Komodo 9.4

[Nordlandia]

In which case of scenario is contempt generally a good idea to use when running infinite analysis. Is there any example positions?

[Leto]

In which case of scenario is contempt generally a good idea to use when running infinite analysis. Is there any example positions?

Any time you are playing against someone weaker than you or when you need to win. With contempt Komodo is willing to trade slightly (or a lot depending on how high you set the contempt to) worse in hopes of outplaying the opponent to get the win.

[Nordlandia]

In which case of scenario is contempt generally a good idea to use when running infinite analysis. Is there any example positions?

Any time you are playing against someone weaker than you or when you need to win. With contempt Komodo is willing to trade slightly (or a lot depending on how high you set the contempt to) worse in hopes of outplaying the opponent to get the win.

[d]5k2/ppp2ppp/1b6/8/6b1/2P5/PP4PP/RN2K3 w - - 0 20

Black has a pawn and the bishop pair for the exchange. Is it enough compensation?

Alright so Komodo 9.4 play black with the intension to draw at very best play. What is the correct contempt value in case K9.4 face last stockfish dev with white pieces.

Position is assessed at +/- | +0.55 to 0.75 depending on engine.

--------------------------------------------------------------------------
In accordance with Larry's rule of thumb e.g use one third of the odds as contempt, which translates to ~ "-0.20" may be adequate for black.

Please some advice whether if my proposed contempt is correct or wrong.

[hgm]

These ideas are reasonable, but even without castling the move ..Kf8 avoids the two safe checks; the problem is much more general than just castling option.

The point is that if a single opponent move, be it O-O or Kf8, can make the safe checks disappear before you ever get the opportunity to cash them, and as this does not replace them by similar or worse safe checks, these original checks cannot be worth very much. It could very well be that in a large fraction of the cases where this eval term kicks in there is no way to evade the threat, and that in these cases the results are so devastating that it Elo-wise pays to gamble and sac a significant amount of other eval points (usually material) to create this situation, even though there is a significant chance that it is a dud. This would explain why it tunes to such a high value. But that doesn't mean you couldn't gain even more by not make it a gamble, but only go for it when you are sure that it will stick, and invest the few extra nodes to verify that. Otherwise it will often happen that when such a position is horizon node, the search will completely restructure the tree to force it, while in the next iteration it has to completely restructure the tree again, because it now sees it was all wishful thinking. Discarding it in the same iteration would then have been much cheaper.

In this particular case white has suddenly pocketed a large amount of eval in the last move before the horizon, and that is always suspect. Would he have caused a similar rise in eval by capturing a minor rather than creating safe checks, we would always consider the recapture, and I think everyone agrees it would be totally wrecking to prune that recapture if it seems good (SEE-wise).

To carry through the analogy with QS: a capture at the horizon makes us run SEE on all recaptures as a preliminary investigation of whether this gain is illusory, and when this indicates it might be (the recapture qualifies as 'good capture'), you search the recapture. A sudden huge raise in eval score of non-material eval terms could similarly trigger a 'FEE': feature-exchange evaluation. Like SEE restricts the retaliation(s) to a single square, FEE would restrict the retaliation to the same evaluation feature, in this case king safety through save checks. So if the previous move jumped up king safety, all moves should be screened for their potential to undo that (FEE). E.g. by stepping the King out of those intended checks, blocking the check rays or protecting the check squares. The moves could then be sorted according to how well they do this (and perhaps on how attractive they are PST-wise), and those that look good enough then searched in QS. E.g. a6 would prevent Qb5+, but not Qa4+, and not be so hot PST-wise, so it would be sorted late or fail the FEE test. While O-O, Kf8, Bd7, Qe7, Nc6 or c6 would all prevent both checks. Of which Kf8 would score very poor (giving up castling rights), and O-O probably very good. So you try O-O first, and would probably fail high, frustrating the only thing that made Qb3 attractive. If you did not have any castling rights, and Bd7 would have seemed best, you would have tried that. But it would fail low due to Qxb7. So you get to try another one, Nc6 (also good PST-wise), and as there is no tactical backlash on that one, it would make you happy and prove the futility of Qb3.

[lkaufman]

In which case of scenario is contempt generally a good idea to use when running infinite analysis. Is there any example positions?

Any time you are playing against someone weaker than you or when you need to win. With contempt Komodo is willing to trade slightly (or a lot depending on how high you set the contempt to) worse in hopes of outplaying the opponent to get the win.

[d]5k2/ppp2ppp/1b6/8/6b1/2P5/PP4PP/RN2K3 w - - 0 20

Black has a pawn and the bishop pair for the exchange. Is it enough compensation?

Alright so Komodo 9.4 play black with the intension to draw at very best play. What is the correct contempt value in case K9.4 face last stockfish dev with white pieces.

Position is assessed at +/- | +0.55 to 0.75 depending on engine.

--------------------------------------------------------------------------
In accordance with Larry's rule of thumb e.g use one third of the odds as contempt, which translates to ~ "-0.20" may be adequate for black.

Please some advice whether if my proposed contempt is correct or wrong.

If we consider Stockfish and Komodo as equals, there is no need for Contempt in your example. It is different with a handicap game, when the presumption is that the odds-giver is stronger and that the handicap is at least roughly correct.

[Nordlandia]

Not sure whether "-0.20" will make Komodo play that particular position differently than "0"

Negative contempt may be used when Komodo face stronger opponent right, so my idea i had was that negative contempt somehow offset white's trivial edge, i may be wrong though.

[lkaufman]

Not sure whether "-0.20" will make Komodo play that particular position differently than "0"

Negative contempt may be used when Komodo face stronger opponent right, so my idea i had was that negative contempt somehow offset white's trivial edge, i may be wrong though.

Both sides should theoretically play the unequal position properly without contempt. Using it will influence play, but it only makes sense if you have a reason to expect one player to find better moves than the other one. I suppose you could also use it if you were convinced that Komodo was misevaluating the position.