New Houdini

[cdani]

I guess much higher eval granularity is a nice way to simultaneously pick up the best move and avoid unwelcome early draws.

+1
Even Andscacs has a (not much but growing) more complex evaluation.

[egiovannotti]

An engine could determine how strong is another analyzing deviations from the evaluations regarding to the own, this a once being outside the book.

[Milos]

An engine could determine how strong is another analyzing deviations from the evaluations regarding to the own, this a once being outside the book.

This would be a pretty useless approach. You can measure deviation only through eval, if eval difference is small it means position is more/less equal so the other engine would be assumed strong even if it wasn't. If there was a more significant difference in eval stronger engine would be already winning so any contempt there would be useless.

[corres]

[quote="mjlef"]My theory of the strong engines having too many draws is that they simply outsearch most opponents, and they assume the opponent's will see the draw too. The search in chess engines always take the "safest" move sequence that they find.
[/quote]
I agree with your views about that what causes too many draws in the play of strong engines.
But in this dispute around "contempt" is mixed the evaluation of material with the real idea of contempt. Based on the real idea of contempt this is a constant bias to shift the null of evaluation therefore the tendency of search to be modified.

[Lyudmil Tsvetkov]

I am an idealist, and for me the best service of a top engine is to always provide the best move.

But, if one's fans/customers want additional features like contempt, please do provide them.

I guess there 2 types of repetition draws: search ones and eval ones. Search ones, where avoiding the draw would lead to huge/losing disadvantage, are obviously unavoidable, while eval ones, where an engine evals its 3 best moves with 0.00, and the 4th-best with -20cps, in case some of the 0.00 scores do not lead to huge disadvantage, but are just evaluated as fully equal because of lack of sufficient eval factors, might of course be avoided by adding a few new (useful) eval features, even with minimal scores. In that case, you will have a best-move line to pick and continue the game even without contempt.

But if you would like to play on even with negative score, well, I do not know..., how negative migth that score become?

When a top engine knows who it is playing against, of course it will see its contempt to 100 for Skipper, and 3cps for a bit lower-rated opponent, and in that case in rating lists the struggle will be not only to see what engine plays best, but also what engine has the best-implemented contempt. I kind of do not like this, I do not know why. What if a top engine has not implemented contempt, or does not want to use it as a default?

Anyway, I am not very much concerned with rating lists, but simply what the best move in a position is, and besides I get a headache from all this contempt stuff.

Blessed are the weak (engines), that do not need contempt. ...

[corres]

I think that there is an imagination of the developers about working and controlling of chess engines and there is an expecting of users about the power of an engine. These two opinion does not meet each other in all cases. Moreover the commercial interests affect the picture too.
It is obvious that the "contempt" of Komodo is not the contempt in the understanding of common sense. Maybe it would be called "dynamism".

[Laskos]

Houdini keeps the Queen longer to win (average number of Queens on the board at the win adjudication, significant result):

Houdini: 0.95
Stockfish: 0.75
Komodo: 0.67

Interesting because Houdini is playing without contempt in the TCEC tournament.
I'm not sure whether the numbers above are just a statistical fluke, or whether the (rather big) changes to the H5 evaluation function have produced this side-effect.

Seems even more convincing now, after 52/62 games:

Average number of Queens on the board at the Win adjudication:

Houdini: 1.00
Stockfish: 0.70
Komodo: 0.67

[Laskos]

Houdini keeps the Queen longer to win (average number of Queens on the board at the win adjudication, significant result):

Houdini: 0.95
Stockfish: 0.75
Komodo: 0.67

Interesting because Houdini is playing without contempt in the TCEC tournament.
I'm not sure whether the numbers above are just a statistical fluke, or whether the (rather big) changes to the H5 evaluation function have produced this side-effect.

Seems even more convincing now, after 52/62 games:

Average number of Queens on the board at the Win adjudication:

Houdini: 1.00
Stockfish: 0.70
Komodo: 0.67

Just computed how reliable these results are: 94% likelihood that the average of Houdini is higher than the average of Stockfish, and 95% likelihood higher than Komodo average.

[Rebel]

What about point 12, calculating the initiative. I only tested it on a large set of positions and the outcome (derived from the numbers) was always right.

I haven't tested it in real games since I am on another Sabbath year.

[Lyudmil Tsvetkov]

Houdini keeps the Queen longer to win (average number of Queens on the board at the win adjudication, significant result):

Houdini: 0.95
Stockfish: 0.75
Komodo: 0.67

Interesting because Houdini is playing without contempt in the TCEC tournament.
I'm not sure whether the numbers above are just a statistical fluke, or whether the (rather big) changes to the H5 evaluation function have produced this side-effect.

Seems even more convincing now, after 52/62 games:

Average number of Queens on the board at the Win adjudication:

Houdini: 1.00
Stockfish: 0.70
Komodo: 0.67

Just computed how reliable these results are: 94% likelihood that the average of Houdini is higher than the average of Stockfish, and 95% likelihood higher than Komodo average.

another possible explanation: H overvalues the queen, which is a matter of fact, though this migth be simply an engine feature and not a weakness.