TalkChess.com

We have released Dragon by Komodo Chess at komodochess.com today. It uses the NNUE technology (but not the code) developed for the game of shogi. The improvement in strength over the latest standard Komodo, version 14.1, is enormous, roughly 160 elo on four threads, 190 elo on one thread, based on 2' +1" matches. Gains are similar in standard and MCTS mode. Standard mode remains the strongest, while MCTS mode is the strongest when MultiPV is used. The net used is embedded, so no need for the user to do anything differently than with previous Komodo versions, although there is an option (Use Regular Eval) to turn off the use of the net. The net was trained on billions of positions and evals from Komodo games. We may offer other nets in the future but not now. We think that the style of play is more human-like than with engines that don't use any NN, because Dragon is not so reliant on arbitrary rules for evaluation.
Dragon is priced at $74.98, a one year subscription is $119.98. Customers who purchased Komodo 13 or later can get a 20% discount; if the system doesn't give the discount just send an email after purchase requesting the 20% as a refund. Komodo 14.1 is $49.98, Komodo 13 is on sale for $19.98, and Komodo 12 becomes free.
Testing shows that Dragon is now stronger than all non-nnue engines as well as all GPU-based engines on most computers. Against SF12 (NNUE) it is generally losing by only about 25 elo points in fast games (1 or 4 threads); we don't yet know how they'll do in long games. In chess960 (FRC) Dragon was able to defeat Stockfish 12 in blitz testing. When both engines were set to use MultiPV=4 for standard chess, Dragon MCTS was able to defeat SF12 decisively in a 3600 games blitz match, so we believe that on most computers it is the world's strongest engine when looking at four or more lines of play at once.
We hope to continue to improve Dragon, both search and nets, in the near future. Special thanks to Dietrich Kappe for his work on the nets, to Dmitry Pervov for his writing of the NNUE code and for many improvements to the Dragon engine, and to chess.com for supporting this project financially.

Many thanks. Makes fun watching it.

Hi Larry, Mark, Dmitry & Dietrich,

a stronger team now!

Congratulation for Dragon by Komodo Chess and good luck for the future!

I am absolutly sure that many games vs. Stockfish & CFish will be follow by many users in the next minutes, hours, days, weeks!
I will follow ...

Best
Frank

May the people rejoice while originality fades away.

lkaufman wrote: ↑Tue Nov 10, 2020 7:42 pm When both engines were set to use MultiPV=4 for standard chess, Dragon MCTS was able to defeat SF12 decisively in a 3600 games blitz match, so we believe that on most computers it is the world's strongest engine when looking at four or more lines of play at once.

This marketing claim "forgets" to acknowledge that this increased strength of the first move is achieved by reducing how deep the engines looks at lines beyond the first one. Indeed, the ability to report multiple PVs with eval, and dedicating similar resources to explore each of the lines, are separate concepts.

Play with MultiPV=4 only reflects the quality of the first move, not that of the 2nd, 3rd or 4th, and is a fundamentally flawed way to evaluate how useful a MultiPV implementation is.

The increased quality of the first move in Komodo (and Leela) MCTS comes at the cost of a lowered quality and reliability for the moves, PV and evals of the 2nd and further lines come from smaller search trees. Turning MultiPV on in these engines doesn't increase the likelihood that they discover an initially disliked move is much better than first thought.

Most people that do use MultiPV for analysis expressed strong opposition to this trade-off when reducing 2nd and further PV lines was considered for Stockfish, though in some circumstances (moves that are much worse than the main move) it can be useful.

For deep analysis, multiPV isn't going to be used on positions with one clear forced move anyway, and the effective time-handicap that makes KMCTS look good at blitz matter much less while the search/eval deficit remain, making Stockfish the best MultiPV engine for e.g. correspondence chess.

lkaufman wrote: ↑Tue Nov 10, 2020 7:42 pm Customers who purchased Komodo 13 or later can get a 20% discount;

Thanks for komodo Dragon and for the discount for 13+ customers too, whilst at sales- site only 14+ are mentioned only so far.
Looking forward to buy my shiny new Dragon by Komodo Chess soon now regards

Alayan wrote: ↑Tue Nov 10, 2020 8:42 pm May the people rejoice while originality fades away.

On the contrary, according to Mark Lefter, the codebase for Komodo and Dragon have been merged, and just as in Stockfish, if one wishes, one could disable the NNUE evaluation and only use Komodo's handcrafted evaluation function.

mjlef wrote: ↑Tue Nov 10, 2020 1:52 pm Yesterday, we merged the code base for Komodo and Dragon by Komodo Chess, so we retain the ability to make compiles of either version. Given the large elo gain NNUE has given us, we do not plan on releasing new versions of Komodo for now. Note that Dragon has a UCI option "Use Regular Eval" which tells Dragon to use the Komodo regular (not NNUE) evaluation. So basically Komodo still exists and can be switched to at any time you want. It is not quite the same as say Komodo 14.1, because some search parameters have changed with Dragon, but it is close. In a sense it is 8 programs: Dragon, Komodo, Dragon MCTS, and Komodo MCTS, with both NNUE or Regular evals depending on how you set it. Komodo will still exist in some things like Apps or commercial chess boards which do not have sufficient CPUs or memory to run Dragon by Komodo Chess well.

A short test Dragon NNUE. You still have problems with time, unfortunately the same bug is in Komodo.

http://www.mediafire.com/file/zwmyya23n ... E.pgn/file

Very early chess960 results here are incredibly impressive. Massive jump in performance.

Alayan wrote: ↑Tue Nov 10, 2020 8:42 pm May the people rejoice while originality fades away.

lkaufman wrote: ↑Tue Nov 10, 2020 7:42 pm When both engines were set to use MultiPV=4 for standard chess, Dragon MCTS was able to defeat SF12 decisively in a 3600 games blitz match, so we believe that on most computers it is the world's strongest engine when looking at four or more lines of play at once.

This marketing claim "forgets" to acknowledge that this increased strength of the first move is achieved by reducing how deep the engines looks at lines beyond the first one. Indeed, the ability to report multiple PVs with eval, and dedicating similar resources to explore each of the lines, are separate concepts.

Play with MultiPV=4 only reflects the quality of the first move, not that of the 2nd, 3rd or 4th, and is a fundamentally flawed way to evaluate how useful a MultiPV implementation is.

The increased quality of the first move in Komodo (and Leela) MCTS comes at the cost of a lowered quality and reliability for the moves, PV and evals of the 2nd and further lines come from smaller search trees. Turning MultiPV on in these engines doesn't increase the likelihood that they discover an initially disliked move is much better than first thought.

Most people that do use MultiPV for analysis expressed strong opposition to this trade-off when reducing 2nd and further PV lines was considered for Stockfish, though in some circumstances (moves that are much worse than the main move) it can be useful.

For deep analysis, multiPV isn't going to be used on positions with one clear forced move anyway, and the effective time-handicap that makes KMCTS look good at blitz matter much less while the search/eval deficit remain, making Stockfish the best MultiPV engine for e.g. correspondence chess.

It is true that with Komodo MCTS (or Lc0 and other MCTS engines) the second, third, and later lines are analyzed less than the first line. However it's based on how close they are to each other. In positions where several moves appear to be nearly equally good, the analysis of the later moves will be nearly as good as the analysis of the best move. In positions where there are say two plausible moves and the rest are obvious blunders, then the two plausible moves will be analyzed properly and the others will get short analysis. This is how it should work in my opinion. Spending equal time on obvious blunders is obviously not a sensible use of resources. Occasionally it will pay off, but that's not the usual case.

Krzysztof Grzelak wrote: ↑Tue Nov 10, 2020 9:45 pm A short test Dragon NNUE. You still have problems with time, unfortunately the same bug is in Komodo.

http://www.mediafire.com/file/zwmyya23n ... E.pgn/file

The file shows a ten game match with Lc0 with a score of one win each and eight draws, and the sole loss was by adjudication, not time. Could you please explain what you think is the problem with Komodo or Dragon time management? It's not obvious from the file.