Position in Rybka vs. Diep

[Rolf]

Better for black from the game would have been 39. Kh1 b3 40. Rf1 Qe4+ 41. Kh2

And black still has a slight advantage.

[d]1r4k1/5pp1/7p/3p4/Q1p1q3/1p5P/PP5K/1R3R2 b

Who agrees with me that it's not the question who finds either Rxf3 or at first Qg6 then after Rg1 Rxg2!, the question is who sees it that with the three connected Pawns Black has compensation and even more, a win? I ffor one see this as the influence of GM Kaufman on Rybka tuning.

Of course also the hardware is relevanrt. On my 2 core with Rybka 3 1-cpu Rybka sees a slight advantage just like the other engines who would also play Rxf3 intending to seek a Draw.

So this is a fantastic example for a computer software finding chess moves for the humanly correct reasons! That alone puts Rybka on top of all other actual machine entities. IMO.

[michiguel]

How long does it take your program to find Rxf3?

In the game Rybka vs. Diep the following position was reached:

[d] 1r4k1/5pp1/7p/3p4/Qpp1q3/5N1P/PP3rP1/1R1R3K b

The Baron plays Rxf3 instantly, settling for a draw.
Richard.

Belka, Crafty, Spike would play Rxf3 for a draw too.

And Gaviota

There is nothing spectacular about tihis move.

Miguel

[Dan]

yep...

[M ANSARI]

I think the main issue here is not the actual exchange sac, but the resulting position after. It would be impossible for even a cluster to search deep enough for an advanatage with resulting passed pawns against rook. I have actually seen this where it is a weakness in Rybka and many times these pawns will be weak and will fall and Rybka will lose. I know that Vas has been working very hard on this and it seems to have paid off here.

The exchange sac is only impressive is the one sacrificing is going for a win and in Rybka's case it does look like the engine though it was getting a tangible advantage, and that is something that I don't think any other engine would do. Anyway let us see if we can get the evaluation profiles from Rybka in all the games, I am really interested to see what changes Vas has made in evaluation.

[Rolf]

The exchange sac is only impressive is the one sacrificing is going for a win and in Rybka's case it does look like the engine though it was getting a tangible advantage, and that is something that I don't think any other engine would do. Anyway let us see if we can get the evaluation profiles from Rybka in all the games, I am really interested to see what changes Vas has made in evaluation.

Please, let's care for that nobody ever mentions Larry Kaufman and his GM title. Let's only examine the output and the evaluations.

Let's not praise creativity but let's try to look and see and then copy and paste. Ok, more or less. Humbly yours truly.

[Pressie]

This specific position may not have been the best example of Rybka's play in the tournament. Over time we are able to more closely examine positions and determine whether a move truly is "brilliant" or perhaps in this case the loser blundered and simply lost? There are other qualities in the development of a program that can be taken into consideration, as pointed out byAnsari and Rolf.

One thing that is indisputable, "Rybka is clearly head and shoulders above it's competition, and has been for a number of years now." And that alone speaks volumes about the strength and uniqueness of the program.

[Rolf]

Thanks so far, Terry. Let me ask a general question that could perhaps become a seperate thread if others are interested. But before I ask I want to mention that of course I am no expert in computerchess programming nor in programming at all. But in my studies in psychology, what others might confound with psychoanalysis, I got a basic education in stats, tests and en plus out of my own initiative a lot of methodology and the science of science so to speak. So everytime you are talking on the base of something specifically rooted in the long history of computerchess programing, then I'm a lay, but, if you discuss, how a display might show with its output, what the engine is doing and for all what this could mean for a single chess position and then for a whole game, then I as an observer can well enough judge the quality of arguments. End of preface.

Now my question.

At what moment the experts here would accept in a position like the given just before the Rxf3 move, that the machine player had seen enough advantages for a practical game against a particular opponent -

if we know that a machine has a limited horizon other than a human expert player?

Isnt it already enough, if the Rybka side does know that they have implemented enough knowledge for Pawn chain structures, that gives them advantages over all competitors?

See the Diep game and compare it with the Sjeng game. I read (in the Rybka forum) that the win in the latter game came out of whatever nowhere. As a surprise. But I can well compare the 3 vs 2 chains in Diep game with the much more crowded chains in the other.

IMO it's weak to argue that Diep should have played Kh1 or g1 to save the Draw, if the knowledge at hand told Diep that the chain fight 2 vs 3 could be drawn anyway. Here we have exactly the difference. Rybka like others has a similar machine horizon limitation but what does this matter, if Rybka then has the joker of the better P handling and can well rely on faulty evaluations by its opponents.

We human chess players always make the same mistake if we comment on machines. We know the outcome or better the decisive point in the end and sort of back-roll the problem to the position where perhaps another K move mattered. But how could that matter for a machine that is thinking/because being told so, that the unbalanced P structure on the other side would be drawish?

The reports about also Rybka already losing such games does only show to me that Rybka is by far not yet the perfect player. Let me end with a little joke. Like Philidor already said, the secret of the game of chess is in the power of the Pawns!

To all competitors of Rybka! Please dont take this all too seriously what Rolf says, for all delete it in your memory! <cough>

[Eelco de Groot]

Better for black from the game would have been 39. Kh1 b3 40. Rf1 Qe4+ 41. Kh2

And black still has a slight advantage.

[d]1r4k1/5pp1/7p/3p4/Q1p1q3/1p5P/PP5K/1R3R2 b

In the line as given by Blueberry Qe4+ is not possible of course but in the game the Queen was on f4, not on f2. With the Queen on f2, after 39. .. b3 Blueberry still insists on 40. a3 instead of any Rook moves, but I think at this point Black does not have any winning line. Maybe Rybka's Qf4 was better than Qf2?


After 1...Rxf3 2.gxf3 Qxf3+ 3.Kh2

[d]1r4k1/5pp1/7p/3p4/Qpp5/5q1P/PP5K/1R1R4 b - -

Engine: Blueberry Beta 4 DM70 Build 375 (Athlon 2009 MHz,256 MB)
by F. Letouzey, T. Gaksch, E. de Groot

10/60 166:00 +1.02 3...Qf2+ 4.Kh1 b3 5.a3 Qf3+ 6.Kh2 Qe2+
7.Kh1 Re8 8.Rd2 Qf3+ 9.Kg1 Qe3+
10.Rf2 Re6 11.Qa8+ Kh7 12.Qxd5 Rg6+
13.Kf1 Qxh3+ 14.Ke2 Re6+ 15.Kd1 (5.346.896.163) 536

After entering the next moves

[FEN "1r4k1/5pp1/7p/3p4/Qpp1q3/5N1P/PP3rP1/1R1R3K b - -"]

1... Rxf3 2. gxf3 Qxf3+ 3. Kh2 Qf2+ 4. Kh1 b3 *

while the hash results are still in memory I got a glitch and Blueberry wants to play axb3?? at first, I think this must have something to do with bad hash results leading the search astray for too long, finding a3 again takes too much time even though it is still found in iteration five. I hope this only happens if I quickly go through the moves in analysis mode:

[d]1r4k1/5pp1/7p/3p4/Q1p5/1p5P/PP3q2/1R1R3K w - -

Engine: Blueberry Beta 4 DM70 Build 375 (Athlon 2009 MHz,256 MB)
by F. Letouzey, T. Gaksch, E. de Groot

1/15 0:00 -0.24 5.a3 Qf3+ 6.Kh2 Rc8 (1.958)

1/17 0:00 0.00 5.axb3 Qf3+ 6.Kh2 Qe2+ 7.Kh1 Qe4+
8.Kg1 Qe3+ 9.Kg2 Qe4+ 10.Kg1 (5.027)

2/26 0:00 -0.44 5.axb3 Qf3+ 6.Kh2 Qf4+ 7.Kh1 Rxb3
8.Qd7 Qf3+ 9.Kg1 Qe3+ 10.Kh1 Qxh3+
11.Qxh3 Rxh3+ 12.Kg2 (172.053)

3/26 0:00 -0.44 5.axb3 Qf3+ 6.Kh2 Qf4+ 7.Kh1 Rxb3
8.Qd7 Qf3+ 9.Kg1 Qe3+ 10.Kh1 Qxh3+
11.Qxh3 Rxh3+ 12.Kg2 (273.804)

4/29 0:01 -0.44 5.axb3 Qf3+ 6.Kh2 Qf4+ 7.Kh1 Rxb3
8.Qd7 Qe4+ 9.Kh2 Qe2+ 10.Kh1 Qf3+
11.Kg1 Qe3+ 12.Kh1 Qxh3+ 13.Qxh3 Rxh3+
14.Kg2 (497.368) 453

5/52 0:19 -3.25 5.axb3 Qf3+ 6.Kg1 Qg3+ 7.Kf1 Rxb3
8.Qa8+ Kh7 9.Qxd5 Rf3+ 10.Qxf3 Qxf3+
11.Kg1 f5 12.Rd2 Qxh3 13.Re1 Qg4+
14.Kf2 Qf4+ 15.Ke2 (9.713.080) 504

5/52 0:27 -1.34 5.Rg1 bxa2 6.Qxa2 Qf3+ 7.Rg2 Qxh3+
8.Kg1 Qe3+ 9.Kh2 Qf4+ 10.Kg1 Qd4+
11.Kh2 Rc8 12.Rf1 Qe5+ 13.Kh1 Qh5+
14.Kg1 (13.205.486) 483

5/52 0:31 -1.34 5.Rg1 bxa2 6.Qxa2 Rb3 7.Qa8+ Kh7
8.Ra1 Rxb2 9.Rxg7+ Kxg7 10.Rg1+ Kf6
11.Qd8+ Ke6 12.Qe8+ Kf5 13.Qxf7+ Ke5
14.Qe8+ Kd6 15.Rg6+ Kc5 16.Qe7+ Kd4
17.Qa7+ Kc3 18.Qxf2 (16.066.920) 514

5/52 1:24 -1.34 5.Rg1 Qf5 6.Kh2 bxa2 7.Rbf1 Rxb2+
8.Rg2 Rxg2+ 9.Kxg2 Qe4+ 10.Kg3 Qe3+
11.Kg2 Kh7 12.Qxa2 Qe4+ 13.Kg1 c3
14.Qa3 d4 15.Rxf7 Qe1+ 16.Rf1 (45.521.293) 541

5/52 1:30 -0.78 5.a3 Qf3+ 6.Kh2 Qe2+ 7.Kh1 Qe4+
8.Kh2 Qe5+ 9.Kh1 Qf5 10.Kh2 c3
11.Qd4 Qc2+ 12.Kh1 cxb2 13.Qxd5 (48.139.370) 534

5/52 1:33 -0.80 5.a3 Qf3+ 6.Kh2 Qe2+ 7.Kh1 Qe4+
8.Kh2 Re8 9.Rd2 c3 10.Qxe4 dxe4
11.bxc3 e3 12.Rxb3 (49.509.304) 532

5/52 1:40 -0.80 5.a3 Qf3+ 6.Kh2 Qe2+ 7.Kh1 Qe4+
8.Kh2 Qe5+ 9.Kh1 c3 10.bxc3 Qxc3
11.Qg4 Qc2 12.Rg1 Qe4+ 13.Qg2 b2
14.a4 (52.827.290) 523

6/52 1:46 -0.66 5.a3 Qf3+ 6.Kh2 Qf4+ 7.Kh1 Qe4+
8.Kh2 Re8 9.Qd7 Qf4+ 10.Kg1 (55.083.188) 519

7/52 1:52 -0.69 5.a3 Qf3+ 6.Kh2 Qe2+ 7.Kh1 Re8 8.Rd2 Qf3+
9.Kg1 Qe3+ 10.Rf2 Re5 11.h4 Qg3+
12.Rg2 Qxh4 13.Qa8+ Kh7 (58.059.308) 514

8/52 2:14 -0.69 5.a3 Qf3+ 6.Kh2 Qe2+ 7.Kh1 Re8 8.Rd2 Qf3+
9.Kg1 Qe3+ 10.Rf2 Re5 11.h4 Qg3+
12.Rg2 Qxh4 13.Qa8+ Kh7 (67.851.466) 503

9/52 2:46 -0.69 5.a3 Qf3+ 6.Kh2 Qe2+ 7.Kh1 Qe4+
8.Kh2 Re8 9.Rd2 Qf4+ 10.Kg1 Qe3+
11.Rf2 Re5 12.h4 Qg3+ 13.Rg2 Qxh4
14.Qa8+ Kh7 (82.078.735) 493

10/56 27:31 -1.02 5.a3 Qf3+ 6.Kh2 Qe2+ 7.Kh1 Qe4+
8.Kh2 Re8 9.Kg1 Qe3+ 10.Kg2 Kh7
11.Rf1 Qe4+ 12.Kg3 Qd3+ 13.Rf3 Qxb1
14.Qxe8 Qxb2 15.Qxf7 (796.784.200) 482

11/59 80:10 -1.02 5.a3 Qf5 6.Kh2 Qf4+ 7.Kh1 Qe4+ 8.Kh2 Qe2+
9.Kh1 Re8 10.Rd2 Qe4+ 11.Kh2 Qf4+
12.Kg1 Qe3+ 13.Rf2 Re6 14.Qa8+ Kh7
15.Qxd5 Rg6+ 16.Kf1 Qxh3+ 17.Ke2 Re6+
18.Kd1 (2.351.797.252) 488

12/59 93:32 -0.98 5.a3 Qf3+ 6.Kh2 Qe2+ 7.Kh1 Qe4+
8.Kh2 Re8 9.Rd2 Qf4+ 10.Kg1 Qe3+
11.Rf2 Re5 12.h4 Qg3+ 13.Rg2 Qxh4
14.Qa8+ Kh7 15.Qa7 Re7 (2.700.390.065) 481

best move: a2-a3 time: 94:34.734 min n/s: 482.010 CPU 100.0% n/s(1CPU): 482.010 nodes: 2.735.280.000

I can't reproduce it when I start the analysis over again without any hash content. Result is still 5. a3 but the eval is a little higher.

10/56 5:21 -1.15 5.a3 Qf3+ 6.Kh2 Qf4+ 7.Kh1 Qe4+
8.Kh2 Qe5+ 9.Kh1 Qe6 10.h4 Qe4+
11.Kg1 Rd8 12.Qa7 Qxh4 13.Rf1 Qg5+
14.Kh1 Qh5+ 15.Kg2 (159.860.645) 497

11/58 39:33 -1.15 5.a3 Qf3+ 6.Kh2 Qf4+ 7.Kh1 Qe4+
8.Kh2 Qe5+ 9.Kh1 Qe6 10.Rf1 Qxh3+
11.Kg1 Qg3+ 12.Kh1 Rb6 13.Qe8+ Kh7
14.Qe2 d4 15.a4 c3 16.Qg2 Qxg2+
17.Kxg2 f5 18.a5 (1.187.870.823) 500

best move: a2-a3 time: 49:03.688 min n/s: 486.595 CPU 100.0% n/s(1CPU): 486.595 nodes: 1.432.300.000

Regards, Eelco

[Eelco de Groot]

One more try on the position, and now I think if the output from this build 419 is just a little trustworthy the draw evaluation that engines may have should be revised upwards. To address the question of whether what Rybka played 38.. Qf4+ is a much better move, strangely at the moment it is neither 38... Qf2+ nor Rybka's 38... Qf4+ that is seen as best, 38... Qf4 was on top for a while but at the moment not in the top three.

However the engine used is experimental in eval and only tested in this position. The long search time should partly make up for that but on the negative side the time taken for some of the PVs was short (partly by design) and multi PV is a bit messy because of IID.

This is an excerpt of what I get with Build 419 after Diep's 38. Kh2 in the game, in multi PV mode:


[d]1r4k1/5pp1/7p/3p4/Qpp5/5q1P/PP5K/1R1R4 b - -

Engine: Blueberry Beta 4 DM70 Build 419 (Athlon 2009 MHz, 256 MB)
by F. Letouzey, T. Gaksch, E. de Groot

6 29:49 +0.94 38...Qf4+ 39.Kg2 Rb6 40.Qe8+ Kh7
41.Rf1 Rg6+ 42.Kh1 Qg3 43.Qe2 Qxh3+
44.Qh2 Qd3 (886.001.183) 492

6 42:05 +0.94 38...Qf2+ 39.Kh1 Qf5 40.Kh2 Qf4+
41.Kg2 Rb6 42.Qe8+ Kh7 43.Rf1 Qd2+
44.Kh1 Qd3 45.Qe5 Rf6 46.Rfd1 Qxh3+
47.Qh2 (1.244.891.042) 492

6 40:48 +0.93 38...Qe2+ 39.Kh1 Qe4+ 40.Kh2 Qf4+
41.Kg2 Rb6 42.Qe8+ Kh7 43.Rf1 Qd2+
44.Kh1 Qd3 45.Qe5 Rf6 46.Rfd1 Qxh3+
47.Qh2 Qf3+ 48.Qg2 Qh5+ 49.Qh2
(1.207.065.053) 492
___________________________________________________________

7 44:13 +0.97 38...Qf4+ 39.Kg2 Qe4+ 40.Kh2 Qe5+
41.Kg2 b3 42.a3 c3 43.bxc3 Qxc3
44.Rbc1 Qe3 (1.304.467.249) 474

7 133:50 +0.90 38...Qf2+ 39.Kh1 Qf5 40.Kh2 Qe5+
41.Kg2 Qe4+ 42.Kh2 Qe2+ 43.Kh1 Qe6
44.Kg1 Qxh3 (3.808.961.013) 474

7 95:02 +0.85 38...Qe2+ 39.Kh1 Qe4+ 40.Kh2 Qe5+
41.Kg2 c3 42.bxc3 Qe4+ 43.Kh2 Qe5+
44.Kh1 Qxc3 45.Kh2 f5 46.Rf1 Qd2+
47.Kh1 f4 48.Kg1 Qe3+ 49.Rf2 Qg3+
50.Rg2 Qxh3 (2.717.000.518) 474
___________________________________________________________

8 157:38 +1.00 38...Qf2+ 39.Kh1 Qf5 40.Kh2 b3
41.a3 Qe5+ 42.Kh1 c3 43.bxc3 Qxc3
44.Qg4 Qa5 (4.480.466.252) 474

8 134:17 +0.97 38...Qf4+ 39.Kg2 Qe4+ 40.Kh2 Qe5+
41.Kg2 b3 42.a3 c3 43.bxc3 Qxc3
44.Rbc1 Qe3 (3.821.089.848) 474

8 169:43 +0.95 38...Qe2+ 39.Kh1 Qe4+ 40.Kh2 Qe5+
41.Kg2 b3 42.a3 c3 43.bxc3 Qxc3
44.Re1 Kh7 (4.830.623.230) 474
___________________________________________________________

9 418:28 +1.15 38...b3 39.a3 Qe2+ 40.Kh1 Re8
41.Rd2 Qf3+ 42.Kg1 Qe3+ 43.Rf2 Re6
44.Qa8+ Kh7 45.Qxd5 Rg6+ 46.Kf1 Qxh3+
47.Ke2 Re6+ 48.Kd1 Qg3 49.Rf1 c3 50.Qd8 cxb2
51.Rxb2 Rd6+ (11.348.829.152) 451

9 188:39 +1.00 38...Qf2+ 39.Kh1 Qf5 40.Kh2 Qe5+
41.Kg2 Qe4+ 42.Kh2 b3 43.a3 Qe5+
44.Kh1 c3 45.bxc3 Qxc3 46.Qg4 Qa5 (5.359.735.896) 451

9 379:13 +1.00 38...Qf4+ 39.Kh1 b3 40.a3 Qe4+
41.Kh2 Qe5+ 42.Kh1 Qf5 43.Kh2 c3
44.Rg1 cxb2 45.Qd4 Qc2+ 46.Rg2 (10.345.223.339) 451
___________________________________________________________

9 423:35 +2.05 {!?!} 38...b3 39.a3 Qe2+ 40.Kh1 Re8
41.Rd2 Qf3+ 42.Kg1 Qe3+ 43.Rf2 Re6
44.Qa8+ Kh7 45.Qxd5 Rg6+ 46.Kf1 Qxh3+
47.Ke2 Re6+ 48.Kd1 Qe3 49.Qd2 Qe4 50.Kc1 Qe1+
51.Qxe1 Rxe1+ (11.486.425.726) 451

9 418:28 +1.15 38...b3 39.a3 Qe2+ 40.Kh1 Re8
41.Rd2 Qf3+ 42.Kg1 Qe3+ 43.Rf2 Re6
44.Qa8+ Kh7 45.Qxd5 Rg6+ 46.Kf1 Qxh3+
47.Ke2 Re6+ 48.Kd1 Qg3 49.Rf1 c3 50.Qd8 cxb2
51.Rxb2 Rd6+ (11.348.829.152) 451

9 188:39 +1.00 38...Qf2+ 39.Kh1 Qf5 40.Kh2 Qe5+
41.Kg2 Qe4+ 42.Kh2 b3 43.a3 Qe5+
44.Kh1 c3 45.bxc3 Qxc3 46.Qg4 Qa5 (5.359.735.896) 451
___________________________________________________________

10 789:31 +1.78 38...Qe2+ 39.Kh1 b3 40.Rf1 bxa2
41.Qxa2 Qe4+ 42.Kh2 Qe5+ 43.Kh1 Rb3
44.Qa8+ Kh7 45.Qc8 d4 46.Qg4 f6
47.Rfe1 Qd5+ 48.Qg2 Rxh3+ 49.Kg1 Rh5
50.Qe4+ Rf5 51.Kh2 d3 (21.514.373.072) 455

10 846:02 +1.64 38...Qf2+ 39.Kh1 Qf5 40.Kh2 Qe5+
41.Kh1 Qe4+ 42.Kh2 Qe2+ 43.Kh1 b3
44.Rf1 Qe3 45.a3 Qxh3+ 46.Kg1 Qg4+
47.Kf2 Rc8 (23.128.911.361) 455

10 834:34 +1.02 {a big drop here from iteration 9} 38...b3 39.a3 Qe2+ 40.Kh1 Re8
41.Rd2 Qf3+ 42.Kg1 Qe3+ 43.Rf2 Re6
44.Qa8+ Kh7 45.Qxd5 Rg6+ 46.Kf1 Qxh3+
47.Ke2 Re6+ 48.Kd1 (22.835.031.478) 455
____________________________________________________________

11 846:08 +1.79 38...Qe2+ 39.Kh1 b3 40.Rf1 Qd3
41.Kh2 Qc2+ 42.Kh1 Qe4+ 43.Kh2 Qe5+
44.Kh1 Qg3 45.Qd7 bxa2 46.Rbd1 Rxb2
47.Qxf7+ Kh8 48.Qe8+ Kh7 49.Rxd5 Qxh3+
50.Kg1 (23.131.399.614) 455

Eelco

[Dirt]

11 846:08 +1.79 38...Qe2+ 39.Kh1 b3 40.Rf1 Qd3
41.Kh2 Qc2+ 42.Kh1 Qe4+ 43.Kh2 Qe5+
44.Kh1 Qg3 45.Qd7 bxa2 46.Rbd1 Rxb2
47.Qxf7+ Kh8 48.Qe8+ Kh7 49.Rxd5 Qxh3+
50.Kg1 (23.131.399.614) 455

Eelco

What does the 11 mean here? Surely you searched more than eleven plies in that amount of time. In fact the PV is longer than that.

Anyhow, Stockfish thinks that 46.Ra1 would give white equality, at least as deeply as my poor computer can think in thirty minutes.