Komodo Dragon 2 released.

[gaard]

Dragon 2 is now the favorite against Stockfish 13 on typical modern hardware.


Rgds

Is it? They seem to imply that with a pv of one the improvement is small, but the difference between Dragon 1 and SF 13 is 79 elo:

http://www.cegt.net/40_40%20Rating%20Li ... liste.html

The above is a fragment of the whole sentence. Larry wrote "We believe that the MultiPV search is now more effective than the Stockfish MultiPV search, and that if MultiPV is set to more than 4 Dragon 2 is now the favorite against Stockfish 13 on typical modern hardware."

So if you set both Dragon 2 and Stockfish to use MultiPV of 4 or more, Dragon 2 will be stronger in games, which basically means you will get better analysis quicker with Dragon 2 displaying 4 or more lines. Basically we improved MultiPV a lot in Dragon 2 (and frankly the old MultiPV scheme was not very efficient). I suppose the main reason people use strong programs is for analysis, so the improvements should help.

Mark

If during play (or analysis) an engine choses the move it thinks best...what does it matter that it's displaying 3,4,5 PV? Unless the extra lines displayed 'hurt' the overall evaluation of all 5 PV? I believe that is what you are sayin, yes (?) - that the extra computations hurt the choice of best move for Stockfish more so than KomodoDragon?

If so, is that due to search depth of each line or something with the evaluation function?

The main improvement is speed, so a MultiPV search now normally takes less time to reach a given depth. Note that setting MultiPV to 2 or higher can improve the move selection because sometimes the second search to get the second best move actually returns a move line with a score better than the first. Modern alpha-beta searches are highly selective and move order changes can produce better lines with additional search. So basically you get better analysis with the new Dragon MultiPV search mostly because it is is faster and on average a higher depth is reached in the same search time. Technically, I changed the old alpha-beta search scheme which basically kept the alpha beta window opened for the first MultiPV line to a scheme that does multiple exclusion searches, which ended up being faster.

Can you show that the 2nd, 3rd, 4th, 5th, etc., PV lines are any better than those produced by Stockfish? If not then the claim that Dragon 2 is stronger than Stockfish when MiltiPV >= 5 is argumentative at best.

[pohl4711]

Komodo Dragon 2 is released today at komodochess.com.

7000 games testrun of KomodoDragon 2.0 avx2 finished. Testrun of KomodoDragon 2.0 MCTS is running.

https://www.sp-cc.de

(Perhaps you have to clear your browsercache or reload the website)

[AdminX]

Is there a Dragon / Komodo channel setup on Discord? If so can you share the invite link?

[lkaufman]

For very deep analysis, should we be using MCTS? When should we be using MCTS?

Dragon MCTS is mostly useful for preparing for or playing against human opponents who will not be consulting engines during play. Since it doesn't assume that it can always predict opponents' moves, but only assumes that they will be "good" moves generally speaking, it will choose moves that should work better against humans. Its other advantage is that MultiPV is "free", but since we have improved MultiPV dramatically in standard mode in Dragon 2, I wouldn't choose MCTS solely for this reason anymore. But the more lines looked at with MultiPV, the smaller the elo gap between the two modes becomes, roughly vanishing when all legal moves are displayed.

[Werewolf]

Komodo Dragon 2 is released today at komodochess.com.

7000 games testrun of KomodoDragon 2.0 avx2 finished. Testrun of KomodoDragon 2.0 MCTS is running.

https://www.sp-cc.de

(Perhaps you have to clear your browsercache or reload the website)

Some nice new ideas with the multi PV but ultimately...6 elo over Dragon 1.

[cma6]

Dragon MCTS is mostly useful for preparing for or playing against human opponents who will not be consulting engines during play. Since it doesn't assume that it can always predict opponents' moves, but only assumes that they will be "good" moves generally speaking, it will choose moves that should work better against humans. Its other advantage is that MultiPV is "free", but since we have improved MultiPV dramatically in standard mode in Dragon 2, I wouldn't choose MCTS solely for this reason anymore. But the more lines looked at with MultiPV, the smaller the elo gap between the two modes becomes, roughly vanishing when all legal moves are displayed.
[/quote]

LK: Thanks for detailed explanation.

[Cornfed]

Komodo Dragon 2 is released today at komodochess.com.

7000 games testrun of KomodoDragon 2.0 avx2 finished. Testrun of KomodoDragon 2.0 MCTS is running.

https://www.sp-cc.de

(Perhaps you have to clear your browsercache or reload the website)

Some nice new ideas with the multi PV but ultimately...6 elo over Dragon 1.

I am suprized they called it Dragon 2 instead of 1.5 or some such.

[mjlef]

Dragon 2 is now the favorite against Stockfish 13 on typical modern hardware.


Rgds

Is it? They seem to imply that with a pv of one the improvement is small, but the difference between Dragon 1 and SF 13 is 79 elo:

http://www.cegt.net/40_40%20Rating%20Li ... liste.html

The above is a fragment of the whole sentence. Larry wrote "We believe that the MultiPV search is now more effective than the Stockfish MultiPV search, and that if MultiPV is set to more than 4 Dragon 2 is now the favorite against Stockfish 13 on typical modern hardware."

So if you set both Dragon 2 and Stockfish to use MultiPV of 4 or more, Dragon 2 will be stronger in games, which basically means you will get better analysis quicker with Dragon 2 displaying 4 or more lines. Basically we improved MultiPV a lot in Dragon 2 (and frankly the old MultiPV scheme was not very efficient). I suppose the main reason people use strong programs is for analysis, so the improvements should help.

Mark

If during play (or analysis) an engine choses the move it thinks best...what does it matter that it's displaying 3,4,5 PV? Unless the extra lines displayed 'hurt' the overall evaluation of all 5 PV? I believe that is what you are sayin, yes (?) - that the extra computations hurt the choice of best move for Stockfish more so than KomodoDragon?

If so, is that due to search depth of each line or something with the evaluation function?

The main improvement is speed, so a MultiPV search now normally takes less time to reach a given depth. Note that setting MultiPV to 2 or higher can improve the move selection because sometimes the second search to get the second best move actually returns a move line with a score better than the first. Modern alpha-beta searches are highly selective and move order changes can produce better lines with additional search. So basically you get better analysis with the new Dragon MultiPV search mostly because it is is faster and on average a higher depth is reached in the same search time. Technically, I changed the old alpha-beta search scheme which basically kept the alpha beta window opened for the first MultiPV line to a scheme that does multiple exclusion searches, which ended up being faster.

Can you show that the 2nd, 3rd, 4th, 5th, etc., PV lines are any better than those produced by Stockfish? If not then the claim that Dragon 2 is stronger than Stockfish when MiltiPV >= 5 is argumentative at best.

We are only saying that if you set both programs for a high enough MultiPV, then Drgaon does better. The overall MultiPV is finding better first moves. Our existing tester does not support playing 2nd or 3rd bets moves to do the experiment you propose. But we do know MultiPV is much more efficient in Dragon 2 than in earlier versions. I think a lot of people use these engines in MultiPV mode for analysis, and this should be useful to them. I do not think there will ever be a MultiPV rating list.

[mjlef]

Hi Larry,

the M1 version for Mac does not run.
Any hints?

regards
Chris

Yes, we're working on a fix for the M1 version problem, but if the Mac AVX2 version runs on your machine, that should be nearly as fast; the difference might turn out to be too small to justify separate versions.

A followup: AVX2 versions do not run on the m1 chip. the standard version (dragon-2-macos) does run. The compiler for MacOS m1 chips digitally stamps the executable, and the MacOS that can run on m1 does not allow running any non-stamped or modified executables. Since the system we use for copy protection modifies each file to make is unique, they will not run. Bui I just had an idea for something that might work. Sorry, this will take a few days. If we cannot come up with a fix i say a week, I would be glad to refund any MacOS m1 users.

Mark

[Uri Blass]

Dragon 2 is now the favorite against Stockfish 13 on typical modern hardware.


Rgds

Is it? They seem to imply that with a pv of one the improvement is small, but the difference between Dragon 1 and SF 13 is 79 elo:

http://www.cegt.net/40_40%20Rating%20Li ... liste.html

The above is a fragment of the whole sentence. Larry wrote "We believe that the MultiPV search is now more effective than the Stockfish MultiPV search, and that if MultiPV is set to more than 4 Dragon 2 is now the favorite against Stockfish 13 on typical modern hardware."

So if you set both Dragon 2 and Stockfish to use MultiPV of 4 or more, Dragon 2 will be stronger in games, which basically means you will get better analysis quicker with Dragon 2 displaying 4 or more lines. Basically we improved MultiPV a lot in Dragon 2 (and frankly the old MultiPV scheme was not very efficient). I suppose the main reason people use strong programs is for analysis, so the improvements should help.

Mark

If during play (or analysis) an engine choses the move it thinks best...what does it matter that it's displaying 3,4,5 PV? Unless the extra lines displayed 'hurt' the overall evaluation of all 5 PV? I believe that is what you are sayin, yes (?) - that the extra computations hurt the choice of best move for Stockfish more so than KomodoDragon?

If so, is that due to search depth of each line or something with the evaluation function?

The main improvement is speed, so a MultiPV search now normally takes less time to reach a given depth. Note that setting MultiPV to 2 or higher can improve the move selection because sometimes the second search to get the second best move actually returns a move line with a score better than the first. Modern alpha-beta searches are highly selective and move order changes can produce better lines with additional search. So basically you get better analysis with the new Dragon MultiPV search mostly because it is is faster and on average a higher depth is reached in the same search time. Technically, I changed the old alpha-beta search scheme which basically kept the alpha beta window opened for the first MultiPV line to a scheme that does multiple exclusion searches, which ended up being faster.

Can you show that the 2nd, 3rd, 4th, 5th, etc., PV lines are any better than those produced by Stockfish? If not then the claim that Dragon 2 is stronger than Stockfish when MiltiPV >= 5 is argumentative at best.

We are only saying that if you set both programs for a high enough MultiPV, then Drgaon does better. The overall MultiPV is finding better first moves. Our existing tester does not support playing 2nd or 3rd bets moves to do the experiment you propose. But we do know MultiPV is much more efficient in Dragon 2 than in earlier versions. I think a lot of people use these engines in MultiPV mode for analysis, and this should be useful to them. I do not think there will ever be a MultiPV rating list.

If the target is only to find better first moves then you can simply avoid using multi-pv.
If the target is to find candidates for best move then you need to compare probability that one of your choices in multi-pv become best after a long analysis.

It is possible to do it by analysis with dragon and stockfish with multi-pv=5 a long list of positions everyone for one second and later make a longer analysis of 1 minutes in every position with multi-pv=1

Now if you have 1000 positions and in 980 of them multi-pv=5 contains the choice with multi-pv=1 then your score is 98%

I do not know if dragon is stronger than stockfish with multi-pv=5 by this test but it is possible to test.