Steve Maughan wrote:It makes sense that a higher contempt value should be used for odds games. Do these results also imply that if Komodo's score starts to fall it should increase the contempt value as a game progresses. So maybe when Komodo's score dips below -0.5 pawns it starts to increase the contempt. This will make Komodo take risks (e.g. a risky king side attack) in the hope of averting a slow loss.
Just an idea - Steve
We do have a form of dynamic Contempt in Komodo. It does not work like you propose though.
I think it would be interesting to try a big negative value of Contempt for problem sets. This would encourage Komodo to capture and simplify. I am not sure it will help, but if anyone has a good tactical position set and could make some runs of different Contempt values, it would be very interesting.
Steve Maughan wrote:So maybe when Komodo's score dips below -0.5 pawns it starts to increase the contempt. This will make Komodo take risks (e.g. a risky king side attack) in the hope of averting a slow loss.
Just an idea - Steve
This must be a relative score change then, compared to a previous move, because the score is already below -2,
(depends on the handicap of course, but currently we talk about a Knight) or more at the beginning of the handicap game.
Steve Maughan wrote:It makes sense that a higher contempt value should be used for odds games. Do these results also imply that if Komodo's score starts to fall it should increase the contempt value as a game progresses. So maybe when Komodo's score dips below -0.5 pawns it starts to increase the contempt. This will make Komodo take risks (e.g. a risky king side attack) in the hope of averting a slow loss.
Just an idea - Steve
We do have a form of dynamic Contempt in Komodo. It does not work like you propose though.
I think it would be interesting to try a big negative value of Contempt for problem sets. This would encourage Komodo to capture and simplify. I am not sure it will help, but if anyone has a good tactical position set and could make some runs of different Contempt values, it would be very interesting.
Since Kai reported a nice improvement on a tactical test set using a large positive Contempt, it would be very strange if a large negative value helped as well. Anyway it would be good to know the optimum Contempt value for this purpose.
Steve Maughan wrote:It makes sense that a higher contempt value should be used for odds games. Do these results also imply that if Komodo's score starts to fall it should increase the contempt value as a game progresses. So maybe when Komodo's score dips below -0.5 pawns it starts to increase the contempt. This will make Komodo take risks (e.g. a risky king side attack) in the hope of averting a slow loss.
Just an idea - Steve
We do have a form of dynamic Contempt in Komodo. It does not work like you propose though.
I think it would be interesting to try a big negative value of Contempt for problem sets. This would encourage Komodo to capture and simplify. I am not sure it will help, but if anyone has a good tactical position set and could make some runs of different Contempt values, it would be very interesting.
Since Kai reported a nice improvement on a tactical test set using a large positive Contempt, it would be very strange if a large negative value helped as well. Anyway it would be good to know the optimum Contempt value for this purpose.
I am getting contradictory results for mostly tactical Arasan 16 and AFAIK exclusively tactical ECM suite. One is solved better with large positive Contempt, another with large negative. Some statistical flukes are not excluded, although pretty unlikely.
Steve Maughan wrote:It makes sense that a higher contempt value should be used for odds games. Do these results also imply that if Komodo's score starts to fall it should increase the contempt value as a game progresses. So maybe when Komodo's score dips below -0.5 pawns it starts to increase the contempt. This will make Komodo take risks (e.g. a risky king side attack) in the hope of averting a slow loss.
Just an idea - Steve
We do have a form of dynamic Contempt in Komodo. It does not work like you propose though.
I think it would be interesting to try a big negative value of Contempt for problem sets. This would encourage Komodo to capture and simplify. I am not sure it will help, but if anyone has a good tactical position set and could make some runs of different Contempt values, it would be very interesting.
Since Kai reported a nice improvement on a tactical test set using a large positive Contempt, it would be very strange if a large negative value helped as well. Anyway it would be good to know the optimum Contempt value for this purpose.
I am getting contradictory results for mostly tactical Arasan 16 and AFAIK exclusively tactical ECM suite. One is solved better with large positive Contempt, another with large negative. Some statistical flukes are not excluded, although pretty unlikely.
Both make nice bell curves, one peaking at +200, the other at -100. Probably it means that the two tests differ in some fundamental way. I suppose that a set with lots of sacrifices might favor negative contempt, since it makes the computer's pieces less valuable. One that featured subtle positional moves to win material or mate might favor positive contempt to keep more material on the board. I guess we can't give any general rule for tactical sets.
I suppose we could have a kind of Multi-PV mode, but instead of finding 2nd, 3rd,... best move it would display PVs for different Contempt values. Unfortunately, changing Contempt means the program should clear the hash table, to not use results from other searches with a varying contempt, but I could just clear the depth portion of the hash and save the best move and bounds to make the extra searches more efficient.
