Komodo Dragon 2 released.

[lkaufman]

Komodo Dragon 2 is released today at komodochess.com. As usual, it is free for subscribers, priced the same as Dragon (1), which itself will be marked down 20% as will Komoro 14.1 shortly. Anyone who bought Komodo 14 or later version gets 20% discount from these prices.
The improvements over Dragon 1 are primarily in two areas. MultiPV search (in standard mode) has been completely rewritten, and now is
about a hundred elo stronger than in Dragon 1 (or older Komodo versions). It is sufficiently improved that we no longer advocate using MCTS mode
just for the MultiPV benefits it provides. The MCTS mode is also improved in Dragon 2, dramatically so when using Windows, where the elo gain
over the original Dragon is also about a hundred elo. There is a new UCI option called MCTS Optimism, which when set to a positive number
(default is +10) makes Dragon "trust" its own internal move preferences more while exploring the opponent's choices more; a higher setting
might be appropriate for a human opponent. MCTS mode remains the best option for super-strong but human-like play and for preparing openings for
human opponents who will not have access to engines while playing. The standard mode when not using MultiPV has also been improved, but only by
an estimated ten elo. It is nearly the same as the version that won the recent TCEC Swiss, and is a slight upgrade from the version that beat
Stockfish in the fifty game final match of the TCEC chess960 (FRC) championship.
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.
Dragon 2 comes with an AVX2 version (suitable for most PCs since 2013) and a standard one. The AVX2 version is quite a bit faster and
stronger, and we hope that the testing groups will limit testing to AVX2 machines if at all possible. It probably doesn't matter when running
against Stockfish 12 or later, which also gets a lot out of AVX2, but it makes a big difference against programs not using NNUE. There is no
BMI2 version for Dragon 2 as we have not found any meaningful benefit to BMI2 beyond the benefit of AVX2.
Finally, for those who bought Komodo 14 or 14.1 but not Dragon, the strength gains from Komodo 14 in either MCTS or standard mode are nearly
200 elo, only slightly less from Komodo 14.1. All estimated elo gains are at the CCRL blitz time control of 2' + 1" increment, and use Ordo rather
than BayesElo.

[bastiball]

Komodo Dragon 2 is released today at komodochess.com. As usual, it is free for subscribers, priced the same as Dragon (1), which itself will be marked down 20% as will Komoro 14.1 shortly. Anyone who bought Komodo 14 or later version gets 20% discount from these prices.
The improvements over Dragon 1 are primarily in two areas. MultiPV search (in standard mode) has been completely rewritten, and now is
about a hundred elo stronger than in Dragon 1 (or older Komodo versions). It is sufficiently improved that we no longer advocate using MCTS mode
just for the MultiPV benefits it provides. The MCTS mode is also improved in Dragon 2, dramatically so when using Windows, where the elo gain
over the original Dragon is also about a hundred elo. There is a new UCI option called MCTS Optimism, which when set to a positive number
(default is +10) makes Dragon "trust" its own internal move preferences more while exploring the opponent's choices more; a higher setting
might be appropriate for a human opponent. MCTS mode remains the best option for super-strong but human-like play and for preparing openings for
human opponents who will not have access to engines while playing. The standard mode when not using MultiPV has also been improved, but only by
an estimated ten elo. It is nearly the same as the version that won the recent TCEC Swiss, and is a slight upgrade from the version that beat
Stockfish in the fifty game final match of the TCEC chess960 (FRC) championship.
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.
Dragon 2 comes with an AVX2 version (suitable for most PCs since 2013) and a standard one. The AVX2 version is quite a bit faster and
stronger, and we hope that the testing groups will limit testing to AVX2 machines if at all possible. It probably doesn't matter when running
against Stockfish 12 or later, which also gets a lot out of AVX2, but it makes a big difference against programs not using NNUE. There is no
BMI2 version for Dragon 2 as we have not found any meaningful benefit to BMI2 beyond the benefit of AVX2.
Finally, for those who bought Komodo 14 or 14.1 but not Dragon, the strength gains from Komodo 14 in either MCTS or standard mode are nearly
200 elo, only slightly less from Komodo 14.1. All estimated elo gains are at the CCRL blitz time control of 2' + 1" increment, and use Ordo rather
than BayesElo.

Congrats for Dragon 2 released!

[Damir]

Thanks for the new version Larry! Unfortunately when I try and load the new version I get this message:
The requested URL dragon-2-ce44fe7f.zip was not found on this server.....

[Werewolf]

Great work, thanks!

So how much elo do you expect in normal mode (PV=1) over Dragon 1? 10 elo?

Also, did you upgrade the net?

[criko]

Hi,

thx for your engine.
The Mac version is running on my Mac Mini M1, but the M1 version does not run

regards
Chris

[cucumber]

Hi, I have a few questions regarding the MCTS mode that I was wondering if you could answer:

1. Does it support tree-reuse?
In Leela, the search tree might look like
1. e4
-> ...e5
-> ...c5
1. d4
-> ...nf6
-> ...d5

Rather than simply rely on the TT/NN cache, if e4 is played, Leela will just drop d4 (and other) branches of the search tree and preserve the entirety of the 1. e4 search tree. This can occasionally save a lot of compute.

2. Does it support smart pruning?
Leela can calculate whether--at the current nodes per second, with some margin--whether it's remotely probable that the MCTS algorithm will ever switch to a new move within the allocated time budget. If it's not possible (if the current best move has too many visits for any other move to catch up) it will simply play the best move.

This just helps improve TM a bit. It's nothing super useful as an end user, but I was curious.

3. Do you have any plans to support Nibbler-compatible output?
I've asked this before, but I figure that it's worth asking again. Sorry if this is obnoxious

Nibbler is a chess GUI for MCTS engines that displays a lot of useful analysis information. It is very similar to Go analysis GUIs, like Lizzie. Support for Nibbler would be really nice.

Thank you so much for your patience answering these questions, and congratulations to you and the K team on another release!

[lkaufman]

Hi, I have a few questions regarding the MCTS mode that I was wondering if you could answer:

1. Does it support tree-reuse?
In Leela, the search tree might look like
1. e4
-> ...e5
-> ...c5
1. d4
-> ...nf6
-> ...d5

Rather than simply rely on the TT/NN cache, if e4 is played, Leela will just drop d4 (and other) branches of the search tree and preserve the entirety of the 1. e4 search tree. This can occasionally save a lot of compute.

2. Does it support smart pruning?
Leela can calculate whether--at the current nodes per second, with some margin--whether it's remotely probable that the MCTS algorithm will ever switch to a new move within the allocated time budget. If it's not possible (if the current best move has too many visits for any other move to catch up) it will simply play the best move.

This just helps improve TM a bit. It's nothing super useful as an end user, but I was curious.

3. Do you have any plans to support Nibbler-compatible output?
I've asked this before, but I figure that it's worth asking again. Sorry if this is obnoxious

Nibbler is a chess GUI for MCTS engines that displays a lot of useful analysis information. It is very similar to Go analysis GUIs, like Lizzie. Support for Nibbler would be really nice.

Thank you so much for your patience answering these questions, and congratulations to you and the K team on another release!

Short answers to 1,2,3: No, Yes, No. We would look into (1) or (3) if we get requests from several customers.

[cucumber]

Hi, I have a few questions regarding the MCTS mode that I was wondering if you could answer:

1. Does it support tree-reuse?
In Leela, the search tree might look like
1. e4
-> ...e5
-> ...c5
1. d4
-> ...nf6
-> ...d5

Rather than simply rely on the TT/NN cache, if e4 is played, Leela will just drop d4 (and other) branches of the search tree and preserve the entirety of the 1. e4 search tree. This can occasionally save a lot of compute.

2. Does it support smart pruning?
Leela can calculate whether--at the current nodes per second, with some margin--whether it's remotely probable that the MCTS algorithm will ever switch to a new move within the allocated time budget. If it's not possible (if the current best move has too many visits for any other move to catch up) it will simply play the best move.

This just helps improve TM a bit. It's nothing super useful as an end user, but I was curious.

3. Do you have any plans to support Nibbler-compatible output?
I've asked this before, but I figure that it's worth asking again. Sorry if this is obnoxious

Nibbler is a chess GUI for MCTS engines that displays a lot of useful analysis information. It is very similar to Go analysis GUIs, like Lizzie. Support for Nibbler would be really nice.

Thank you so much for your patience answering these questions, and congratulations to you and the K team on another release!

Short answers to 1,2,3: No, Yes, No. We would look into (1) or (3) if we get requests from several customers.

That's understandable. Again, thank you so much for the answers!

It would be interesting to see if tree reuse gained elo.

[lkaufman]

Hi, I have a few questions regarding the MCTS mode that I was wondering if you could answer:

1. Does it support tree-reuse?
In Leela, the search tree might look like
1. e4
-> ...e5
-> ...c5
1. d4
-> ...nf6
-> ...d5

Rather than simply rely on the TT/NN cache, if e4 is played, Leela will just drop d4 (and other) branches of the search tree and preserve the entirety of the 1. e4 search tree. This can occasionally save a lot of compute.

2. Does it support smart pruning?
Leela can calculate whether--at the current nodes per second, with some margin--whether it's remotely probable that the MCTS algorithm will ever switch to a new move within the allocated time budget. If it's not possible (if the current best move has too many visits for any other move to catch up) it will simply play the best move.

This just helps improve TM a bit. It's nothing super useful as an end user, but I was curious.

3. Do you have any plans to support Nibbler-compatible output?
I've asked this before, but I figure that it's worth asking again. Sorry if this is obnoxious

Nibbler is a chess GUI for MCTS engines that displays a lot of useful analysis information. It is very similar to Go analysis GUIs, like Lizzie. Support for Nibbler would be really nice.

Thank you so much for your patience answering these questions, and congratulations to you and the K team on another release!

Short answers to 1,2,3: No, Yes, No. We would look into (1) or (3) if we get requests from several customers.

That's understandable. Again, thank you so much for the answers!

It would be interesting to see if tree reuse gained elo.

I don't know whether tree reuse as defined in Leela would make sense with Dragon MCTS. They are pretty different. With Dragon MCTS, once we play a move the analysis done is not very relevant as now there is some asymmetry between how we handle opponent's moves and engine moves as well as root moves vs. nonroot moves.

[Chessqueen]

Komodo Dragon 2 is released today at komodochess.com. As usual, it is free for subscribers, priced the same as Dragon (1), which itself will be marked down 20% as will Komoro 14.1 shortly. Anyone who bought Komodo 14 or later version gets 20% discount from these prices.
The improvements over Dragon 1 are primarily in two areas. MultiPV search (in standard mode) has been completely rewritten, and now is
about a hundred elo stronger than in Dragon 1 (or older Komodo versions). It is sufficiently improved that we no longer advocate using MCTS mode
just for the MultiPV benefits it provides. The MCTS mode is also improved in Dragon 2, dramatically so when using Windows, where the elo gain
over the original Dragon is also about a hundred elo. There is a new UCI option called MCTS Optimism, which when set to a positive number
(default is +10) makes Dragon "trust" its own internal move preferences more while exploring the opponent's choices more; a higher setting
might be appropriate for a human opponent. MCTS mode remains the best option for super-strong but human-like play and for preparing openings for
human opponents who will not have access to engines while playing. The standard mode when not using MultiPV has also been improved, but only by
an estimated ten elo. It is nearly the same as the version that won the recent TCEC Swiss, and is a slight upgrade from the version that beat
Stockfish in the fifty game final match of the TCEC chess960 (FRC) championship.
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.
Dragon 2 comes with an AVX2 version (suitable for most PCs since 2013) and a standard one. The AVX2 version is quite a bit faster and
stronger, and we hope that the testing groups will limit testing to AVX2 machines if at all possible. It probably doesn't matter when running
against Stockfish 12 or later, which also gets a lot out of AVX2, but it makes a big difference against programs not using NNUE. There is no
BMI2 version for Dragon 2 as we have not found any meaningful benefit to BMI2 beyond the benefit of AVX2.
Finally, for those who bought Komodo 14 or 14.1 but not Dragon, the strength gains from Komodo 14 in either MCTS or standard mode are nearly
200 elo, only slightly less from Komodo 14.1. All estimated elo gains are at the CCRL blitz time control of 2' + 1" increment, and use Ordo rather
than BayesElo.

Do you believe it would be a good idea to create a site where Dragon 2 can be rented per day, lets say $10.00 for 1 day