TalkChess.com

Komodo 14.1 Release and Dragon Announcement

Komodo 14.1 is released today at komodochess.com. It is a normal strength upgrade from Komodo 14,
about ten elo in normal mode and twelve in MCTS mode. It is free for subscribers and also this time for
anyone who bought Komodo 14. It will not run on processors older than the year 2008. Some of the elo gains
are from speedups, some from parameter tuning, LMR changes, revised aspiration, revised time management, etc.

We are also announcing our new "Dragon" version of Komodo, which is now playing in the chess.com CCC
tournament as "Mystery" and which we expect to release soon. It uses the new NNUE technology that was
developed for the game of shogi, but not the NNUE code. The search is Komodo search (with some parameters
tuned), and the nets we use are all trained on Komodo games and Komodo evals. The net is embedded so the
user need not do anything special to use it (though it can be turned off).

The improvement from Komodo 14.1 to Dragon is enormous, even larger than the gain Stockfish got from
this NNUE idea. Depending on the time control and number of threads, it should fall somewhere in the 150
to 200 elo range (assuming a 2013 or later processor); we'll be more precise with actual release. The big
gains apply to both standard and MCTS mode. We believe that for most users Dragon will be stronger than
any GPU based engine and any CPU based engine except for Stockfish with NNUE. Dragon is not yet equal to
Stockfish NNUE in blitz; with more time and more threads the gap shrinks and may disappear at some point,
we don't yet know. We are making steady progress with Dragon and hope to catch Stockfish even in fast chess
before long. As with normal Komodo, it is best to use MCTS mode when you want to look at the top three or more
lines at once.

Regarding pricing, we are reducing the price of Komodo 14.1 by ten dollars to $49.98 (less 20% if you
bought Komodo 13 or later). We are pricing Dragon at $74.97 due to the massive improvement and to extra
costs associated with it borne by our partner chess.com. Subscriptions will now be $119.97 (less 20% for
current subscribers or purchasers of Komodo 14 or 14.1) since they will include a year's worth of Dragon
upgrades. Current subscribers will get the Dragon upgrades for free for the remainder of their subscriptions.

Special thanks to Dietrich Kappe for his expertise in training the nets, to Dmitry Pervov of chess.com
for writing the NNUE code we use and for many other enhancements to Dragon, and to our webmaster Jesse Gersenson
who unleashed hundreds of computers to generate the billions of positions needed to train the nets.

Rather pricey indeed, but glad to hear of the release of Dragon.

During the course of a year, about how many revisions to the Nets do you expect to push out? Would one just go to the Komodo website and download them?

Would one need to have a subscription to do so...or by buying the initial Dragon, will one be able to run subsequent Komodo specific Dmitry Pervov Net tweaks (or even experiment with non-Komodo specific/non-Pervov nets) with that version?

How is the playing style of your nnue implementation?

Hi Larry,

thanks for Komodo 14.1.

Note:
Dragon by Bruno LUCAS (France) was in the past a partner program from Arena Chess GUI.
At this time I had many contacts to french programmers and Bruno is really a very nice guy.
Dragon by Bruno is a very positional program with around 2350 Elo.

Last available UCi Dragon is v4.6.
After I remember I wrote the readme file.
Program are a long time available on my Arena site as partner program with SOS for Arena, the French program's AnMon, Nejmet and others.

To have now a second program with the name "Dragon" isn't a good idea.

Such a situation we have with Nightmare in the past.
The German Nightmare and the Dutch Nightmare.

The Dutch Nightmare programmer selling his program to an other programmer. This programmer made around the year 2000 first experience with ideas around self "learning" and used the sources from Nighmare for his experiments. This programmer wrote a chess book about it and gave it programmers played the Dutch-open for free. Don Daily played this year the Dutch-ch in Leiden and got the book too (after all I know). I gave my book Dr. Ulrich Türke (Comet). Years later many of that ideas are available in Leela and so one.

Not the topic!
But 2 times the same name isn't a good idea!

In my opinion better to give your "Dragon" an other name!

Best
Frank

Maybe call it Komodo Dragon?

Cornfed wrote: ↑Mon Nov 02, 2020 6:32 pm Rather pricey indeed, but glad to hear of the release of Dragon.

During the course of a year, about how many revisions to the Nets do you expect to push out? Would one just go to the Komodo website and download them?

Would one need to have a subscription to do so...or by buying the initial Dragon, will one be able to run subsequent Komodo specific Dmitry Pervov Net tweaks (or even experiment with non-Komodo specific/non-Pervov nets) with that version?

We haven't made decisions yet about whether to release nets individually or just new Komodo Dragon versions with new nets at the usual intervals. One correction: Dmitry Pervov helped greatly in making the new Dragon work with the nets without using NNUE code and in finding ways to speed it up (and many other things), but the actual nets are primarily the work of Dietrich Kappe.

Good side : Leela is going to get kicked out of the next TCEC SuFi.
Bad side : training on Komodo data won't prevent Komodo Dragon from losing a lot of K's originality, for the most part becoming "Stockfish but worse".

Alayan wrote: ↑Mon Nov 02, 2020 8:41 pm Good side : Leela is going to get kicked out of the next TCEC SuFi.
Bad side : training on Komodo data won't prevent Komodo Dragon from losing a lot of K's originality, for the most part becoming "Stockfish but worse".

+1

Alayan wrote: ↑Mon Nov 02, 2020 8:41 pm Bad side : training on Komodo data won't prevent Komodo Dragon from losing a lot of K's originality, for the most part becoming "Stockfish but worse".

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

dkappe wrote: ↑Mon Nov 02, 2020 8:45 pm

Alayan wrote: ↑Mon Nov 02, 2020 8:41 pm Bad side : training on Komodo data won't prevent Komodo Dragon from losing a lot of K's originality, for the most part becoming "Stockfish but worse".

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

Someone told me the other day that you said in TCEC's Twitch chat something along the lines of "There is no NNUE training code that does not use SF's qsearch". Is this something you said, or am I hearing it second hand incorrectly? Can you elaborate. Are the networks you trained using SF's NNUE Trainer, and thus SF's qsearch?