Equal opponent is it ever wise to use Komodo contempt?

[Laskos]

We did make a change to Contempt (for Komodo 10) based on your findings, but it is a small change and appears to be worth just one or two elo. But we'll take it, thanks!

It's normal that the eval change brings not very much, as Contempt is intended to be an external factor accounting for the strength of the opponent, an information unavailable to eval. It's only that Komodo Contempt has a little defect, and correcting for this defect brings a couple of ELO points. As for Contempt itself, I just tested a bit Komodo in self-play for simulating handicap matches, we saw long time ago that self-play is a better ELO estimator of human-engine differences. And it seems that Contempt helps a lot Komodo to get better results against humans:

Code: Select all

Komodo 12s+0.12s vs Komodo 2s+0.02s 
Disadvantage 200-240cp for stronger side

Contempt=0
+290 =382 -328    48.1%  

Contempt=+60
+429 =275 -296    56.7%

Contempt=+75
+459 =247 -294    58.3%    +71 ELO points

Contempt=+90
+444 =238 -318    56.3%    

Contempt=+120
+432 =219 -349    54.2%

Contempt=75 for 220cp handicap adds 71 ELO points to Komodo against humans (if simulation is meaningful). It also brings 58% more wins. Seems a very successful implementation of Contempt.

[Nordlandia]

<Kai Laskos>

It it wise to adjust contempt in proposed Time Odds match?

We might consider big time odds (maybe 2 hours + 1 min vs 2 min + 1 sec) combined with White pieces for human, no opening book (except maybe one or two moves for variety), and no ponder. But this only makes sense if we get another top-ten player like Nakamura, and even then Komodo would be heavily favoured unless we also specify Komodo must run on laptop computer.

[Laskos]

<Kai Laskos>

It it wise to adjust contempt in proposed Time Odds match?

We might consider big time odds (maybe 2 hours + 1 min vs 2 min + 1 sec) combined with White pieces for human, no opening book (except maybe one or two moves for variety), and no ponder. But this only makes sense if we get another top-ten player like Nakamura, and even then Komodo would be heavily favoured unless we also specify Komodo must run on laptop computer.

Larry had very good estimation of Contempt for handicap matches as a fraction of pawns given (He had 40%, here I get once 40% another time 35%). My doubt about this match is that 2min+1s is too much to give to his multicore computer. But who knows, maybe all my simulations were off. Also, it doesn't seem to me very much fun to see an almost instantly playing engine against a tortured super GM.

[Laskos]

Contempt=75 for 220cp handicap adds 71 ELO points to Komodo against humans (if simulation is meaningful). It also brings 58% more wins. Seems a very successful implementation of Contempt.

I compared that to what Stockfish Contempt does in the same conditions. AFAIK Stockfish Contempt is simply minus Drawscore. And it gives 21 ELO points in the same conditions. So Mark and Larry work on Contempt can give additional 50 ELO points compared to simple Drawscore in handicap matches against humans.