Dragon engine

[cma6]

I see LK has introduced Dragon but with little explanation. Is it a successor to Komodo or just a different engine.
What is Dragon's strength vs. Komodo 14 or vs. SF dev?
And how does one get updated nets to use in Dragon?

[Modern Times]

https://komodochess.com/

[Andrew]

No updated nets yet, however I expect a lot of work has been done on then since the initial release and
we'll see some updates soon! I believe I read the TCEC version was an update?

Andrew

[Andrew]

Also many posts by Larry on this forum re Dragon Komodo if you search!

Andrew

[cma6]

Andrew, thanks for reply. I found a few posts about Dragon in the thread " How much work is it to train an NNUE?" but nothing by Larry. From his site it was unclear if improved nets would be regularly provided for download or if they will be incorporated in updated versions of Dragon.

[Ozymandias]

This was the original thread.

[cma6]

Ozymandias, thanks for the reference. Now that 3 months have passed, what has been the experience of users doing very deep (correspondence chess) analysis of Dragon vs. latest SF dev?

Also, for LK or Mark or other experts: what changes in default Dragon parameters for deep analysis?

[cma6]

[quote=cma6 post_id=882335 time=1613152521 user_id=7478]
Ozymandias, thanks for the reference. Now that 3 months have passed, what has been the experience of users doing very deep (correspondence chess) analysis of Dragon vs. latest SF dev?

Also, for LK or Mark or other experts: what changes in default Dragon parameters for deep analysis?
For SF dev, I have found that "x64 for Modern computers" were faster than "AVX2" versions. What about for Komodo/Dragon?

My Komodo 14 parameters, presumably based on Mark and LK recommendations on this form:
Contempt 0; Dynamism 122; King safety 75; MCTS Explore 45; MCTS Hash 128; Progress Threshold 30; Reduction -20;
Selectivity 180; Skil 25; Syzygy Probe Depth 2; Syzygy Probe Limit 6; Table Memory 1024; Use LMR yes;

[cma6]

I believe the neural net used by SF is 20 MB. How big is Dragon's NN?

[gaard]

Ozymandias, thanks for the reference. Now that 3 months have passed, what has been the experience of users doing very deep (correspondence chess) analysis of Dragon vs. latest SF dev?

Also, for LK or Mark or other experts: what changes in default Dragon parameters for deep analysis?

Dragon is a great addition to complement SF. I frequently see very different PVs from them which gives you more ideas to pursue as you analyze.

I believe the neural net used by SF is 20 MB. How big is Dragon's NN?

Komodo 14.1 is approximately 1.56 MB while Dragon is 13.4 MB. If we assume the difference is close to the NN size, then 11.84 MB is your answer, although I'm not sure that's a very useful metric.