GM Kaufman speaks about his pet dragon — Komodo 13

by Davide Nastasio
8/20/2019 – Some people keep a Komodo dragon in their yard, and some keep it in their living room, safely inside their computer! In this interview with Grandmaster Larry Kaufman, one of the creators of Komodo, the World Chess Software Champion, we gain new insight into the latest release of the popular chess program. Komodo has gained strength both in the standard version, but particularly in the Monte Carlo Tree Search (MCTS), making the engine a valuable research tool for correspondence games, and openings.

Komodo Chess 13 Komodo Chess 13

Komodo 13 thinks like no other chess program. Inspired by AlphaZero, Komodo developers GM Larry Kaufman and Mark Lefler have reinvented their engine from scratch over the last two years. The result speaks for itself: The new Komodo 13 MCTS ("Monte Carlo Tree Search") searches for candidate moves in an incredibly innovative way and finds solutions most engines never see!

More...

Komodo 13 is the ICGA the new Champion

Last week Komodo won the ICGA World Chess Software Championship with 6½/10 (3 wins, 7 draws, no losses). Replay the games. Davide Nastasio interviewed one of the creators of Komodo, Grandmaster Larry Kaufman prior to the tournament.

Book covers

Larry Kaufman has been playing chess for a long time. He is also one of the best western Shogi (Japanese Chess) players in the world. Author of the: "Kaufman opening repertoire," a true opening bible used by many tournament chess players. In this interview he kindly answers some questions upon the new version of Komodo, shedding new light on the Monte Carlo Tree Search, and how to better use the program.

Davide Nastasio: I was listening to a podcast a while back, and the founder of ChessBase mentioned a very interesting and smart idea, nowadays the challenge is not to create a stronger engine, but to create a human one, giving the chance to amateurs to enjoy playing against an engine. Because the truth is: if I play against any engine top strength (3300 but could be also 2900), by move 25 I've lost for some blunders I make 3-5 moves ahead, and I don't see. What's your opinion about it?

Activating Komodo

Larry Kaufman: I think most people buy Komodo to analyze games and/or to help them with their correspondence games, where their opponents will also use top of the line software. For playing against the engine, there are basically three options:

1. Use ancient software on weak hardware.
2. Get Komodo, and set it to one of the (currently) 20 levels below the unlimited one

The current estimated Elo range for those levels is 800 to 3000 for ten minute games. We make an effort to weaken the levels in ways that are at least somewhat human, in that both the tactical and positional strength deteriorate as the level goes lower. Of course some other engines also offer many levels, we just think we do it better!

Rated game screen

3. Play Komodo at full strength with a material handicap, like b1 knight or a1 rook. Curiously, this works better for stronger players than for weaker ones, because if you need much more than rook odds, the engines don't really know what to do and play rather poorly. On the high side, I feel confident in saying that there is no human on earth with the possible exception of Magnus Carlsen who could win a ten game match from Komodo at f7 pawn handicap in 45' + 15" games, and even Carlsen would lose such a match at f2 pawn handicap. This is based on extensive match experience between Komodo and GMs with handicaps. Personally I find Komodo to be an excellent opponent for me with b1 knight odds at 5' + 5" level. I'm no longer near GM level at age 71, but I haven't forgotten how to play!

The Fritz GUI has an handicap menu for playing against the engine with odds


Here I'll pause the interview, because the average reader may not be informed, but throughout the years, with different versions of Komodo, there have been a number of different matches, with different top grandmasters, Hikaru Nakamura being one, which have played matches against Komodo with different types of handicap, be they material, or tempi (moves).

Here some of those games I selected, there are many more:

 

DN: Always in this line of questioning: Often a drawish endgame position is reached, and a human would concede the draw, because he understands there is no reason in wasting 20 minutes playing for the 50 moves rule, instead most engines refuse the draw. What's your opinion about it? Should this be addressed by programmers?

LK:  I think that the popularity of games with rather short increments (relative to the base time) makes this not so important, but at least the NN [neural network] engines should find a way to stop trying to make the games go as many moves as possible! Komodo MCTS doesn't do that at least. If engines could reliably tell which unequal endgames are hopelessly drawn they would gain Elo points by seeking or avoiding them as appropriate, so what you are asking for goes well with making engines stronger.

I always dreamed of trying to discover the best moves over a certain opening I use, and my dream would be to go to work in the morning, let the engine/computer work, return home in the evening, and the engine should have created an opening library which maybe goes 20 moves deep, and I can consult or add my comments. Does this new Komodo help to do that?

We believe that Komodo 13 MCTS is especially good at "MultiPV", so if you want it to analyze a long time and display the best lines of play for the top five (or even all) moves, it is arguably the best CPU based engine for that task. We also have a lot of evidence that it gains in strength with longer time limits relative to standard engines.

Would this Komodo version be able to improve itself? For example after losing a game to analyze it, discover the mistake and avoid it the next time the engine plays the same game?

No, it does not have that feature. It wouldn't be allowed in rating lists and most tournaments, and for games with humans it will probably never lose a game at full strength. So it's not a feature many people ask for.

Different programs use clearly distinctive "styles"; why is that?

Probably it's mostly that the relative weight put on static features like material and pawn structure versus the weight put on dynamic features like mobility and king attack potential varies from engine to engine, so some will play more like Tal and some will play more like Karpov.

Author Cyrus Lakdawala suggested I ask: In what respect are the program's move choices human?

All the features of the engines' evaluation function have been based on how some human (in the case of Komodo, me) thinks they should be defined. The weights were originally my subjective ones, but gradually got "tuned" by testing and automated methods. So in theory, if the search depth were the same as mine (which of course is variable so this is unrealistic) it should play somewhat like I do, although I don't add up hundreds of numbers in my head when I play, I just estimate everything. The main reason Komodo is a thousand or more Elo stronger than I am is that it searches so much further ahead in nearly every line than I could possibly do. Perhaps if I spent 24 hours per move on a game (moving pieces around freely, but not consulting any engine) I could play as well as Komodo plays a blitz game, but that's just a guess. Note that this does not apply to NN engines, which (in pure form) don't have any human knowledge input. 

Also, many engines pitted against each other seem to wait till the very last minute and blitz the endgame. Why is this the optimal strategy?

It depends on the time limit of course, but generally speaking most games are pretty much decided by the endgame, so it may be too late to think by then. It depends on their being enough increment for the engine to still play pretty high quality in the endgame. Having endgame tablebases makes this strategy even more appropriate, since those take almost no time to access.

The advertisement of the new Komodo says: "With AlphaZero technology" but AZ used some special GPU which were likely very expensive made by Google, some say the DeepMind team spent between 4 and 25 million dollars for the AlphaZero project. So is the new Komodo using the GPU of the computer on which is mounted? Or the normal CPU?

Just the CPU. We use the MCTS part of AlphaZero, which is done by CPU, but not the NN part, which needs a GPU to be fast enough to compete with CPU engines.



If you're interested in a GPU-nased neural net engine, check out Fat Fritz, the newest addition to the ChessBase engine cloud!




I was reading a chess book by an IM on using engines to analyze one's own games, and he said one should always let the engine reach depth 27 in order to have a reliable answer. What's your opinion about it?

Well, that was written before MCTS engines were relevant. The "depth" for MCTS engines is arbitrary and is not comparable to normal engine depths. I would say now that time is a more useful metric than depth, as time has the same meaning for any engine.

In which way is best to use this new version of Komodo, is there a depth one should reach?

Again, with Komodo MCTS (which I recommend for use when looking for two or more lines of analysis, not just the best one), time is a better guide. It needs at least a few seconds to produce useful analysis, and of course the quality goes up with more time. Once you see the evals stabilize with just small fluctuations you can probably move on to the next task, but there will always be positions where a long time is needed to discover something.

Another point I think it's missed by amateurs and professional alike is the use of "rating" for engines. For example let's say an engine is rated 2900, one could think the engine is strong like Magnus, but my understanding is that also a 2900-rated engine can calculate a checkmate in 12 moves in a fraction of a second, while Magnus surely can't. Maybe he can feel it, but doesn't mathematically know it like an engine. So I believe human ratings are "subjective" while engine's ratings are relative only to other engines, since they don't commit the blunders humans play. Is this idea correct?

I think that there is some effort made by most rating lists to display ratings that are reasonably realistic in FIDE terms. The CCRL 40/40 list is probably the one that is closest to estimating what ratings the engines would have in FIDE competition if allowed. A 2900-rated engine on that list won't make the oversights Magnus would make, but it also won't have his level of judgment, so in theory they should be fairly matched at standard time controls. The way you can check this is by looking up ratings for ancient engines that actually played drawn matches with Kasparov and Kramnik near the dawn of the current century. But there are a lot of variables like hardware and opening books, so we can't be too precise. Overall I would say that engines rated below 2900 on that list would be good matches for similarly rated humans, and that engines rated 3400 for example can beat those 2900 engines by huge enough margins to justify their ratings. Of course if a 3400-rated engine played in a human tournament, someone would have to create a special opening book for it to avoid draws as Black.

The latest TCEC ended few days ago. Leela won, but some people say the win was controversial because there were some crashes for the Stockfish hardware. The question which came to my mind is the following: finding a sponsor would the Komodo team be interested in a public match against Leela?

Komodo isn't really competitive with Leela when Leela is running on $1000 GPUs. I think that if both engines are required to run on a $1000 laptop for example, which might have a $100 to $200 GPU, at a standard time limit, it could be an interesting match.

When analyzing some of the games by AlphaZero some players found the same amazing moves played by AlphaZero with a depth around 35. If finding the right move, or the "revolutionizing move" is just a question of depth (because the horizon effect could trick an engine into considering a move good at depth 25 and bad at depth 35) then wouldn't the difference between AlphaZero and other engines just based on how good the hardware is?

There are some real differences. Basically AlphaZero and Lc0 (Leela) and Komodo MCTS try to find moves that set problems for the opponent, rather than just assuming that he will see everything they see. Also the NN engines "know" things that are too complex to teach to a standard engine, we don't even know what they are.

Since this is the last question, I'd like to ask about a different topic: ChessBase 15 but also previous versions are wonderful for chess lovers because allow to comment and store endless number of games. I know the market is quite smaller, but is there any chance you guys could produce something similar for Shogi or XiangQi (Chinese chess). Maybe also Komodo could be adapted for those two games.

The late Don Dailey and I already made a Komodo Shogi back in or around 2012, but we only spent two weeks on it so it wasn't competitive with existing programs. As the top ranked American-born shogi player I would like to work on shogi, but breaking into the Japanese market with the language barrier might be challenging. I don't know the status of database programs for shogi now. I know less about the XiangQi situation, although I was a decent player long ago.


I'd like to thank GM Larry Kaufman for being so detailed in his answers, and his partner dealing with the programming part of Komodo: Mark Lefler who also helped me with this interview.

ChessBase 15 - Premium package

The very best: the ChessBase package for professional standards. With Mega Database 2019, new endgame turbo on a USB stick and Corr Database 2018. Plus ChessBase Magazine (DVD + magazine) and CB Premium membership for 1 year!

ChessBase is a personal, stand-alone chess database that has become the standard throughout the world. Everyone uses ChessBase, from the World Champion to the amateur next door. It is the program of choice for anyone who loves the game and wants to know more about it. Start your personal success story with ChessBase and enjoy the game even more.

Links


Davide is a novel chess aficionado who has made chess his spiritual tool of improvement and self-discovery. One of his favorite quotes is from the great Paul Keres: "Nobody is born a master. The way to mastery leads to the desired goal only after long years of learning, of struggle, of rejoicing, and of disappointment..."

Discuss

Rules for reader comments

 
 

Not registered yet? Register