Komodo 9 Announcement

[mclane]

Larry, is it true that you and Mark are obsessed beyond imagination with Stockfish and that you cleverly stole some code from it? Ofcourse, you'll deny like the deceitful grandmaster you are.

Do you still beat your wife ? Of course you deny it.

[Astatos]

Whenever Stockfish or earlier Houdini or Critter where doing a step forward Komodo follows in short course. Of course all this is totally accidental, nobody ever thought that might be a correlation. I am guessing that whenever Stockfish hit the wall, by sheer coincidence Komodo will lose steam also.

The world is always full of conspiracy theories, most of which are utter nonsense.

It's quite clear to me that Stockfish and Komodo play differently.

They were clear to you many things that happened not to be the case...

[Werewolf]

Larry, is it true that you and Mark are obsessed beyond imagination with Stockfish and that you cleverly stole some code from it? Ofcourse, you'll deny like the deceitful grandmaster you are.

Vas quit it.
Vas go away unless you are going to improve your own program and leave Larry alone.

Vas?! Not Vas...Rajlich?

[APassionForCriminalJustic]

Larry, is it true that you and Mark are obsessed beyond imagination with Stockfish and that you cleverly stole some code from it? Ofcourse, you'll deny like the deceitful grandmaster you are.

Taking code directly from SF wouldn't make sense, as the programs are too different for copy and paste to help. We certainly have felt free to use general ideas from Stockfish; I think one that was a definite help to us was the concept of "improving" moves. But we don't pay close attention to every development in Stockfish, because most of them simply don't apply to us. They may be tuning some parameter (we tune our own parameters, never relying on anyone else for this), or rewriting code (again, not applicable to us), or even changing something to make it more like the way we already do it (that happens quite a bit, but I'm not accusing anyone of decompiling Komodo). We have tried making the search and/or parts of the eval very similar to Stockfish, but that always just makes us weaker.
Frankly, we cannot explain why Stockfish outsearches Komodo 9 by nearly four ply at normal blitz levels, even on single core and despite Komodo 9 outsearching Komodo 8 by a ply or more. One ply difference is easily explained by some extensions we do, but the rest is a total mystery. If we try to duplicate the Stockfish search as much as is practical in Komodo, a large depth difference remains. Despite our inability to account for this, Komodo 9 appears to be as strong as or stronger than Stockfish.
So while we don't mind using new ideas that appear in Stockfish, in practice so far this has been of only marginal value to Komodo.

Larry, do you guys have any test runs between the latest development version of Stockfish versus Komodo 9? We all know that the latest development version is about or close to 25 ELO over Stockfish 6 at single core - and the difference is nearly 60+ ELO when the cores equal approximately 16 or greater.

I'm doing a run now against the latest SF on single core, at 2' + 1". As you would expect, we're running a bit behind so far; I'll post final result later. Which would win at a "normal" time control remains to be seen. As already reported, on 16 cores we won by 14 elo against current SF, but of course few people have 16 cores.

Okay cool. That is a great ELO number with the 16-core test. I am looking forward to the single core test result.

[lkaufman]

Final result after 2000 games at 2' + 1" on single core against April 18 Stockfish was a loss by 15 elo points. However I'm currently running the same pairing at 10 times the time limit (20' + 10") and so far Komodo 9 leads by 101 to 89, which is +22 elo, but still within error margin. So it appears that K9 passes SF April 18 somewhere between 2 and twenty minutes (plus inc), although more games are needed to have confidence in this claim.

[Uri Blass]

Final result after 2000 games at 2' + 1" on single core against April 18 Stockfish was a loss by 15 elo points. However I'm currently running the same pairing at 10 times the time limit (20' + 10") and so far Komodo 9 leads by 101 to 89, which is +22 elo, but still within error margin. So it appears that K9 passes SF April 18 somewhere between 2 and twenty minutes (plus inc), although more games are needed to have confidence in this claim.

I think that it is possible that the last change from 26.4 improved stockfish more at long time control.

If I understand correctly the change means that stockfish does less LMR
because cmh <= 0 && h < 0 happens in less times than cmh+h<0 so stockfish does less times extra ply in LMR

Here is the change from the stockfish framework

Author: VoyagerOne
Date: Sun Apr 26 20:12:25 2015 +0100
Timestamp: 1430075545

Change extra ply LMR condition to: cmh <= 0 && hist < 0

Extra ply LMR condition is now cmh <= 0 && h < 0
Instead of cmh + h < 0

STC:
LLR: 2.96 (-2.94,2.94) [-1.50,4.50]
Total: 55210 W: 10812 L: 10557 D: 33841

LTC:
LLR: 2.95 (-2.94,2.94) [0.00,4.00]
Total: 13212 W: 2239 L: 2045 D: 8928

Bench: 8420865

Resolves #339

[APassionForCriminalJustic]

Final result after 2000 games at 2' + 1" on single core against April 18 Stockfish was a loss by 15 elo points. However I'm currently running the same pairing at 10 times the time limit (20' + 10") and so far Komodo 9 leads by 101 to 89, which is +22 elo, but still within error margin. So it appears that K9 passes SF April 18 somewhere between 2 and twenty minutes (plus inc), although more games are needed to have confidence in this claim.

I think that it is possible that the last change from 26.4 improved stockfish more at long time control.

If I understand correctly the change means that stockfish does less LMR
because cmh <= 0 && h < 0 happens in less times than cmh+h<0 so stockfish does less times extra ply in LMR

Here is the change from the stockfish framework

Author: VoyagerOne
Date: Sun Apr 26 20:12:25 2015 +0100
Timestamp: 1430075545

Change extra ply LMR condition to: cmh <= 0 && hist < 0

Extra ply LMR condition is now cmh <= 0 && h < 0
Instead of cmh + h < 0

STC:
LLR: 2.96 (-2.94,2.94) [-1.50,4.50]
Total: 55210 W: 10812 L: 10557 D: 33841

LTC:
LLR: 2.95 (-2.94,2.94) [0.00,4.00]
Total: 13212 W: 2239 L: 2045 D: 8928

Bench: 8420865

Resolves #339

But a change that would account for 22 ELO? If Larry's test here is true - error of margin of course - then perhaps Komodo really is the king of LTC. If anything the ELO turnaround for Komodo with increased time limits is absolutely insane.

[lkaufman]

Final result after 2000 games at 2' + 1" on single core against April 18 Stockfish was a loss by 15 elo points. However I'm currently running the same pairing at 10 times the time limit (20' + 10") and so far Komodo 9 leads by 101 to 89, which is +22 elo, but still within error margin. So it appears that K9 passes SF April 18 somewhere between 2 and twenty minutes (plus inc), although more games are needed to have confidence in this claim.

I think that it is possible that the last change from 26.4 improved stockfish more at long time control.

If I understand correctly the change means that stockfish does less LMR
because cmh <= 0 && h < 0 happens in less times than cmh+h<0 so stockfish does less times extra ply in LMR

Here is the change from the stockfish framework

Author: VoyagerOne
Date: Sun Apr 26 20:12:25 2015 +0100
Timestamp: 1430075545

Change extra ply LMR condition to: cmh <= 0 && hist < 0

Extra ply LMR condition is now cmh <= 0 && h < 0
Instead of cmh + h < 0

STC:
LLR: 2.96 (-2.94,2.94) [-1.50,4.50]
Total: 55210 W: 10812 L: 10557 D: 33841

LTC:
LLR: 2.95 (-2.94,2.94) [0.00,4.00]
Total: 13212 W: 2239 L: 2045 D: 8928

Bench: 8420865

Resolves #339

But a change that would account for 22 ELO? If Larry's test here is true - error of margin of course - then perhaps Komodo really is the king of LTC. If anything the ELO turnaround for Komodo with increased time limits is absolutely insane.

After 250 games K9's lead at 20' +10" over SFApril 18 has increased to 24 elo. The results shown above for SF April 26 when combined show a gain of between 2 and 3 elo, so the results imply a lead of 21 elo for K9 over SF April 26. Still the margin of error is more than that though.

[Leto]

Komodo 9 does go through depths quicker thank K8, I like that a lot.

In this position on my 12 core machine with 5 piece syzygy installed K8 is at depth 27 at the 1 minute mark, and depth 30 at the 2 minute mark.

2r5/R1p2r1p/3pR1n1/2pN3k/2P2p1P/1P3P2/5KP1/8 b - - 0 43

Komodo 9 is at depth 30 at the 1 minute mark, and depth 32 at the 2 minute mark.

For analysis I prefer engines to go through depths quicker so that I see more activity on the analysis pane rather than staring at it waiting for it to produce results.

In comparison the April 26 2015 Stockfish dev version gets to depth 35 at the 1 minute mark, and depth 37 at the 2 minute mark. So in future releases I'd like to see Komodo continue to increase the speed at which it goes through depths.

I also like the addition of options to change Komodo's pruning, this might be helpful under certain situations such as correspondence play or solving tactical positions. Also the persistent hash feature is very welcome.

I will begin testing for CEGT now on my 12 core machine.

[Damir]

Hi Larry

Can either you or Marc put more parameters in Komodo, like pawn, Knight, Bishop, Rook and Queen value, like Sugar 5.1. If I remember this correctly one of the older version of Komodo or was it Doch had these parameters to tune and many more. Can you please put them back in Komodo 9 or later versions ?