Komodo 9.2

[FWCC]

Hello Larry,it seems that Komodo 9.2 picks up the endgame tablebases at a faster rate and with more pieces on the board.In other words I get TB hits much sooner with the board almost "full of pieces" and the hits go deeper than most engines.Can you explain?




FWCC

[lkaufman]

Hello Larry,it seems that Komodo 9.2 picks up the endgame tablebases at a faster rate and with more pieces on the board.In other words I get TB hits much sooner with the board almost "full of pieces" and the hits go deeper than most engines.Can you explain?

No, we haven't worked on tablebases for a while. What earlier version of Komodo are you comparing 9.2 to when you say the above? If you are just comparing to other engines it would probably be due to the depth at which we probe, which I think is lower than others.




FWCC

[Leto]

I sent 1750 games with two versions of Komodo 9.2 running with 12 cores to CEGT 40/4, one with the default 15 contempt setting, and one with contempt set to 0. After 1000 games the default 15 contempt version scored 68.6% after 1000 games, the contempt 0 version scored 72.1% after 750 games.

The average opposition for the opponents was around 141 elo lower than Komodo 9.1 x64 12CPU. Shouldn't the default version score higher than the contempt 0 version?

my thread with the scores. The CEGT 40/4 rating list should be updated soon:
http://www.talkchess.com/forum/viewtopic.php?t=57353

[Leto]

I sent 1750 games with two versions of Komodo 9.2 running with 12 cores to CEGT 40/4, one with the default 15 contempt setting, and one with contempt set to 0. After 1000 games the default 15 contempt version scored 68.6% after 1000 games, the contempt 0 version scored 72.1% after 750 games.

The average opposition for the opponents was around 141 elo lower than Komodo 9.1 x64 12CPU. Shouldn't the default version score higher than the contempt 0 version?

my thread with the scores. The CEGT 40/4 rating list should be updated soon:
http://www.talkchess.com/forum/viewtopic.php?t=57353

Rating list updated. The default contempt 15 version after 1000 games is 4 elo higher than K9.1 and 31 elo higher than K9. The contempt 0 version after 750 games is 5 elo higher than the default contempt 15 version, so 9 elo higher than K9.1 and 36 elo higher than K9

CEGT Blitz: http://www.husvankempen.de/nunn/40_4_Ra ... liste.html

[lkaufman]

I sent 1750 games with two versions of Komodo 9.2 running with 12 cores to CEGT 40/4, one with the default 15 contempt setting, and one with contempt set to 0. After 1000 games the default 15 contempt version scored 68.6% after 1000 games, the contempt 0 version scored 72.1% after 750 games.

The average opposition for the opponents was around 141 elo lower than Komodo 9.1 x64 12CPU. Shouldn't the default version score higher than the contempt 0 version?

my thread with the scores. The CEGT 40/4 rating list should be updated soon:
http://www.talkchess.com/forum/viewtopic.php?t=57353

Rating list updated. The default contempt 15 version after 1000 games is 4 elo higher than K9.1 and 31 elo higher than K9. The contempt 0 version after 750 games is 5 elo higher than the default contempt 15 version, so 9 elo higher than K9.1 and 36 elo higher than K9

CEGT Blitz: http://www.husvankempen.de/nunn/40_4_Ra ... liste.html

A difference of 5 elo after 750 games is not statistically meaningful, so probably there is nothing wrong with contempt, since default k9.2 seems to have done well generally. Maybe there is some obscure reason that contempt doesn't work as well with many cores running mp, but I don't think this is likely.

[shrapnel]

Then I set contempt to 5, recommended by Larry when playing against SF. The result in 100 games was 50,5-49,5 in favour of SF.

Yes ! Your results confirm my experience in online matches. The much-recommended Contempt 5 does NOTHING for Komodo in Matches against Stockfish. 15 is good as White and 0 is good as Black when using Komodo 9.2.

[lkaufman]

Then I set contempt to 5, recommended by Larry when playing against SF. The result in 100 games was 50,5-49,5 in favour of SF.

Yes ! Your results confirm my experience in online matches. The much-recommended Contempt 5 does NOTHING for Komodo in Matches against Stockfish. 15 is good as White and 0 is good as Black when using Komodo 9.2.

Well, although I doubt that color makes that much of a difference for proper contempt, if you are right then you are saying that 7.5 contempt is the best average value to use against Stockfish, not so far from my estimate of 5.

[shrapnel]

Well, although I doubt that color makes that much of a difference for proper contempt, if you are right then you are saying that7.5 contempt is the best average value to use against Stockfish, not so far from my estimate of 5.

I suppose we see the issue with different perspectives.
As a gamer, if I use C=15 I usually win with White, very few draws. If I use C=5, very few wins and many draws.
Obviously, THIS difference is very significant for me.
As Black, if I use Contempt 0, I NEVER lose, even against dual-Xeon users, let alone the average user ; and still manage to win a few, as I don't think Komodo's Contempt 0 is as defensive-minded ( so to speak) as say Houdini 4's Contempt 0.
With Contempt 5 as Black, there is a definite possibility of losing to equal or slightly stronger opponents.
So, again, this difference is important for me.

[lkaufman]

Well, although I doubt that color makes that much of a difference for proper contempt, if you are right then you are saying that7.5 contempt is the best average value to use against Stockfish, not so far from my estimate of 5.

I suppose we see the issue with different perspectives.
As a gamer, if I use C=15 I usually win with White, very few draws. If I use C=5, very few wins and many draws.
Obviously, THIS difference is very significant for me.
As Black, if I use Contempt 0, I NEVER lose, even against dual-Xeon users, let alone the average user ; and still manage to win a few, as I don't think Komodo's Contempt 0 is as defensive-minded ( so to speak) as say Houdini 4's Contempt 0.
With Contempt 5 as Black, there is a definite possibility of losing to equal or slightly stronger opponents.
So, again, this difference is important for me.

If you are correct, it implies that we should be able to improve Komdo's eval quite a bit; we shouldn't have to rely on changing contempt based on color. It gives me something to think about.

[lkaufman]

I sent 1750 games with two versions of Komodo 9.2 running with 12 cores to CEGT 40/4, one with the default 15 contempt setting, and one with contempt set to 0. After 1000 games the default 15 contempt version scored 68.6% after 1000 games, the contempt 0 version scored 72.1% after 750 games.

The average opposition for the opponents was around 141 elo lower than Komodo 9.1 x64 12CPU. Shouldn't the default version score higher than the contempt 0 version?

my thread with the scores. The CEGT 40/4 rating list should be updated soon:
http://www.talkchess.com/forum/viewtopic.php?t=57353

I checked all the CCRL and CEGT rating lists with various numbers of cores, considering only those lists that had established ratings for k9.0, k9.1, and k9.1. Averaging these six lists, I get that the elo gain from Komodo 9.0 to 9.1 was 20 elo, from 9.1 to 9.2 (using the default value for 9.2 in your test) was again another 20 elo. Since this is somewhat better than what I believe was the elo gain from 9.1 to 9.2 excluding contempt, the evidence on balance is that contempt did help against the range of opponents tested by CCRL and CEGT, but only by a few elo points. Stockfish 6 came out ten elo above K9.0 but ten below 9.1 and thirty below 9.2.