couple of test positions from WMCCC Pamplona

[jdart]

I've verified these are the best moves but they're not easy:

Johnny-Junior, bm Nh5:

[D] 8/k4r2/1p3q2/p1p1n3/P3Pnp1/3p1PQ1/1P4PP/R2B1R1K b - -

Rybka-Shredder, bm g3:

[D] r1b2rk1/pp1p2pR/8/1pb2p2/5N2/7Q/qPPB1PPP/6K1 w - -

The preceding Rybka move Nd5 was also very impressive - especially if Rybka had a winning score:
[D] r1b2rk1/pppp2pp/4n1q1/1Bb2p2/7R/2N4Q/PPPB1PPP/4R1K1 w - -

--Jon

[Eelco de Groot]

I've verified these are the best moves but they're not easy:

Johnny-Junior, bm Nh5:

[D] 8/k4r2/1p3q2/p1p1n3/P3Pnp1/3p1PQ1/1P4PP/R2B1R1K b - -

Ancalagon can find the first of these, from Johnny - Junior ! Would not find it in a Blitz game but knowing that Junior was playing with the newest Israeli made Intel hardware, as always, the move could have been fairly deep under the surface. It takes Ancalagon a while to get a big score but the move appears to be winning outright.


8/k4r2/1p3q2/p1p1n3/P3Pnp1/3p1PQ1/1P4PP/R2B1R1K b - -

Engine: Ancalagon 1.3 WS180 Build 171 (Athlon 2009 MHz, 256 MB)
by Romstad, Costalba, Kiiski, de Groot

2.00 0:00 -0.72 1...gxf3 2.gxf3 d2 3.Bc2 (7.692) 32

3.00 0:00 -0.70 1...gxf3 2.Bxf3 Rg7 3.Qf2 Nxf3
4.gxf3 (191.659) 322

4.00 0:00 -0.07 1...gxf3 2.Bxf3 Nxf3 3.Qxf3 Qxb2
4.Qf2 Qxf2 5.Rxf2 (335.750) 412

5.01 0:03 -0.07 1...gxf3 2.Bxf3 Nxf3 3.Qxf3 Qxb2
4.Qf2 Qxf2 5.Rxf2 (1.678.280) 494

6.01 0:04 -0.13 1...gxf3 2.Bxf3 Rg7 3.Qf2 Nxf3
4.Qxf3 Qxb2 5.Qf2 Qxf2 6.Rxf2 (2.137.278) 506

7.01 0:11 -0.33 1...gxf3 2.Bxf3 (6.300.879) 531

7.08 0:12 -0.11 1...Kb7 2.b3 Ne2 3.Bxe2 dxe2 4.Rfe1 gxf3
5.gxf3 Nxf3 (6.822.410) 541

8.01 0:16 -0.11 1...Kb7 2.b3 Ne2 3.Bxe2 dxe2 4.Rfe1 gxf3
5.gxf3 Nxf3 (8.894.755) 555

9.01 0:53 -0.31 1...Kb7 2.Rb1 (31.059.445) 579

9.02 0:54 -0.30 1...c4 (31.575.627) 580

10.01 1:49 +0.07 1...c4 (64.584.605) 590

11.01 1:55 +0.05 1...c4 2.fxg4 Nh5 3.Rxf6 Nxg3+
4.hxg3 Rxf6 5.Kg1 Kb7 6.g5 Rg6 7.Kf2 Rxg5
8.Ke3 Kc6 (68.343.831) 590

11.06 4:43 +0.39 1...Nh5 2.Qe1 g3 3.h3 Nf4 4.b3 Qg5
5.Qd2 Rh7 6.Ra2 Nc6 7.Rg1 Nd4 (162.471.011) 574

12.01 11:42 +0.43 1...Nh5 2.Qe1 g3 3.h3 Nf4 4.b3 Qg5
5.Ra2 Rh7 6.Rb2 Nc6 7.Rg1 Nd4 (425.231.515) 605

13.01 14:33 +0.50 1...Nh5 2.Qe1 g3 3.h3 Nf4 4.b3 Qg5
5.Ra2 Rh7 6.Kg1 Nc6 7.Rd2 Nd4 (536.011.433) 613

14.01 23:36 +0.70 1...Nh5 (863.726.575) 609

15.01 28:13 +1.56 1...Nh5 (1.019.217.384) 601

16.01 37:25 +5.01 1...Nh5 (1.327.486.199) 591

17.01 46:10 +5.31 1...Nh5 2.Qe1 g3 3.h3 Qh4 4.f4 Nxf4
5.Qd2 Rf6 6.Bh5 Nxh5 7.Rf5 Qxe4
8.Rxf6 Nxf6 9.Qg5 Nd5 10.Qxg3 Ne3
11.Qf2 (1.639.403.196) 591

best move: Nf4-h5 time: 53:20.922 min n/s: 590.145 nodes: 1.889.010.248

Eelco

[diep]

I've verified these are the best moves but they're not easy:

Johnny-Junior, bm Nh5:

[D] 8/k4r2/1p3q2/p1p1n3/P3Pnp1/3p1PQ1/1P4PP/R2B1R1K b - -

Rybka-Shredder, bm g3:

[D] r1b2rk1/pp1p2pR/8/1pb2p2/5N2/7Q/qPPB1PPP/6K1 w - -

The preceding Rybka move Nd5 was also very impressive - especially if Rybka had a winning score:
[D] r1b2rk1/pppp2pp/4n1q1/1Bb2p2/7R/2N4Q/PPPB1PPP/4R1K1 w - -

--Jon

Good post Jon,

Even though it is just tactical positions. Mating extensions really help there. The problem is not so much finishing such positions by finding tactical shots (mating extensions really helpful), the real problem is getting those positions of course.

The first 2 positions are blitz level for Diep, especially at 8 core hardware, even though it has no mating extensions at all. Probably lucky with its kingsafety. It has trouble finding Nd5 though. Nd5 is a real good shot at this position as the alternative would be going into some sort of endgame that is good for white (and probably winning) yet Nd5 is directly game over of course.

What time and/or plydepth does rybka3 take to find Nd5 in that final position? don't forget to clear hashtable and permanent learning before trying. Don't analyze first in that position to figure it out. Interesting is knowing what time it takes a cleared hashtable to find this.

Note i don't really agree with the score condition you mention. It is not about the score you display as that score you could add up a constant of 10 pawns or something, to always look happy.

It is about the move you play.

[jdart]

> Note i don't really agree with the score condition you mention.

What I meant here is that there are some positions where you can play the right move for the wrong reason, just because it scores a fraction of a pawn higher than alternatives. Getting a high plus score here though means the program sees not just the first move but the significant parts of the PV that leads to a win. That's very difficult in this case.

[Spock]

What time and/or plydepth does rybka3 take to find Nd5 in that final position? don't forget to clear hashtable and permanent learning before trying. Don't analyze first in that position to figure it out. Interesting is knowing what time it takes a cleared hashtable to find this.

Rybka 3 on my Intel QX6700, just over 10 mins
On an Octal or better, significantly less probably

11.00 0:01 +0.19 1.Bd3 c6 2.Rh5 d5 3.Ne2 Be7 4.Rxf5 Rxf5 5.Qxf5 Qxf5 6.Bxf5 Nc7 7.Bd3 Bd7 8.Nd4 (224.653) 186
12.01 0:01 +0.19 1.Bd3 c6 2.Rh5 d5 3.Ne2 Be7 4.Rxf5 Rxf5 5.Qxf5 Qxf5 6.Bxf5 Nc7 7.Bd3 Bd7 8.Nd4 (291.745) 182
13.01 0:06 +0.19 1.Bd3 h6 2.Re5 b6 (1.364.834) 200
14.01 0:10 +0.19 1.Bd3 h6 2.Re5 b6 3.Rxf5 Rxf5 (2.111.105) 208
15.01 0:45 +0.29 1.Bd3 c6 2.b4 Bb6 3.Rh5 d5 4.Rxf5 Rxf5 5.Qxf5 Qxf5 6.Bxf5 Nc7 7.Bd3 Ne6 8.Na4 g6 9.c4 d4 10.g3 Bc7 (10.000.274) 225
16.01 0:55 +0.12 1.Bd3 c6 2.Na4 d6 3.b4 Bd4 4.Rxe6 Bxf2+ 5.Kxf2 Bxe6 6.Nc3 Rae8 7.Bf4 Rf7 8.Bxd6 Qf6 9.Ne2 g6 10.Qg3 Bxa2 11.Bc5 (12.061.493) 224
17.01 1:36 +0.20 1.Bd3 c6 2.Na4 d6 3.b4 Bd4 4.Rxe6 Bxf2+ 5.Kxf2 Bxe6 6.Nc3 Rae8 7.a3 h6 8.Ne2 Qf6 9.Bc3 Qg6 10.Qg3 Qxg3+ 11.Nxg3 (21.416.394) 227
18.01 2:04 +0.20 1.Bd3 c6 2.Na4 d6 3.b4 Bd4 4.Rxe6 Bxf2+ 5.Kxf2 Bxe6 6.Nc3 Rae8 7.a3 h6 8.Ne2 Qf6 9.Bc3 Qg6 10.Qg3 Qxg3+ 11.Nxg3 (26.949.479) 220
19.01 6:50 +0.09 1.Bd3 c6 2.Na4 d6 3.Bc3 Bb6 4.Nxb6 axb6 5.a4 h6 6.Rb4 Nc5 7.b3 Ne4 8.Ba1 b5 9.axb5 Qg5 (93.557.753) 233
19.04 10:43 +0.40++ 1.Nd5 (142.964.526) 227

[tano-urayoan]

What time and/or plydepth does rybka3 take to find Nd5 in that final position? don't forget to clear hashtable and permanent learning before trying. Don't analyze first in that position to figure it out. Interesting is knowing what time it takes a cleared hashtable to find this.

If I remember correctly this game was played by cluster Rybka.

[Gian-Carlo Pascutto]

Yes. From the comments of Hans van der Zijden the Cluster hadn't actually seen the win and still had a 0 score for that move.

[diep]

What time and/or plydepth does rybka3 take to find Nd5 in that final position? don't forget to clear hashtable and permanent learning before trying. Don't analyze first in that position to figure it out. Interesting is knowing what time it takes a cleared hashtable to find this.

Rybka 3 on my Intel QX6700, just over 10 mins
On an Octal or better, significantly less probably

11.00 0:01 +0.19 1.Bd3 c6 2.Rh5 d5 3.Ne2 Be7 4.Rxf5 Rxf5 5.Qxf5 Qxf5 6.Bxf5 Nc7 7.Bd3 Bd7 8.Nd4 (224.653) 186
12.01 0:01 +0.19 1.Bd3 c6 2.Rh5 d5 3.Ne2 Be7 4.Rxf5 Rxf5 5.Qxf5 Qxf5 6.Bxf5 Nc7 7.Bd3 Bd7 8.Nd4 (291.745) 182
13.01 0:06 +0.19 1.Bd3 h6 2.Re5 b6 (1.364.834) 200
14.01 0:10 +0.19 1.Bd3 h6 2.Re5 b6 3.Rxf5 Rxf5 (2.111.105) 208
15.01 0:45 +0.29 1.Bd3 c6 2.b4 Bb6 3.Rh5 d5 4.Rxf5 Rxf5 5.Qxf5 Qxf5 6.Bxf5 Nc7 7.Bd3 Ne6 8.Na4 g6 9.c4 d4 10.g3 Bc7 (10.000.274) 225
16.01 0:55 +0.12 1.Bd3 c6 2.Na4 d6 3.b4 Bd4 4.Rxe6 Bxf2+ 5.Kxf2 Bxe6 6.Nc3 Rae8 7.Bf4 Rf7 8.Bxd6 Qf6 9.Ne2 g6 10.Qg3 Bxa2 11.Bc5 (12.061.493) 224
17.01 1:36 +0.20 1.Bd3 c6 2.Na4 d6 3.b4 Bd4 4.Rxe6 Bxf2+ 5.Kxf2 Bxe6 6.Nc3 Rae8 7.a3 h6 8.Ne2 Qf6 9.Bc3 Qg6 10.Qg3 Qxg3+ 11.Nxg3 (21.416.394) 227
18.01 2:04 +0.20 1.Bd3 c6 2.Na4 d6 3.b4 Bd4 4.Rxe6 Bxf2+ 5.Kxf2 Bxe6 6.Nc3 Rae8 7.a3 h6 8.Ne2 Qf6 9.Bc3 Qg6 10.Qg3 Qxg3+ 11.Nxg3 (26.949.479) 220
19.01 6:50 +0.09 1.Bd3 c6 2.Na4 d6 3.Bc3 Bb6 4.Nxb6 axb6 5.a4 h6 6.Rb4 Nc5 7.b3 Ne4 8.Ba1 b5 9.axb5 Qg5 (93.557.753) 233
19.04 10:43 +0.40++ 1.Nd5 (142.964.526) 227

A score drop at 19 ply for Bd3 it seems. So all it has to find is this drop in score at 19 ply for Bd3 in order to get time extensions anyway i suppose.

Already within a game with a filled hashtable this should be no problem to find at a fast quadcore box for rybka 3 at a time control of 60 in 2.

In Diep i'm seeing at 90 30 time control already that shots similar like this with filled hashtable get found within 1 minute whereas with cleaned hashtable it is 5 minutes.

Vincent

[diep]

> Note i don't really agree with the score condition you mention.

What I meant here is that there are some positions where you can play the right move for the wrong reason, just because it scores a fraction of a pawn higher than alternatives. Getting a high plus score here though means the program sees not just the first move but the significant parts of the PV that leads to a win. That's very difficult in this case.

Thanks for trying to make Rybka look like a fool, as it plays every move for the wrong reason, but does that count?

Vincent

[Eelco de Groot]

What time and/or plydepth does rybka3 take to find Nd5 in that final position? don't forget to clear hashtable and permanent learning before trying. Don't analyze first in that position to figure it out. Interesting is knowing what time it takes a cleared hashtable to find this.

If I remember correctly this game was played by cluster Rybka.

If I had a cluster with so many nodes as Lukas has, I think it would be worthwhile to try to skip ahead with one or a few of the cluster-nodes; when at move n, try to anticipate your own move n, opponent's reply to move n and then your first choice for best move n+1 and try to start early on your own second-best n+1 move. If you can anticipate that Bd3 will be your first choice, you could deepen the search for the second or other third, fourth alternative moves and have a good chance of finding Nd5, if your anticipated moves turn out to be correct. With so many nodes, you can risk a few of them going on a wild goose chase if once in a while you are correct. Not being able to exchange hash results in a cluster does not matter very much in that case, the node that has "jumped ahead" will be able to look deeply enough on its own, freeing other roving nodes for the same process. I don't know if that is actually happening, but it would be thinkable. If the master node is still displaying a 0.0 score for Nd5 it is not certain that there is not another node with a better evaluation, but that the master is just maintaining its own search, checking the alternatives against this conservative 0.0 score, because that is something the other nodes can't do as well as the master. The master has to search as broadly as possible also, like Vas said, because it has to act as a failsafe in case the communications to the other cluster-nodes break down

Just pondering here a bit about that 0.0 score, in theory it could be more of a lower bound, at least that is what I would like to design, for a cluster type search and masternode type of output. But maybe Vas has some logs that could tell you more, some of the logs where made publicly available by Vas for those interested

Eelco