Houdini 5 dev in TCEC

Discussion of anything and everything relating to chess playing software and machines.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
User avatar
hgm
Posts: 23772
Joined: Fri Mar 10, 2006 9:06 am
Location: Amsterdam
Full name: H G Muller
Contact:

Re: Houdini 5 dev in TCEC

Post by hgm » Sun Sep 11, 2016 8:29 pm

Friendly moderation request:

Please stay on topic in this thread, which is about TCEC and Houdini. The dissatisfaction of some with extremely long threads has been noticed now, and the problem will be addressed in a more appropriate place, where we will seek a solution to make everyone happy.

So please do not dwell on it here.

User avatar
lucasart
Posts: 3041
Joined: Mon May 31, 2010 11:29 am
Full name: lucasart
Contact:

Re: Houdini 5 dev in TCEC

Post by lucasart » Sun Sep 11, 2016 11:00 pm

Jouni wrote:Robert has really improved engine. In CCRL, CEGT and IPON Houdini 4 was 119, 154 and 137 points below Komodo 10.1. So progress has been around 140 ELO!
You're jumping to conclusions too quickly. You need to account for error bars. IMO, only rating lists will tell if Houdini is on par with Komodo 10.1, not TCEC.
Theory and practice sometimes clash. And when that happens, theory loses. Every single time.

Lyudmil Tsvetkov
Posts: 6052
Joined: Tue Jun 12, 2012 10:41 am

Re: Houdini 5 dev in TCEC

Post by Lyudmil Tsvetkov » Mon Sep 12, 2016 5:55 am

For people that do not believe new Houdini has thrashing-good evaluation, here the opening position from game 28, Houdini-Rybka, of stage 3 of TCEC:

[d]rnbq1rk1/1p3pbp/p2p1np1/3Pp3/4P3/2N1BP2/PP1Q2PP/R3KBNR w KQ - 0 10

Believe it or not, Houdini plays here g4!, with an eval of 80cps, while Rybka is at 40cps.

Do you see the major difference in eval?

g4, after f3, is precisely the way to counter effeciently the mainline King's Indian, the only way, but quite convincing. The purpose of g4 is of course to prevent black f7-f5 with building a long chain pointing toward the white king.

With g4 above, Houdini leaves its king in the center, with no pawn shelter, only to counter f5.

I also noticed goodish KID and related structures treatment in a range of other Houdini games, so one could say H 5 will be the first engine to start getting grasp of the King's Indian.

How many engines would play g4 above?

peter
Posts: 1786
Joined: Sat Feb 16, 2008 6:38 am
Full name: Peter Martan

Re: Houdini 5 dev in TCEC

Post by peter » Mon Sep 12, 2016 6:44 am

Hi Lyudmil!
Lyudmil Tsvetkov wrote: How many engines would play g4 above?
Tried SF only, here in 4 MV mode and depth 31:
rnbq1rk1/1p3pbp/p2p1np1/3Pp3/4P3/2N1BP2/PP1Q2PP/R3KBNR w KQ - 0 1

Analysis by Stockfish 030916 64 POPCNT:

1. +/= (0.70): 10.Ld3 Sbd7 11.Sge2 Sc5 12.Lc2 a5 13.0-0 Ld7 14.a4 Tc8 15.Sc1 Sh5 16.Sd3 Sxd3 17.Lxd3 f5 18.exf5 Lxf5 19.Tac1 Lxd3 20.Dxd3 Sf4 21.Dd2 Tc4 22.b3 Tb4 23.Tb1 Dd7 24.Tfc1 Tc8 25.Sb5 Txc1+ 26.Txc1 Txb3 27.Lxf4 exf4 28.Dxf4 Td3 29.Sxd6

2. +/= (0.65): 10.g4 h5 11.g5 Se8 12.Sh3 Sc7 13.Sf2 Ld7 14.a4 a5 15.Le2 Sca6 16.Sb5 Lxb5 17.Lxb5 Sd7 18.0-0 Sdc5 19.Ta3 b6 20.Sd3 De7 21.Kh1 f5 22.gxf6 Dxf6 23.Sxc5 Sxc5 24.Lg5 Df7 25.Lc6

3. +/= (0.59): 10.Sge2 Sbd7 11.g4 h5 12.g5 Se8 13.Sc1 a5 14.Le2 Sc5 15.Sd3 b6 16.Tb1 La6 17.0-0 Sc7 18.b3 Sxd3 19.Lxd3 Lxd3 20.Dxd3 Sa6 21.Sb5 Sc5 22.De2 De7 23.Tbc1 Dd7 24.a4 Tab8 25.Tc3 Tfc8 26.Tfc1

4. +/= (0.41): 10.Tc1 b5 11.g4 h5 12.g5 Sh7 13.h4 f6 14.b4 Ld7 15.a3 fxg5 16.hxg5 Tf7 17.Th2 Lf8 18.Ld3 Le8 19.Tb1 Le7 20.Tg2 Ld7 21.Tc1 Df8 22.Kd1 De8 23.Tb1 Dd8 24.Kc1 Le8 25.Kb2 Sd7 26.Tc1 Sb6 27.Lxb6

There are some good moves at this certain ply, g4 doesn't have to come at once as 10th, time enough for it later on too.
E.g. 10. 0-0-0 isn't bad neither, is it? With main cont. ...Nd7 11.g4 does work as well, doesn't it?
So you can't give this position for testing without looking at the output- lines at least more than just one ply deep.
In the line after 10.Nge2 above you see g4 as 11. too

Played 10.Nge2 Nd7 (main cont. after Nge2):
r1bq1rk1/1p1n1pbp/p2p1np1/3Pp3/4P3/2N1BP2/PP1QN1PP/R3KB1R w KQ - 0 1

Analysis by Stockfish 030916 64 POPCNT:

11.g4 h5 12.g5 Se8 13.Sc1 Sc7 14.Sd3 a5 15.Le2 Tb8 16.a4 b6 17.b4 Sa6 18.Sa2 f5 19.gxf6 Dxf6 20.0-0 axb4 21.Saxb4 Sac5 22.Tab1 Sxd3 23.Lxd3 Sc5 24.Sc6 Ta8 25.Lg5 Df7 26.Lc2 Lh3 27.Tf2 Lf6 28.Txb6 Sxa4
+/= (0.64) Tiefe: 31/45 00:01:11 1266MN
I'd say not until this 11th, if ever, g4 is a more or less "best move", still difficult to prove it to be the one and only one.

Given 10.Nge2 b5 (also played sometimes):
rnbq1rk1/5pbp/p2p1np1/1p1Pp3/4P3/2N1BP2/PP1QN1PP/R3KB1R w KQ - 0 1

Analysis by Stockfish 030916 64 POPCNT:

11.g4 h5 12.gxh5 Sxh5 13.Sg3 Sf4 14.h4 Sd7 15.h5 Sb6 16.b3 Ld7 17.Sd1 Tc8 18.Dh2 Te8 19.Tb1 Sa8 20.Tb2 Kf8 21.Td2 Ke7 22.hxg6 Th8 23.Dg1 fxg6 24.Se2 Txh1 25.Dxh1 g5 26.Sg3 Kf7 27.Sh5 Dg8 28.Th2 Tc7 29.Sxg7 Dxg7 30.Dg1
+/- (0.75) Tiefe: 32/60 00:02:04 2152MN
Peter.

Henk
Posts: 5832
Joined: Mon May 27, 2013 8:31 am

Re: Houdini 5 dev in TCEC

Post by Henk » Mon Sep 12, 2016 7:53 am

Lyudmil Tsvetkov wrote:For people that do not believe new Houdini has thrashing-good evaluation, here the opening position from game 28, Houdini-Rybka, of stage 3 of TCEC:

[d]rnbq1rk1/1p3pbp/p2p1np1/3Pp3/4P3/2N1BP2/PP1Q2PP/R3KBNR w KQ - 0 10

Believe it or not, Houdini plays here g4!, with an eval of 80cps, while Rybka is at 40cps.

Do you see the major difference in eval?

g4, after f3, is precisely the way to counter effeciently the mainline King's Indian, the only way, but quite convincing. The purpose of g4 is of course to prevent black f7-f5 with building a long chain pointing toward the white king.

With g4 above, Houdini leaves its king in the center, with no pawn shelter, only to counter f5.

I also noticed goodish KID and related structures treatment in a range of other Houdini games, so one could say H 5 will be the first engine to start getting grasp of the King's Indian.

How many engines would play g4 above?
Maybe that's the reason not to castle so soon with black. Otherwise black could play h5 and g4 would be a bad move.

By the way after 1 d4 I would play g6 for Nf6 is hindering f5.

Lyudmil Tsvetkov
Posts: 6052
Joined: Tue Jun 12, 2012 10:41 am

Re: Houdini 5 dev in TCEC

Post by Lyudmil Tsvetkov » Mon Sep 12, 2016 9:49 am

peter wrote:Hi Lyudmil!
Lyudmil Tsvetkov wrote: How many engines would play g4 above?
Tried SF only, here in 4 MV mode and depth 31:
rnbq1rk1/1p3pbp/p2p1np1/3Pp3/4P3/2N1BP2/PP1Q2PP/R3KBNR w KQ - 0 1

Analysis by Stockfish 030916 64 POPCNT:

1. +/= (0.70): 10.Ld3 Sbd7 11.Sge2 Sc5 12.Lc2 a5 13.0-0 Ld7 14.a4 Tc8 15.Sc1 Sh5 16.Sd3 Sxd3 17.Lxd3 f5 18.exf5 Lxf5 19.Tac1 Lxd3 20.Dxd3 Sf4 21.Dd2 Tc4 22.b3 Tb4 23.Tb1 Dd7 24.Tfc1 Tc8 25.Sb5 Txc1+ 26.Txc1 Txb3 27.Lxf4 exf4 28.Dxf4 Td3 29.Sxd6

2. +/= (0.65): 10.g4 h5 11.g5 Se8 12.Sh3 Sc7 13.Sf2 Ld7 14.a4 a5 15.Le2 Sca6 16.Sb5 Lxb5 17.Lxb5 Sd7 18.0-0 Sdc5 19.Ta3 b6 20.Sd3 De7 21.Kh1 f5 22.gxf6 Dxf6 23.Sxc5 Sxc5 24.Lg5 Df7 25.Lc6

3. +/= (0.59): 10.Sge2 Sbd7 11.g4 h5 12.g5 Se8 13.Sc1 a5 14.Le2 Sc5 15.Sd3 b6 16.Tb1 La6 17.0-0 Sc7 18.b3 Sxd3 19.Lxd3 Lxd3 20.Dxd3 Sa6 21.Sb5 Sc5 22.De2 De7 23.Tbc1 Dd7 24.a4 Tab8 25.Tc3 Tfc8 26.Tfc1

4. +/= (0.41): 10.Tc1 b5 11.g4 h5 12.g5 Sh7 13.h4 f6 14.b4 Ld7 15.a3 fxg5 16.hxg5 Tf7 17.Th2 Lf8 18.Ld3 Le8 19.Tb1 Le7 20.Tg2 Ld7 21.Tc1 Df8 22.Kd1 De8 23.Tb1 Dd8 24.Kc1 Le8 25.Kb2 Sd7 26.Tc1 Sb6 27.Lxb6

There are some good moves at this certain ply, g4 doesn't have to come at once as 10th, time enough for it later on too.
E.g. 10. 0-0-0 isn't bad neither, is it? With main cont. ...Nd7 11.g4 does work as well, doesn't it?
So you can't give this position for testing without looking at the output- lines at least more than just one ply deep.
In the line after 10.Nge2 above you see g4 as 11. too

Played 10.Nge2 Nd7 (main cont. after Nge2):
r1bq1rk1/1p1n1pbp/p2p1np1/3Pp3/4P3/2N1BP2/PP1QN1PP/R3KB1R w KQ - 0 1

Analysis by Stockfish 030916 64 POPCNT:

11.g4 h5 12.g5 Se8 13.Sc1 Sc7 14.Sd3 a5 15.Le2 Tb8 16.a4 b6 17.b4 Sa6 18.Sa2 f5 19.gxf6 Dxf6 20.0-0 axb4 21.Saxb4 Sac5 22.Tab1 Sxd3 23.Lxd3 Sc5 24.Sc6 Ta8 25.Lg5 Df7 26.Lc2 Lh3 27.Tf2 Lf6 28.Txb6 Sxa4
+/= (0.64) Tiefe: 31/45 00:01:11 1266MN
I'd say not until this 11th, if ever, g4 is a more or less "best move", still difficult to prove it to be the one and only one.

Given 10.Nge2 b5 (also played sometimes):
rnbq1rk1/5pbp/p2p1np1/1p1Pp3/4P3/2N1BP2/PP1QN1PP/R3KB1R w KQ - 0 1

Analysis by Stockfish 030916 64 POPCNT:

11.g4 h5 12.gxh5 Sxh5 13.Sg3 Sf4 14.h4 Sd7 15.h5 Sb6 16.b3 Ld7 17.Sd1 Tc8 18.Dh2 Te8 19.Tb1 Sa8 20.Tb2 Kf8 21.Td2 Ke7 22.hxg6 Th8 23.Dg1 fxg6 24.Se2 Txh1 25.Dxh1 g5 26.Sg3 Kf7 27.Sh5 Dg8 28.Th2 Tc7 29.Sxg7 Dxg7 30.Dg1
+/- (0.75) Tiefe: 32/60 00:02:04 2152MN
Bd3 is not so good because of Ne8, followed by f5, and black is much better in comparison to g4.

I checked SF 7 and Komodo 10.1 at low depths, and both would like to play Bd3, eval for SF some 55cps, for Komodo some 50cps. Latest SF migth be a bit better, but it still shows lower score than Houdini and seemingly likes enough also Bd3 and castling short, which is not the way to play.

But you are rigth, I did not pick the best possible position, as white has large advantage, c6 for black, and then cd5 cd5, is simply a bad line.

But if you pick just about any other position from TCEC, you will see that Houdini indeed shows much more realistic scores than both Komodo and SF.

Btw., if you look at TCEC games and Houdini evals, in most games SF and Komodo show +40cps, while Houdini + 80cps, there was a rule about checking all opening positions with the 2 top engines to ensure score is within reasonable bounds (was it 40cps ?), maybe now they should discard half of their positions, since another top engine claims advantage is twice as high?

User avatar
Harvey Williamson
Posts: 1820
Joined: Sun May 25, 2008 9:12 pm
Location: Media City, UK
Contact:

Re: Houdini 5 dev in TCEC

Post by Harvey Williamson » Mon Sep 12, 2016 1:43 pm

MikeB wrote:
mclane wrote:
Lyudmil Tsvetkov wrote:I forgot to say that H still keeps its good attacking style from the past, above features only enriching it, while making it significantly more positional

PS. Is not it possible to delete that very-stupid benchmark-new-smartphones thread? It has already reached 2000 posts, fully unreadable, dating from a couple of years, fully unrelated to the forum content, and, most importantly, always appearing/being renewed(maybe on purpose) to make difficult normally reading other threads???
What is your problem with it ?
It relates hardware with each other. And gives information to computer chess nerds which hardware is best for chess.

That seems a proper target for a computerchess forum IMO.
+1 It so so easy to ignore unwelcome posts/topics. Making a mountain out of a molehill. Seriously doubt that folks are doing that purposely to annoy you. I guess that if one has an ego-centric view of the world it easy to come to that conclusion, i.e., everything that annoys you in life is being done on purpose. Hope your day gets better.
Agreed just switch to flat view if you do not like the way thread view shows the thread.

And if people do not like a particular thread please do not hijack another thread to complain about it.

Post Reply