Komodo 14.1 Release and Dragon Announcement

[carldaman]

How is the playing style of your nnue implementation?

Well, it's hard to make general statements about playing style. I would say it is more human-like, even though it was trained on engine games, because it is less reliant on arbitrary rules that might not be appropriate for a given position. With very short searches the difference is obvious, but as searches become deeper the differences become smaller.

A better question would be: does K-Dragon retain the style of the old Komodo?

[Werewolf]

Komodo 14.1 Release and Dragon Announcement

The improvement from Komodo 14.1 to Dragon is enormous, even larger than the gain Stockfish got from
this NNUE idea.

Strange.
For ages the marketing line / general understanding is that SF has the better search, and Komodo the better EF.

[Madeleine Birchfield]

Strange.
For ages the marketing line / general understanding is that SF has the better search, and Komodo the better EF.

That changed when Stockfish adopted neural networks for its evaluation function earlier this year. Then all of a sudden Stockfish had a much superior evaluation function over Komodo and still its much better search. However, Stockfish's own neural network is extremely unoptimised, so there is still a possibility that if Komodo Dragon trained its neural network well enough it would be able to surpass Stockfish with a better evaluation function again.

[Andrew]

I'd love to see these engines train networks specifically for KP and KPR endings. They are the most common in normal play.

However the big question is how big they would need to be useful. 1Mb? 10 Mb?

ELO advantage, tiny or reasonable?

Maybe useful studying certain middle or endgame positions/studies?

Lots of questions, but I hope we see experiments to get some answers!

Andrew

[Madeleine Birchfield]

I'd love to see these engines train networks specifically for KP and KPR endings. They are the most common in normal play.

However the big question is how big they would need to be useful. 1Mb? 10 Mb?

ELO advantage, tiny or reasonable?

Maybe useful studying certain middle or endgame positions/studies?

Lots of questions, but I hope we see experiments to get some answers!

Andrew

I think Andrew Grant's Ethereal and Jonathan Kreuzer's SlowChess both use networks for king and pawn endgames and other endgames, with a moderate amount of elo gain. Most other engines are moving towards using a neural network for the entire evaluation function, entirely sidestepping the need for specialised networks for endgames.

[Raphexon]

Komodo 14.1 Release and Dragon Announcement

The improvement from Komodo 14.1 to Dragon is enormous, even larger than the gain Stockfish got from
this NNUE idea.

Strange.
For ages the marketing line / general understanding is that SF has the better search, and Komodo the better EF.

Don't believe marketing.
I don't think anyone has comprehensive proof of Komodo's eval having been better in the past x years.

There used to be "Komodo scaling" on multicores when they were among the first to use LazySMP and the rest were still using YBWC, but those days are long gone.

[marsell]

This “blandness blender” theory isn’t supported by any evidence.

+1

[Marcus9]

Thanks for freeing the Dragon!
Are there any Dragon vs. Human matches scheduled?
I find them so exciting

[Thomas Lagershausen]

Sorry to say, but where is the data?

Till now Komodo Dragon is a papertiger.

[Ozymandias]

Sorry to say, but where is the data?

Till now Komodo Dragon is a papertiger.

After 54 games at CCC, it's only one and a half point behind SF @ 10+5. I imagine the final version will be on par.