Komodo 9.4

[lkaufman]

Komodo 9.4 was released this morning at komodochess.com. As usual, it is free for subscribers and 20% discounted for purchasers of Komodo 8 or later version. Elo gains from 9.3 are estimated at 15 on single thread (based on over 10,000 games at 2' + 1") and perhaps a bit more with MP. Gains from version 9.0 are over 50 elo based on our tests. Gains come from both search and eval, especially king safety and passed pawns. There are some new eval terms.
Komodo 9.4 has two new UCI options. "Contempt for White" allows the user to specify that contempt is just for one color (for Black just use negative sign), which is important in analysis mode. It should be left unchecked for normal play. The other new option, "Dynamism", gives the user full control over the style of Komodo; a low setting makes it prioritize static advantages like material and pawn structure, while a high setting prioritizes short-term ("dynamic") advantages like mobility and threats to the king. Although the default (100) is believed to be best for actual play, a lower value around 80 seems to produce more realistic evaluations in most positions so may be better for opening analysis.

[Werewolf]

Komodo 9.4 was released this morning at komodochess.com. As usual, it is free for subscribers and 20% discounted for purchasers of Komodo 8 or later version. Elo gains from 9.3 are estimated at 15 on single thread (based on over 10,000 games at 2' + 1") and perhaps a bit more with MP. Gains from version 9.0 are over 50 elo based on our tests. Gains come from both search and eval, especially king safety and passed pawns. There are some new eval terms.
Komodo 9.4 has two new UCI options. "Contempt for White" allows the user to specify that contempt is just for one color (for Black just use negative sign), which is important in analysis mode. It should be left unchecked for normal play. The other new option, "Dynamism", gives the user full control over the style of Komodo; a low setting makes it prioritize static advantages like material and pawn structure, while a high setting prioritizes short-term ("dynamic") advantages like mobility and threats to the king. Although the default (100) is believed to be best for actual play, a lower value around 80 seems to produce more realistic evaluations in most positions so may be better for opening analysis.

Since I spend a lot of time analysing openings in IDeA, this interests me.

It's not that I doubt you, but I don't see how there can be a discrepancy between that which gives the best results (d=100) and the most accurate value for opening analysis (d=80).

Please explain.

Thanks.

[lkaufman]

Komodo 9.4 was released this morning at komodochess.com. As usual, it is free for subscribers and 20% discounted for purchasers of Komodo 8 or later version. Elo gains from 9.3 are estimated at 15 on single thread (based on over 10,000 games at 2' + 1") and perhaps a bit more with MP. Gains from version 9.0 are over 50 elo based on our tests. Gains come from both search and eval, especially king safety and passed pawns. There are some new eval terms.
Komodo 9.4 has two new UCI options. "Contempt for White" allows the user to specify that contempt is just for one color (for Black just use negative sign), which is important in analysis mode. It should be left unchecked for normal play. The other new option, "Dynamism", gives the user full control over the style of Komodo; a low setting makes it prioritize static advantages like material and pawn structure, while a high setting prioritizes short-term ("dynamic") advantages like mobility and threats to the king. Although the default (100) is believed to be best for actual play, a lower value around 80 seems to produce more realistic evaluations in most positions so may be better for opening analysis.

Since I spend a lot of time analysing openings in IDeA, this interests me.

It's not that I doubt you, but I don't see how there can be a discrepancy between that which gives the best results (d=100) and the most accurate value for opening analysis (d=80).

Please explain.

Thanks.

This issue is very important to me. I can't say that I fully understand what's going on, but I think that the evaluation can be optimized for the search without being good for actual game positions. In other words, a "proper" evaluation may not be very good for a search, because the positions arising in a search are not typical of positions arising in a real game. I welcome any informed comments on this subject.

[cdani]

Thanks for the new version!

Although the default (100) is believed to be best for actual play, a lower value around 80 seems to produce more realistic evaluations in most positions so may be better for opening analysis.

You mean more realistic evaluations from human point of view? Or is consequence of some tests?

[lkaufman]

Thanks for the new version!

Although the default (100) is believed to be best for actual play, a lower value around 80 seems to produce more realistic evaluations in most positions so may be better for opening analysis.

You mean more realistic evaluations from human point of view? Or is consequence of some tests?

More realistic from human point of view, but also produces better results on very short searches (fixed depth or game in one second plus 10 ms). I'm not sure why the value needs to be higher for real game play.

[cdani]

More realistic from human point of view, but also produces better results on very short searches (fixed depth or game in one second plus 10 ms). I'm not sure why the value needs to be higher for real game play.

Higher values of king safety tend to be better for longer time control games. Is 100% like this for Andscacs. Don't know for mobility.

[lkaufman]

More realistic from human point of view, but also produces better results on very short searches (fixed depth or game in one second plus 10 ms). I'm not sure why the value needs to be higher for real game play.

Higher values of king safety tend to be better for longer time control games. Is 100% like this for Andscacs. Don't know for mobility.

Any idea why this is so?

[Henk]

Sometimes you need deep searches to make use of bad king safety opponent.

[Pio]

Hi Larry!

I think mainly the reason why more aggresive values for king's safety works for longer time controls is because it helps the search prune entire sub trees when finding draws due to three fold repitition in otherwise inferior positions or when finding wins/losses.

I think you could gain a lot from sorting the moves differently depending on the remaining depth. For example might moves that reduces the branching factor a lot like equal captures with respect to SEE might be better to sort in the front when close to the root while quiet moves if they seem to be better should be placed before some of the captures close to the leaves. You could do the same with agressive king threatening moves as well. Killer moves might be better to do close to the leaves but not close to the root since they do not reduce the branching factor.

Best Regards
Pio

[cma6]

LK:
Thanks to Mark and LK for standing by 12-month subscription model.