losing move? I don't really know whether Monte Carlo searching can answer this? 

After 37. Bc7

1r3r1k/2B1b2p/Q1Bp2p1/1p2ppq1/8/P1P2Pn1/1P1R4/1K1R4 b - -


Engine: Glaurung 2.0.1 default (64 MB)
by Tord Romstad

21.01  50:02  -0.72    37...f4 38.Bxb8 Rxb8 39.Rxd6 Bxd6
                       40.Rxd6 Nf5 41.Rd7 Rd8 42.Rxd8+ Qxd8
                       43.Qxb5 Qd1+ 44.Ka2 Qd6 45.Qb7 h5
                       46.Be4 Qf6 47.b4 Nd6 48.Qa8+ Kg7
                       49.a4 g5 50.Qa7+ Qf7+ (1.096.450.099) 365

22.01  165:49 -0.90    37...f4 38.Rxd6 Bxd6 39.Rxd6 Rbc8
                       40.Rd7 Qf5+ 41.Ka1 Nh5 42.Re7 Rxc7
                       43.Rxc7 Nf6 44.Bxb5 Nd5 45.Rc5 Ne3
                       46.Qd6 Nc2+ 47.Ka2 Ne1 48.Be2 Re8
                       49.Rc7 (3.767.375.428) 378

23.01  290:34 -0.86    37...f4 38.Rxd6 Bxd6 39.Rxd6 Rbc8
                       40.Rd7 Qf5+ 41.Ka1 Rxc7 42.Rxc7 Rf7
                       43.Rxf7 Qxf7 44.Qxb5 Qe7 45.b4 Kg7
                       46.Kb2 h5 47.Qd3 Kf6 48.a4 g5 49.a5 Nf5 (6.617.253.884) 379

24.01  594:54 -0.92    37...f4 38.Rxd6 Bxd6 39.Rxd6 Rbc8
                       40.Rd7 Qf5+ 41.Ka1 Rxc7 42.Rxc7 Rf7
                       43.Rxf7 Qxf7 44.Qxb5 Qe7 45.b4 Kg7
                       46.Kb2 g5 47.a4 e4 48.fxe4 Kf6
                       49.Qc4 h5 50.a5 (13.644.295.393) 382

best move: f5-f4 time: 716:43.640 min  n/s: 381.218  nodes: 16.393.770.000

Engine: Belka 1.8.11 (64 MB)
by Yuri Osipov, Igor Korshunov

17.35  4:33   -0.99    37...Rbc8 38.Bxd6 Bxd6 39.Rxd6 Rb8
                       40.Qa7 Qh4 41.Qe3 Qf4 42.Qc5 Kg7
                       43.Re6 e4 44.Rd7+ Kh6 45.Qe7 Rh8
                       46.fxe4 Qf1+ 47.Kc2 Qe2+ 48.Rd2 (140.639.102) 514

18.35  26:53  -1.00    37...Qe3 38.Bxb8 Rxb8 39.Bxb5 e4
                       40.a4 exf3 41.Rd3 Qe4 42.Qa7 Rb7
                       43.Qf2 f4 44.Qxf3 d5 45.Qxe4 dxe4
                       46.Rd7 Rxd7 47.Rxd7 Bc5 (827.672.698) 512

19.35  46:44  -1.08    37...Qe3 38.Rd3 Qf2 39.Bxb8 Rxb8
                       40.Bxb5 Qc5 41.Rd5 Qc7 42.Qc6 Qxc6
                       43.Bxc6 Kg7 44.a4 Rc8 45.Bb7 Rc7
                       46.Rb5 h5 47.a5 h4 (1.419.094.595) 506

20.35  96:42  -1.31    37...Qe3 38.Rd3 Qf2 39.Bxb8 Rxb8
                       40.Bxb5 Qc5 41.Rd5 Qc7 42.a4 e4
                       43.Qc6 Qxc6 44.Bxc6 Rc8 45.Bd7 Rc7
                       46.Bb5 exf3 47.R5d3 Kg7 48.Rxf3 (2.912.214.053) 501

21.35  370:13 -1.29    37...Qe3 38.Rd3 Qf2 39.Bxb8 Rxb8
                       40.Bxb5 Qc5 41.Rd5 Qc7 42.a4 e4
                       43.Qc6 Qxc6 44.Bxc6 Rc8 45.Bb5 exf3
                       46.R5d3 Kg7 47.Rxf3 Ne4 48.a5 h5 (11.159.049.069) 502

best move: Qg5-e3 time: 716:43.703 min  n/s: 501.710  nodes: 21.575.430.401

Eelco
Parent - - By Felix Kling (Gold) [de] Date 2007-12-30 17:35
"by Yuri Osipov, Igor Korshunov"

Are you sure you didn't forget one author? ;)
Parent - - By Eelco de Groot (***) Date 2007-12-30 18:14
Felix, I think you are absolutely right, I will change it in the engine details at least for any new analysis output. But I'm not sure Vas will really want to be associated with the other two programmers...

Regards, Eelco


Russian Belka squirrel-monster
Parent - - By turbojuice1122 (Gold) [us] Date 2007-12-30 19:47
Awesome picture!
Parent - - By Eelco de Groot (***) Date 2007-12-30 21:42

> Awesome picture!


Belka clearly is some form of hybrid! :) But which parts are from which animal?

24 ply result now in from Glaurung, Nf5 is still a draw if this is correct:

3r1r1k/4b2p/Q1Bp1pp1/1p2p1q1/8/P1P2Pn1/1P1R1B2/1K1R4 b - -


Engine: Glaurung 2.0.1 default (256 MB)
by Tord Romstad

24     398:01 0.00     35...Nf5 36.Qxb5 Ng7 37.Qb7 Nf5
                       38.Qb5 Ng7 (15.830.128.594) 662

24     398:01 -0.03    35...Rg8 36.Bd5 Rg7 37.Qxb5 Nh5
                       38.Bc6 Nf4 39.a4 Qf5+ 40.Be4 Qg5
                       41.a5 f5 42.Bb7 Ng2 43.Bb6 Rf8 44.a6 Ne3
                       45.a7 Nxd1 46.Rxd1 Rgg8 47.Qa4 Qf4
                       48.Qxf4 exf4 (15.830.128.594) 662

Eelco
Attachment: Sampleshoot-out.pgn (8k)
Parent - - By turbojuice1122 (Gold) [us] Date 2008-01-01 20:31
Belka clearly is some form of hybrid!

It sounds like thus far, we have something like the following:

70% fish
15% fruit
5% cat
5% dragon
5% other
Parent - - By Harvey Williamson (*****) Date 2008-01-01 20:36
I feel a recipe coming on ;-)
Parent - - By turbojuice1122 (Gold) [us] Date 2008-01-01 21:38
How to make a Belka Dinner:

5 rybka fillets
2 apples
1 tablespoon cinnamon
1 fillet of glaurung tail steak
1 small wildcat thigh
1 carrot
1 head broccoli
1/3 cup water
1/4 cup olive oil

Preheat oven to 400 degrees Fahrenheit.  Pour the water into a large baking pot and place the rybka fillets in the pot.  Dice the apples and add to the pot; sprinkle the cinnamon on the apples.  In smaller separate pot, add olive oil and then add the cat thigh.  Slice the carrot and add to the pan with the cat.  Finally, add the glaurung tail steak.  Chop up the broccoli and add to the pot.  Place both pots in oven and bake for 20 minutes.
Parent - - By Harvey Williamson (*****) Date 2008-01-01 21:40
How about garnishing with a few deranged posts ;-)
Parent - By turbojuice1122 (Gold) [us] Date 2008-01-01 21:42
That would only occur during a match between Belka and Rocket in the Ultimate Clone Computer Chess Championship.
Parent - - By Banned for Life (Gold) Date 2008-01-01 22:09
I think this recipe needs a little bit of rat to make it complete.

Alan
Parent - By turbojuice1122 (Gold) [us] Date 2008-01-01 22:10
Ha!  Classic!  :-)
Parent - - By Vasik Rajlich (Silver) [hu] Date 2008-01-03 14:55
The PV looks very unconvincing, 35. .. Nf5 36. Qxb5 Ng7 37 a4!

I'll be shocked if black survives this.

Vas
Parent - - By Eelco de Groot (***) Date 2008-01-03 15:59
Thanks Vas and Ansari,

Okay, I'm some more convinced now about White's winning chances. I can only give very low quality analysis, with some moves that I tried myself but it was just to get an impression myself how good White's chances were here, the programs I tried seemed reluctant to try pushing pawns where it seemed fairly safe, so my own coffee house analysis following the original game but with the better 36... Rc8 is something like this:

[36... Rc8 37. Qxb5 f4 38. a4 

{The diagrammed position in my other post}

38... Rb8 {Useful for pinning the Bishop I suppose}

{
1r3r1k/4b2p/1BBp2p1/1Q2p1q1/P4p2/2P2Pn1/1P1R4/1K1R4 w - -


Engine: HIARCS 11.2 SP (256 MB)
by Mark John Uniacke

20/48  258:18 +0.30    39.a5 Nf5 40.Qd3 Ne3 41.Bxe3 fxe3
                       42.Rh2 Rbc8 43.Be4 Kg7 44.a6 Rc7
                       45.Rhh1 Qf4 46.Rdg1 e2 47.Qxe2 Rfc8
                       48.Bd5 (1.802.695.374) 116
}

39. a5 Nf5 40. b4 { was just my of own suggestions as far as I remember } Ne3 41. Rc1 Qf5+
42. Kb2 {  Kb2 and Rc1 are just my own suggestions, HIARCS played something different but I don't see why this is very wrong } 42... Bd8 ( { Or for instance something like } 42... Qe6 43. Bd5 Qc8 44. c4 Nxd5 45. Qxd5 h5 ( { Or maybe } 45... Rf5 46. c5 dxc5 47. Bxc5 Qe8 48. Qe6 Bxc5 49. Qxe8+ Rxe8 50. Rxc5 e4 51. Rxf5 gxf5 52. a6 h5 53. b5 {This is now easily shown won for White}) 46. b5 h4 47. Bf2 g5 48. Ra1 Kg7 49. a6 {Black can't do much against this threatening advance, and on the other side of the board it is harder to push pawns down for Black. White is winning } ) { this position gives me some problems again, 42... Bd8! seems like a strong resource, in the variation that I would play myself I can't find any advantage for White anymore. Maybe that is because my suggestions were not good enough for White. I present them anyway as a counterweight to all the computer output. I'm still doing just part one in the Coffee house training course :)



http://www.schaakcursus.com/

}  43. Rxd6 Bxb6 44. axb6
Rxb6 45. Qc5 Rc8 {computer analysis below *1}]

Eelco

*1

2r4k/7p/1rBR2p1/2Q1pq2/1P3p2/2P1nP2/1K6/2R5 w - -


Engine: HIARCS 11.2 SP (256 MB)
by Mark John Uniacke

7/19   0:00   -0.79    46.c4 Rbb8 47.b5 Rc7 48.b6 Kg7
                       49.Qb5 Re7 (80.230) 146

8/21   0:00   -0.56    46.c4 Rbb8 47.b5 Rc7 48.b6 Rcc8
                       49.Qb5 (115.869) 145

9/25   0:02   -0.56    46.c4 Rb7 47.b5 h5 48.Qb4 (303.422) 145

10/25  0:04   -0.51    46.c4 Rbb8 47.b5 Rc7 48.b6 Kg7
                       49.Qb4 Rf7 50.c5 (704.303) 145

11/29  0:10   -0.43    46.c4 Rbb8 47.b5 Rc7 48.b6 Qg5
                       49.Re6 Rcc8 50.Qb5 Rd8 51.Rxe5 Qg3

(1.475.187) 146

12/31  0:21   -0.55    46.c4 Rbb8 47.b5 Rc7 48.b6 Qg5
                       49.Qb5 Re7 50.Rd2 Qf6 51.c5 e4+
                       52.Ka3 exf3 53.Bxf3 (3.120.499) 144

13/33  0:52   -0.54    46.c4 Rbb8 47.Kb3 Rc7 (7.591.053) 144

14/40  2:17   -0.29++  46.c4 (19.510.305) 142

14/40  2:25   -0.28    46.c4 Rbb8 47.b5 Rc7 48.b6 Qg5
                       49.Re6 Rcc8 50.Qb5 Qf5 (20.750.444) 142

15/40  5:46   -0.12    46.c4 Rb7 47.b5 Rg7 48.b6 Qg5 49.Qb5 Rb8
                       50.Be4 Re7 51.c5 Qg2+ 52.Ka3 Ree8
                       53.Rd7 Rec8 (48.479.098) 139

16/45  12:46  0.00     46.c4 Rb7 47.b5 Rg7 48.b6 Qf7
                       49.Qxe5 Nxc4+ 50.Rxc4 (103.595.458) 135 TB:1

17/45  23:59  -0.01    46.c4 Rb7 47.b5 Rg7 48.b6 Qf8
                       49.Qxe5 Nf5 50.Rxg6 Qb4+ 51.Ka1 Qxb6
                       52.Rxg7 Nxg7 53.Bd5 Qa7+ 54.Kb2 Qe3
                       55.Qxe3 fxe3 56.Re1 (195.944.300) 136 TB:4

18/49  48:32  0.00     46.c4 Rb7 47.b5 Rg7 48.Kb3 h5 49.b6 Qg5
                       50.Re6 Qd8 51.Qb4 Qg8 52.Rf6 Qd8
                       53.Re6 (400.534.207) 137 TB:13
Parent - By Vasik Rajlich (Silver) [hu] Date 2008-01-04 10:37
I'd probably prefer including 38. Bb7 in this line, to discourage .. Rb8. Then 38. .. Qf5+ 39. Rc2 Rce8 40. a4 should be very good for white. The one thing white has to watch out for is a good exchange sacrifice, perhaps something like 39. .. Rb8. This would take some analysis.

Generally, it's difficult to analyze a position which your engines don't understand, it takes a lot more manual work. You can't just trust that automatic 15-ply search to fill in the details.

Vas
Parent - - By Arrière Pensée (Gold) Date 2007-12-30 22:39
That is some nasty chess chick forum member after someones nuts.
Parent - By Banned for Life (Gold) Date 2007-12-30 23:27
:-) :-) :-)
Parent - - By Eelco de Groot (***) Date 2008-01-01 14:56 Edited 2008-01-01 15:02
I have a little more analysis for the die hards that suggests that 35... f5 was not that bad - pfff, sigh of relief :) , but Rybka did not play the best choice one move later :-P. Now with the best defence the score for 35... f5 is still sinking but not as fast as I'd expected under analysis with my Hiarcs 11.2 SP, with position learning on. I'll just post the analysis, most is in three best moves mode.

First, without retrogade analysis in the position just before 35... f5 I got

20  197:42    0.00   35...h5 36.Bxg3 Qxg3 37.Qb7 Rf7 38.Rxd6 Rxd6 39.Rxd6 Qg1+ 40.Kc2 Rg7 41.Rd7 Qf2+ 42.Rd2 Qg3 43.Rd7 Qf2+ (2.338.310.625) 197

20  197:42   -0.01   35...f5 36.Bb6 Rc8 37.Bb7 Rce8 38.Qxb5 f4 39.c4 Qh5 40.a4 Nf5 41.Rd3 Qg5 42.Rg1 Qh5 43.a5 Bd8 44.Re1 (2.338.310.625) 197

19  197:42   -0.16   35...Nf5 36.Qxb5 Nh4 37.Qa4 Nf5 38.Qe4 Kg7 39.Bd5 Rb8 40.Qd3 Rfe8 41.Rg1 (2.338.310.625) 197

best move: h7-h5 time: 224:36.593 min  n/s: 198.241  nodes: 2.671.621.519

Then I tried some retrogade analysis but the results differed less than I expected. Maybe there is some influence using three best moves that interfere a bit with learning but I don't really see why that should happen. The score for 35... f5 is higher than I expected and you can still doubt whether it is actually losing, at least it is not losing fast!

Starting backtracking after 37. Bc7

1r3r1k/2B1b2p/Q1Bp2p1/1p2ppq1/8/P1P2Pn1/1P1R4/1K1R4 b - -


Engine: HIARCS 11.2 SP (256 MB, Athlon 2009 MegaHertz)
by Mark John Uniacke

19     64:24  -1.07    37...Qe3 38.Rd3 Qf2 39.Bxb8 Rxb8
                       40.Bxb5 Qc5 41.Rd5 Qc7 42.R5d2 Rc8
                       43.Rh2 (984.613.006) 254

19     64:24  -1.14    37...Rbc8 38.Bxd6 Bxd6 39.Rxd6 b4
                       40.cxb4 Rcd8 41.Qd3 Rxd6 42.Qxd6 Qf6
                       43.Qxf6+ Rxf6 44.b5 Ne2 45.Rd8+ Kg7
                       46.Rd7+ Kh8 47.Rc7 Nd4 48.a4 Nxc6
                       49.bxc6 (984.613.006) 254

19     64:24  -1.14    37...b4 38.cxb4 Rbc8 39.Bxd6 Bxd6
                       40.Rxd6 Rcd8 41.Qd3 Rxd6 42.Qxd6 Qf6
                       43.Qxf6+ Rxf6 44.b5 Ne2 45.Rd8+ Kg7
                       46.Rd7+ Kh8 (984.613.006) 254

Taking back one ply but the engine still running so the analysis for 37... Qe3 should have gone into the position learning:

After 36... Rb8

1r3r1k/4b2p/QBBp2p1/1p2ppq1/8/P1P2Pn1/1P1R4/1K1R4 w - -


Engine: HIARCS 11.2 SP (256 MB)
by Mark John Uniacke

20     183:44 +0.97    37.Bc7 Qe3 38.Rd3 Qf2 39.Bxb8 Rxb8
                       40.Bxb5 Qc5 41.Rd5 Qc7 42.Qc6 Qxc6
                       43.Bxc6 Rc8 44.Bd7 Rc7 45.Bb5 h5
                       46.c4 h4 47.R1d2 Kg7 48.b4

(2.725.909.428) 247

20     183:44 +0.91    37.Qa7 Rbc8 38.Bc7 Rf7 39.Bxd6 f4
                       40.Ka1 Qf6 41.Qa6 Rcf8 42.Qxb5 Bxd6
                       43.Rxd6 Qg7 44.Rd7 Qf6 45.Rxf7 Rxf7
                       46.Bd5 Rf8 47.Qb7 Rd8 (2.725.909.428)

247

20     183:44 +0.43    37.Rd5 Rbc8 38.Bb7 Rce8 39.Bc7 b4
                       40.cxb4 e4 41.Bc6 Rc8 42.Bxd6 Bxd6
                       43.Rxd6 (2.725.909.428) 247

Going back one ply:

After 36. Bb6

3r1r1k/4b2p/QBBp2p1/1p2ppq1/8/P1P2Pn1/1P1R4/1K1R4 b - -


Engine: HIARCS 11.2 SP (256 MB)
by Mark John Uniacke

21     264:22 -0.30    36...Rc8 37.Bb7 Rce8 38.Qxb5 f4
                       39.a4 Nf5 40.a5 Bd8 41.Bf2 Ne3
                       42.Bxe3 fxe3 43.Rxd6 Bc7 44.R6d3 Rb8
                       45.a6 Qf5 46.Qd7 Bb6 47.Qxf5 gxf5

(3.901.835.135) 245

21     264:22 -0.97    36...Rb8 37.Bc7 Qe3 38.Rd3 Qf2
                       39.Bxb8 Rxb8 40.Bxb5 Qc5 41.Rd5 Qc7
                       42.Qc6 Qxc6 43.Bxc6 Rc8 44.Bb5 h5
                       45.a4 (3.901.835.135) 245

21     264:22 -1.60    36...f4 37.Bxd8 Rxd8 38.Qxb5 Qf5+
                       39.Ka1 Qf8 40.Bd5 Kg7 41.Rh2 Nf5
                       42.Rdh1 h5 43.Rg1 Ng3 44.c4 g5 45.a4 Rb8

(3.901.835.135) 245

This shows that Rybka's 36... Rb8 was I think worse than 36... Rc8

One ply back

After 35... f5

3r1r1k/4b2p/Q1Bp2p1/1p2ppq1/8/P1P2Pn1/1P1R1B2/1K1R4 w - -


Engine: HIARCS 11.2 SP (256 MB)
by Mark John Uniacke

21     625:53 +0.15    36.Bb6 Rc8 37.Qxb5 f4 (8.834.807.286) 235 TB:10

21     625:53 0.00     36.Ba7 f4 37.Bb6 Rc8 38.Bb7 Rce8
                       39.Bc6 Qf5+ 40.Ka1 b4 (8.834.807.286) 235 TB:10

21     625:53 0.00     36.Bb7 b4 37.axb4 f4 38.Bb6 Rde8
                       39.Bc6 Rc8 40.Bb7 Qf5+ 41.Kc1 (8.834.807.286) 235 TB:10

The evaluation before playing 35... f5 would be about -0.15 for Black if I would take back one more ply and change the sign, while in the first analysis it was -0.01 where the search was not quite so deep, and no learning was involved. I guess HIARCS for some reason does not simply take the evaluation of the deeper search in retrogade analysis, maybe that would disturb the search too much? +0.15 for Toga is not really a winning position! :) But I did not really check any deeper positions in this line. Question remains: how resistant is Rybka's defence if best moves are played?

Eelco
Parent - - By M ANSARI (*****) [kw] Date 2008-01-02 07:14 Edited 2008-01-02 07:17
I also spent a couple of hours on this ... Qh3 was already putting the Queen in a location where it would not do too much good and still get bogged down for a long time, it could possibly be a losing move but that needs hours of analysis as black does seem to have still some resources to defend.  Qxh4 burned the bridges with complications that can only favor white.  It still took some strong play to prove that but really after Qxh4? black is simply lost.  I wouldn't fault Rybka's play for playing Qh3 and later Qxh4 ... I think as an engine it is impossible to show that this is dangerous for white because it feels "like the Queen is over extended".  Also quickly looking at the position after a few pawns are removed it looks like Rybka gets a bunch of connected pawns that could be very dangerous in a resulting endgame once big pieces are exchanged.  Engines need concrete variations and without a deep deep search it seems that the Queen foray is OK.  A strong GM would obviously have avoided Qh3 like the plague without calculating anything.  It just feels like the Queen will get harassed and bogged down for a long long time, and especially taking pawns that are in front of the King would seem to only accelerate white's attack.

This was a perfect position for powerful hardware .... and as we know Toga's hardware was much much stronger than Rybka's.  I would be very interested in Toga's evaluation when Rybka played Qxh4?  Was Toga able to realize after some deep searching that this was a losing move ... or did it just stumble on the winning line due to Rybka's play.
Parent - - By Uri Blass (*****) [il] Date 2008-01-02 10:16
Toga says draw after Qxh4 and expect 33...fxg5 that rybka did not play

399: Rybka 2.3.2a 32-bit, 80'/40+40'/20+20' 2007
3r1r1k/B3b2p/pQ1p1pp1/1p2p1Pn/4B2q/P1P2P2/1P1R4/1K1R4 w - - 0 1


Analysis by Toga II 1.3 Beta1:

33.g5xf6 Nh5xf6 34.Qb6-b7 Nf6xe4 35.f3xe4
  -+  (-1.63)   Depth: 1/7   00:00:00
33.Qb6xa6 f6xg5 34.Qa6xb5
  =  (-0.13)   Depth: 1/7   00:00:00
33.Qb6-c7
  =  (-0.10)   Depth: 1/11   00:00:00
33.Qb6-c7 Rf8-f7
  =  (0.18)   Depth: 2/11   00:00:00
33.Qb6-c7 Rd8-e8 34.Be4-c6 f6xg5 35.Bc6xe8 Rf8xe8
  ³  (-0.45)   Depth: 3/11   00:00:00
33.Qb6-b7 Rd8-e8 34.Qb7xa6 f6xg5 35.Qa6xb5
  =  (-0.13)   Depth: 3/11   00:00:00
33.Qb6-b7 Rd8-e8 34.Qb7xa6 f6xg5 35.Qa6xb5
  =  (-0.13)   Depth: 4/11   00:00:00
33.Qb6-b7 f6-f5 34.Qb7xe7 Rf8-e8 35.Qe7-f7 f5xe4 36.Rd2xd6 Qh4xg5 37.f3xe4
  ³  (-0.31)   Depth: 5/15   00:00:00
33.Qb6-b7 f6-f5 34.Qb7xe7 Rf8-e8 35.Qe7-f7 f5xe4 36.Rd2xd6 Qh4xg5 37.f3xe4
  ³  (-0.31)   Depth: 6/15   00:00:00  22kN
33.Qb6-b7 f6-f5 34.Qb7xe7 f5xe4 35.Ba7-f2 Qh4-f4 36.Bf2-b6
  µ  (-1.00)   Depth: 7/20   00:00:00  54kN
33.Qb6xa6 f6-f5 34.Ba7-f2 Qh4xg5 35.Be4-c6 Rd8-c8 36.Qa6xb5 Nh5-f4
  ³  (-0.33)   Depth: 7/20   00:00:00  101kN
33.Qb6xa6 f6xg5 34.Ba7-b6 Rd8-c8 35.Be4-b7 Rc8-c4 36.Qa6xb5 Qh4-f4 37.Bb7-d5 Qf4-f5+ 38.Bd5-e4
  ³  (-0.37)   Depth: 8/21   00:00:00  482kN
33.Qb6xa6 f6xg5 34.Ba7-b6 Rd8-c8 35.Be4-b7 Rc8-c4 36.Qa6xb5 Nh5-f6 37.Rd2-h2 Qh4xh2 38.Qb5xc4
  ³  (-0.47)   Depth: 9/23   00:00:01  1018kN
33.Qb6xa6 f6xg5 34.Ba7-b6 Rd8-e8 35.Qa6xb5 Nh5-g3 36.Be4-d5 g5-g4 37.f3xg4 Qh4xg4 38.Rd1-g1 Re8-c8 39.Qb5-d3 Be7-h4
  ³  (-0.27)   Depth: 10/31   00:00:03  2272kN
33.Qb6xa6 f6xg5 34.Ba7-b6 Rd8-e8 35.Be4-c6 Re8-b8 36.Bb6-c7 Rb8-c8 37.Bc7xd6 Be7xd6 38.Rd2xd6 Rc8-b8 39.Rd6-d7 Nh5-f6 40.Rd7-e7
  ³  (-0.26)   Depth: 11/33   00:00:07  4719kN
33.Qb6xa6 f6xg5 34.Ba7-b6 Rd8-e8 35.Be4-c6 Re8-c8 36.Qa6xb5 g5-g4 37.f3xg4 Qh4xg4 38.Bc6-d7 Qg4-e4+ 39.Qb5-d3 Qe4xd3+ 40.Rd2xd3 Rc8-a8 41.c3-c4 Nh5-f4 42.Rd3-f3
  =  (-0.25)   Depth: 12/33   00:00:13  8009kN
33.Qb6xa6 f6xg5 34.Ba7-b6 Rd8-e8 35.Be4-c6 Re8-c8 36.Qa6xb5 g5-g4 37.f3xg4 Qh4xg4 38.Bc6-d7 Qg4-e4+ 39.Qb5-d3 Qe4xd3+ 40.Rd2xd3 Rc8-a8 41.Rd3-d2 Nh5-f4 42.Bd7-c6 Ra8-a6
  =  (-0.18)   Depth: 13/34   00:00:23  15442kN
33.Qb6xa6 f6xg5 34.Ba7-b6 Rd8-e8 35.Be4-c6 Re8-c8 36.Bc6-b7 Rc8-c4 37.Qa6xb5 Nh5-f6 38.Rd2-g2 Qh4-f4 39.Bb6-e3 Qf4-f5+ 40.Kb1-a1 d6-d5 41.Be3xg5 Qf5xf3 42.Bg5xf6+ Rf8xf6 43.Qb5xd5 Qf3xd5 44.Bb7xd5
  =  (-0.15)   Depth: 14/42   00:00:46  31463kN
33.Qb6xa6 f6xg5 34.Ba7-b6 Rd8-e8 35.Be4-c6 Re8-c8 36.Bc6-b7 Rc8-e8 37.Bb7-c6
  =  (0.00)   Depth: 15/42   00:02:20  98252kN
33.Qb6xa6 f6xg5 34.Ba7-b6 Rd8-e8 35.Be4-c6 Re8-c8 36.Bc6-b7 Rc8-e8 37.Bb7-c6
  =  (0.00)   Depth: 16/45   00:03:44  158330kN
33.Qb6xa6 f6xg5 34.Ba7-b6 Rd8-e8 35.Be4-c6 Re8-c8 36.Bc6-b7 Rc8-e8 37.Bb7-c6
  =  (0.00)   Depth: 17/47   00:06:30  279555kN
33.Qb6xa6 f6xg5 34.Ba7-b6 Rd8-e8 35.Be4-c6 Re8-c8 36.Bc6-b7 Rc8-e8 37.Bb7-c6
  =  (0.00)   Depth: 18/50   00:11:29  495633kN

(,  02.01.2008)
Parent - By Uri Blass (*****) [il] Date 2008-01-02 16:15
I simply used different toga than the toga that was used in the game

Toga that was used in the game is happy with Qxa6 and expect rybka's move in the game

399: Rybka 2.3.2a 32-bit, 80'/40+40'/20+20' 2007
3r1r1k/B3b2p/pQ1p1pp1/1p2p1Pn/4B2q/P1P2P2/1P1R4/1K1R4 w - - 0 1


Analysis by Toga II 1.3x4:

33.g5xf6 Nh5xf6 34.Be4-c6
  µ  (-0.86)   Depth: 1/7   00:00:00
33.Qb6xa6 f6xg5 34.Qa6xb5
  ²  (0.69)   Depth: 1/7   00:00:00
33.Qb6xa6 f6xg5 34.Qa6xb5
  ²  (0.69)   Depth: 2/11   00:00:00
33.Qb6-c7 f6-f5 34.Qc7xe7 f5xe4 35.Rd2xd6
  ±  (0.76)   Depth: 2/11   00:00:00
33.Qb6-c7 Rd8-e8 34.Be4-c6 Qh4xg5 35.Bc6xe8 Rf8xe8
  ²  (0.30)   Depth: 3/13   00:00:00
33.Qb6-b7 Rf8-f7 34.Ba7-b6
  ²  (0.62)   Depth: 3/13   00:00:00
33.Qb6-b7 Rd8-e8 34.Qb7xa6 f6xg5 35.Qa6xb5
  ²  (0.68)   Depth: 4/13   00:00:00
33.Qb6-b7 Rd8-e8 34.Qb7xa6 f6xg5 35.Qa6xb5
  ²  (0.68)   Depth: 5/14   00:00:00
33.Qb6-b7 f6-f5 34.Qb7xe7 Rf8-e8 35.Qe7-f7 f5xe4 36.Rd2xd6 Qh4xg5 37.f3xe4
  =  (0.18)   Depth: 6/17   00:00:00  28kN
33.Qb6-b7 f6-f5 34.Qb7xe7 Rf8-e8 35.Qe7-f7 Re8-f8 36.Qf7-e7
  =  (0.00)   Depth: 7/21   00:00:00  75kN
33.Qb6xa6 f6-f5 34.Be4-b7 Be7xg5 35.Rd2xd6 Qh4-c4 36.Ba7-e3 Rd8xd6 37.Qa6xd6
  =  (0.15)   Depth: 7/21   00:00:00  92kN
33.Qb6xa6 f6-f5 34.Be4-c6 Be7xg5 35.Rd2xd6 Rd8xd6 36.Rd1xd6 Bg5-e7 37.Rd6-d1
  =  (0.16)   Depth: 8/21   00:00:00  182kN
33.Qb6xa6 f6-f5 34.Be4-b7 Be7xg5 35.Rd2xd6 Qh4-a4 36.Qa6xa4 b5xa4 37.Rd6-d7 Rd8xd7 38.Rd1xd7
  =  (0.23)   Depth: 9/23   00:00:01  356kN
33.Qb6xa6 f6-f5 34.Be4-c6 b5-b4 35.c3xb4 Qh4xg5 36.Ba7-b6 Rd8-c8 37.Bc6-b7 Rc8-b8 38.Bb7-d5
  ²  (0.37)   Depth: 10/25   00:00:01  902kN
33.Qb6xa6 f6xg5 34.Ba7-f2 Qh4-h3 35.Bf2-b6 Rd8-c8 36.Qa6xb5 Nh5-f6 37.Be4-b7 Rc8-b8 38.Bb7-c6 Qh3-f5+ 39.Kb1-a1
  =  (0.21)   Depth: 11/28   00:00:03  1660kN
33.Qb6xa6 f6xg5 34.Ba7-f2 Qh4-h3 35.Bf2-b6 Rd8-c8 36.Qa6xb5 Nh5-f6 37.Be4-b7 Rc8-b8 38.Bb7-c6 Qh3-f5+ 39.Kb1-a1
  =  (0.21)   Depth: 12/32   00:00:04  2743kN
33.Qb6xa6 f6xg5 34.Ba7-b6 Rd8-e8 35.Be4-c6 Re8-c8 36.Bc6-b7 Rc8-c4 37.Bb7-d5 Rc4-f4 38.Qa6xb5 Nh5-f6 39.Bd5-c6
  =  (0.19)   Depth: 13/34   00:00:08  5502kN
33.Qb6xa6 f6xg5 34.Ba7-b6 Rd8-c8 35.Be4-b7 Rc8-c4 36.Qa6xb5 Nh5-f6 37.Bb6-f2 Qh4-f4 38.Bf2-e3 Qf4-f5+ 39.Kb1-a1 Rc4-h4 40.c3-c4
  ²  (0.35)   Depth: 14/38   00:00:14  9495kN
33.Qb6xa6 Nh5-g3 34.Ba7-f2 f6xg5 35.Qa6-a7 Rf8-f7 36.Be4-d5 Rf7-g7 37.Qa7-b6 Qh4-f4 38.Rd2-d3 Rd8-c8 39.Qb6-b7 Rc8-f8 40.Qb7xb5
  ²  (0.36)   Depth: 15/38   00:00:38  25382kN
33.Qb6xa6 Nh5-g3 34.Ba7-f2 Qh4xg5 35.Be4-c6 Ng3-f5 36.Rd2-d3 b5-b4 37.c3xb4 Qg5-f4 38.Bc6-b7 Rd8-d7 39.Rd3-c3 Qf4-h2 40.Bb7-c8 Rd7-d8 41.Bc8xf5 Qh2xf2
  ²  (0.40)   Depth: 16/40   00:01:11  46712kN
33.Qb6xa6 Nh5-g3 34.Ba7-f2 Qh4xg5 35.Be4-d5 Qg5-f5+ 36.Kb1-a1 Ng3-h5 37.Bf2-b6 Rd8-b8 38.Qa6-a7 Rf8-e8 39.Bd5-c6 Re8-g8 40.Bc6-e4 Qf5-h3 41.Be4-d5 Rg8-c8
  ²  (0.36)   Depth: 17/50   00:02:01  80513kN
33.Qb6xa6 Nh5-g3 34.Ba7-f2 Qh4xg5 35.Be4-c6 Ng3-f5 36.Rd2-d3 b5-b4 37.c3xb4 Qg5-g2 38.Bf2-b6 Rd8-c8 39.Rd3-d2 Qg2-g5 40.Bc6-b7 Rc8-b8 41.Bb6-a7 Nf5-e3 42.Ba7xb8 Ne3xd1
  ²  (0.40)   Depth: 18/57   00:04:29  178897kN

(,  02.01.2008)
Parent - By Eelco de Groot (***) Date 2008-01-02 11:17
Hello Ansari,

> It still took some strong play to prove that but really after Qxh4? black is simply lost.


I am not convinced about the simply lost, can't see the programs could so totally misjudge this. I think the following position although critical is still a draw unless you can show me how White wins it, I agree it is a precarious position but if it is really won by pushing pawns it should not be so hard for the programs to see. Or the programmers have some homework to do. Right now this is what Hiarcs is showing:

2r2r1k/4b2p/1BBp2p1/1Q2p1q1/P4p2/2P2Pn1/1P1R4/1K1R4 b - -


Engine: HIARCS 11.2 SP (256 MB) (two best search)
by Mark John Uniacke

18     61:15  -0.14    38...Rb8 39.a5 Nf5 40.a6 Bd8 41.a7 Rxb6
                       42.Qa4 Bc7 43.Be4 Qf6 44.a8Q Rxa8
                       45.Qxa8+ Rb8 46.Qa7 Qe7 47.Bxf5 gxf5
                       48.Rh2 (286.432.988) 77

18     61:15  -0.35    38...Nf5 39.Bb7 Rce8 40.a5 Bd8
                       41.Bf2 (286.432.988) 77

2r2r1k/4b2p/1BBp2p1/1Q2p1q1/P4p2/2P2Pn1/1P1R4/1K1R4 b - -


Engine: Belka 1.8.11 (64 MB)
by Vasik Rajlich, Yuri Osipov, Igor Korshunov

17.32  8:30   -0.92    38...Qf5+ 39.Rd3 Rb8 40.a5 Qf7
                       41.Qd5 Qxd5 42.Bxd5 Nf5 43.Ba7 Rb5
                       44.b4 Rc8 45.Bf2 Kg7 46.a6 Bh4
                       47.Be6 Ra8 (158.113.215) 309

18.32  18:07  -0.89    38...Qf5+ 39.Rd3 Rb8 40.a5 Qf7
                       41.Qd5 Qxd5 42.Bxd5 Nf5 43.Ba7 Rbc8
                       44.c4 h5 45.Rb3 Kg7 46.a6 h4 47.Bf2 Ne3
                       48.Bxe3 fxe3 49.Rxe3 (324.038.081) 297

And this one just in: 19.32  82:11  -0.59    38...Rb8 39.a5 Nf5 40.Rh1 Bd8 41.Qa6 Bxb6
                       42.axb6 h5 43.b7 Kg7 44.Be4 h4
                       45.Rhh2 Rh8 46.Qc6 Qe7 47.Bxf5 gxf5
                       48.Rhg2+ Kf7 49.Rxd6 Rxb7 (1.532.968.698) 310

2r2r1k/4b2p/1BBp2p1/1Q2p1q1/P4p2/2P2Pn1/1P1R4/1K1R4 b - -


Engine: Glaurung 2.0.1 default (64 MB)
by Tord Romstad

19.01  22:30  +0.17    38...Rb8 39.Rxd6 Bxd6 40.Rxd6 Nf1
                       41.Rd5 Rf5 42.a5 Ne3 43.a6 Nxd5
                       44.a7 Rc8 45.Bxd5 Rff8 46.Qa5 Ra8
                       47.Bxa8 Rxa8 (327.990.247) 242

20.01  47:11  -0.07    38...Rb8 39.Rxd6 Bxd6 40.Rxd6 Nf1
                       41.Rd5 Rf5 42.a5 Ne3 43.a6 Nxd5
                       44.a7 Rc8 45.Bxd5 Rff8 46.Qa5 Ra8
                       47.Bxa8 Rxa8 48.b4 (679.076.096) 239

Regards, Eelco
Parent - - By Vasik Rajlich (Silver) [hu] Date 2008-01-03 14:57
I agree with your comments, except the part about Toga's hardware. 2x or even 10x more speed here won't do much. No engine understands this, they're all just rolling dice.

Vas
Parent - - By Banned for Life (Gold) Date 2008-01-03 19:16
This is a perfect example of a position where DPA should give better results than spending the equivalent amount of time performing analysis from the root.

Alan
Parent - - By Vasik Rajlich (Silver) [hu] Date 2008-01-04 10:26
Yes, you're right about this.

I'm starting to get a little bit more open to this type of idea. DPA is a bit similar to a randomizer/Monte Carlo approach, in the sense that you emphasize depth over breadth. It's also easier to divide the work with this type of approach, making it suitable for large machines with large latencies (such as clusters).

Vas
Parent - - By Banned for Life (Gold) Date 2008-01-04 18:44
The randomizer is actually a superset of DPA and has a lot of potentially significant advantages. What it is missing right now is the tool to take the games that are played and construct a tree which attempts to rank the lines that have been analyzed from most promising to least promising based on analysis at the leaf nodes. This is an easier task for DPA since it performs an exhaustive fixed-depth search with constrained branching, but it should also be possible with the randomizer. I suspect this would give better results than looking at the percentages at the root.

Regards,
Alan
Parent - - By Vasik Rajlich (Silver) [hu] Date 2008-01-05 14:02
This is also an interesting question - how to propagate scores up the tree.

Let's say that somewhere inside the tree, two moves were tried. The first move scored 75% after 1,000 games and the second move scored 25% after 1,000 games. What score do you assign to the parent?

The Monte Carlo approach is to assign the average, ie. 50%. There is also something to be said for performing a mini-max and assigning the 75% to the parent. In this case, you could apply a sort of beta cutoff and stop playing games with the second move once it's clear that its performance will be worse.

Some sort of hybrid approach may also be possible. For human-in-the-loop analysis, it makes sense to let the user control this.

It will all need lots of tinkering.

Vas
Parent - - By Banned for Life (Gold) Date 2008-01-06 01:28
Mini-max certainly seems like a better solution than averaging, even with a sparsely analyzed tree. Once you picked the best lines in this manner, you could verify that something important wasn't missed by doing backward analysis with much greater depth on the best candidate lines. This is actually very similar to the procedure I follow for correspondence chess (I'm not sure that's an endorsement though :-)).

Rather than cutting off lines that have a low percentage score, you might reduce their frequency. Having the probability of using each line be directly proportional to its success would ensure that the best line would never be rejected, even if it got off to a very bad start.

Regards,
Alan
Parent - - By Vasik Rajlich (Silver) [hu] Date 2008-01-06 14:46

> Mini-max certainly seems like a better solution than averaging, even with a sparsely analyzed tree.


I'm not so sure about this.

If move 1 scored 75% after two games and move 2 scored 25% after two games, I would think that averaging is better. I also suspect that averaging is better when the scores are close.

> Once you picked the best lines in this manner, you could verify that something important wasn't missed by doing backward analysis with much greater depth on the best candidate lines.


I'm not too worried about missing something important. The point of Monte Carlo is that you have a very "rich" sample, covering all types of different variations. When you do a mini-max on the results, you lose some of this richness.

> Rather than cutting off lines that have a low percentage score, you might reduce their frequency.


Yes, that makes sense.

Vas
Parent - - By ralu (*) [si] Date 2008-01-22 01:05
Instead of just making alpha-beta or avarage could do statistical aprocah whit 2 parameters. 1st mean and 2nd deviation.  That mean that scores whit higher deviation are also more dynamic.  This information can be propagated form deeper to higher levels
Parent - By Vasik Rajlich (Silver) [hu] Date 2008-01-22 14:27
There is probably some sort of mathematically sound procedure for handling this. I haven't tried to work this out.

Vas
Parent - - By Banned for Life (Gold) Date 2008-01-22 03:34
If move 1 scored 75% after two games and move 2 scored 25% after two games, I would think that averaging is better. I also suspect that averaging is better when the scores are close.

Aren't these conclusions influenced by the small sample size you've chosen (2)? Suppose move 1 scored 75% after 2000 games and move 2 scored 25% after 2000 games? I'm not arguing that min-max is necessarily better since its easy to set up an example where the 25% path actually contains the better move sequence (along with a lot of losing moves). Of course if you bias the percentages taking each path rather than cut off the lower percentage path you should eventually arrive at the correct solution.

Alan
Parent - By Vasik Rajlich (Silver) [hu] Date 2008-01-22 14:30
I just wanted to show that the minimax approach is not necessarily better.

Even if you "bias the percentages" and start playing fewer games with the weaker moves, this still leaves the question of what score you assign to the parent.

Vas
Parent - - By Dadi Jonsson (Silver) [is] Date 2008-01-02 23:51

> 1) How quickly does the score rise for white with a really deep normal search? After an hour or so on a core duo laptop, Rybka gets to +0.26 (after I force 33. Qxa6).


Rybka 2.3.2a mp 64bit (8 cores):

3r1r1k/B3b2p/Q2p1pp1/1p2p1Pn/4B2q/P1P2P2/1P1R4/1K1R4 b - -

  [+0.31]  d=29  1...fxg5 2.Bb6 Rb8 3.Bc7 Nf6 4.Bc2 Rbe8 5.Qxb5 Qf4 6.Bxd6 Bxd6 7.Rxd6 e4 8.fxe4 (30:26.52)
  [+0.28]  d=28  1...fxg5 2.Bb6 Rb8 3.Bc7 Nf6 4.Bc2 Rbe8 5.Qxb5 Qf4 6.Bxd6 Bxd6 7.Rxd6 e4 8.fxe4 (28:01.31)
  [+0.40]  d=28  1...Ng3 2.Bf2 Qxg5 3.Bb7 Rb8 4.Qxb5 e4 5.Rd5 Qf4 6.Bxg3 Qxg3 7.fxe4 Qf3 8.R5d4 (20:22.54)
  [+0.31]  d=27  1...Ng3 2.Bf2 Qxg5 3.Bd5 Nh5 4.Qxb5 Nf4 5.Be3 Qf5 6.Be4 Qd7 7.a4 Kg7 8.Bxf4 (4:27.50)
  [+0.32]  d=26  1...Ng3 2.Bc6 fxg5 3.Bf2 Qf4 4.Bxg3 Qxg3 5.Qb7 Rf7 6.Rxd6 Rxd6 7.Rxd6 Bxd6 8.Qxf7 (2:11.11)
  [+0.26]  d=25  1...Ng3 2.Bc6 fxg5 3.Bf2 Qf4 4.Bxg3 Qxg3 5.Qb7 Rf7 6.Rxd6 Rxd6 7.Rxd6 Bxd6 8.Qxf7 (1:20.51)
  [+0.26]  d=24  1...Ng3 2.Qb7 Nxe4 3.fxe4 Rfe8 4.gxf6 Bxf6 5.Qd5 Bg5 6.Re2 Qg4 7.Qxb5 Qd7 8.Qxd7 (0:31.47)
  [+0.22]  d=23  1...Ng3 2.Qb7 Nxe4 3.fxe4 Rfe8 4.gxf6 Bxf6 5.Qd5 Bg5 6.Re2 Qg4 7.Qxb5 Qd7 8.Qxd7 (0:25.46)
  [+0.24]  d=22  1...Ng3 2.Qb7 f5 3.Bc2 Qxg5 4.Qxb5 Ra8 5.Qd7 Nh5 6.a4 Rfd8 7.Qb7 Bf8 8.a5 (0:10.47)
  [+0.18]  d=21  1...Ng3 2.Qb7 f5 3.Bc2 Qxg5 4.Qxb5 Ra8 5.Qd7 Qf6 6.Bb6 h5 7.a4 Qf7 8.a5 (0:03.49)
  [+0.16]  d=20  1...Ng3 2.Qb7 f5 3.Bc2 Qxg5 4.Qxb5 Ra8 5.Qb7 Qf6 6.Bb3 f4 7.Bd5 Kg7 8.Rg1 (0:02.33)
  [+0.21]  d=20  1...fxg5 2.Bb6 Rc8 3.Bb7 Rce8 4.Qxb5 Ng3 5.Rd4 Qh3 6.R4d3 Nf5 7.f4 Qh2 8.fxg5 (0:01.17)
  [+0.16]  d=19  1...fxg5 2.Bb6 Rc8 3.Bb7 Rce8 4.Qxb5 Ng3 5.Rg2 Qh3 6.Rgg1 Qf5 7.Ka1 Qc2 8.Qd5 (0:00.54)
  [+0.15]  d=18  1...fxg5 2.Bb6 Rc8 3.Bb7 Rce8 4.Qxb5 Ng3 5.Rg2 Qh3 6.Rgg1 Qf5 7.Ka1 Qc2 8.Qd5 (0:00.44)
  [+0.15]  d=17  1...fxg5 2.Bb6 Rc8 3.Bb7 Rce8 4.Qxb5 Ng3 5.Rg2 Qh3 6.Rdg1 Nf5 7.Rd2 Qh4 8.Rdd1 (0:00.39)
  [+0.16]  d=17  1...Ng3 2.Qb7 f5 3.Bc2 Qxg5 4.Qxb5 Rfe8 5.Qb7 Qh4 6.Bb3 Nh5 7.Rg2 Bf6 8.Qc6 (0:00.29)
  [+0.15]  d=16  1...Ng3 2.Qb7 f5 3.Bc2 Qxg5 4.Qxb5 h5 5.a4 h4 6.a5 Kg7 7.a6 h3 8.Bb6 (0:00.23)
  [+0.08]  d=15  1...Ng3 2.Qb7 f5 3.Bc2 Qxg5 4.Qxb5 h5 5.Qb7 h4 6.Bb3 Nf1 7.Rc2 Ng3 (0:00.21)
  [+0.19]  d=14  1...Ng3 2.Qb7 f5 3.Bc6 Qxg5 4.Bxb5 f4 5.a4 Qf5 6.Rc2 Qe6 7.a5 Nf5 (0:00.09)
  [+0.20]  d=13  1...Ng3 2.Qb7 f5 3.Bc6 Qxg5 4.Bxb5 f4 5.Bd7 Qh4 6.Be6 Nf5 (0:00.07)

After this I ran the position through Interactive analysis (automatic) in Chess Assistant for about a quarter of the time I used for infinite analysis above. It came up with the exact same main variation. When I stopped it, it had reached depth 22 or 23. See the attachment for the output of the analysis.

Parent - By Vasik Rajlich (Silver) [hu] Date 2008-01-03 15:22
Thanks, that's what I wanted to see. The score rises really slowly.

When I analyze with Rybka, taking her moves for black, I can pretty much always drive up the white score after a couple of tries.

Vas
Up Topic Rybka Support & Discussion / Rybka Discussion / Cluster Toga - Rybka

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill