2

The new Komodo 13 MCTS has been announced with the claim that:

Komodo 13 MCTS thinks like no other chess engine in the world. Inspired by the ideas of AlphaZero and Leela, Larry and Mark set out to reinvent Komodo. What they came up with was Komodo 13 MCTS - a engine which searches for candidate moves in an incredible new way, and finds moves no other engine can

They are clearly trying to cash in on the publicity behind AlphaZero and LeelaZero. But is the hype justified?

There seems to be no mention of an AlphaZero like database of training results. Is the use of MCTS on its own enough to produce worthwhile improvements? More detailed information seems to be hard to find (possibly for commercial reasons).

Brian Towers
  • 96,800
  • 11
  • 239
  • 397
  • Are there any games published with this new version of Komodo? Otherwise there is no way for an outsider to tell whether these claims are valid or not. – Scounged May 17 '19 at 14:48
  • 1
    They are promoting their product, they need to explain in what way their engine is different, as it stands the quoted paragraph is just marketing speak, and not very good one. – Akavall May 17 '19 at 16:26
  • Does it make any difference if Komodo MCTS is on infinite analysis or a timed game for decision making assuming all moves were made with the same thinking time? I’m wondering if infinite analysis will only look for the best score instead of winning chances since it’s able to see possibilities of opponent mistakes. – Ryan Oct 08 '19 at 21:23

3 Answers3

4

Komodo MCTS isn't a neural-network based engine. It just uses a different search algorithm, in particular MCTS instead of AB. Alphazero also used MCTS, so KMCTS is an "Alphazero lite" in that sense; however a distinguishing factor of Alphzero is that it uses a neural network to perform the evaluation, and KMCTS does not do that.

Is MCTS "better" than AB? Yes and no. From what I've seen, MCTS is exceptionally better when you want multiple PVs (because they effectively get it for free, while AB engines need to spend computational power). However MCTS loses firepower in tactical shoot-outs, and it's also really bad at winning a technical position. For example I remember seeing a KMCTS game in TCEC where its eval read "I have a forced checkmate" one move and then "wait, I don't, I only have a very large advantage" another, and then back to "I have a forced checkmate", etc.

Having said all that, KMCTS eliminated regular Komodo from the latest TCEC cup, so it's absolutely a respectable engine. In a straight-up game against another engine, it's not as strong as Stockfish or Leela, but it's comparable to everything else.

Allure
  • 26,534
  • 1
  • 68
  • 146
3

Komodo doesn't use machine learning, it just uses roughly the same search algorithm that A0/L0 uses. It still relies on the human-built heuristics it's used.

Inertial Ignorance
  • 19,818
  • 23
  • 70
  • 3
    Is there any benefit to using MCTS without deep learning or is this just a marketing gimmick? – Qudit May 17 '19 at 19:24
  • @Qudit That's also what I want to know! – Brian Towers May 17 '19 at 20:35
  • 1
    I had thought that the two techniques depended on each other to work well. I do not know for sure, but using one without the other seems suspicious to me. – Qudit May 17 '19 at 21:03
  • Every single chess engine uses machine learning – David May 17 '19 at 23:25
  • 1
    @David No, traditional engines use alpha-beta search which is an adversarial search algorithm from classical AI. It is not considered machine learning and AI and machine learning are not the same thing. – Qudit May 17 '19 at 23:50
  • @Qudit In general I think you're right, for some reason MCTS usually works better with ML. Komodo 12 started using MCTS and it was slightly weaker than standard Komodo 11, so they're still working things out. – Inertial Ignorance May 18 '19 at 00:12
  • 1
    https://chess.stackexchange.com/q/23734/10075 is about why NNs are usually used with MCTS. – konsolas May 18 '19 at 08:52
  • @Qudit I'm pretty sure traditional chess engines also use "machine learning", but it's not machine learning in the conventional sense of the word. In particular traditional engines use gradient descent to tune the parameters used in the evals, e.g. "how much should a hostile pawn two squares away from my king be worth" and "how much should a hostile pawn that attacks a square away from my king be worth" are two related parameters, and gradient descent finds the best combination for them. – Allure May 19 '19 at 00:06
  • @Allure That is interesting if true. Do you have a reference? My quick search didn't come up with anything. – Qudit May 19 '19 at 00:13
  • @Qudit http://talkchess.com/forum3/viewtopic.php?p=772589&sid=411a0a7705d4ddb4578a9925b1c5bdd4#p772589, about tuning parameters in Ethereal. – Allure May 19 '19 at 06:45
  • @Qudit Every single engine built in the last decade uses some form of machine learning. Don't assume that just because someone does not agree with you, it automatically menas he does not know what he's talking about – David May 19 '19 at 14:12
3

There are a number of benefits to using MCTS instead of alpha-beta search, mainly in the realm of analysis.

Firstly, Alpha-Beta Search only attempts to find the best move in a position. As a result, MultiPV is implemented in alpha-beta engines by searching all the root moves, then searching the root again, skipping the root move, then searching again skipping the two already-found root moves, etc.

This significantly reduces the strength of alpha-beta engines in analysis, where you often want to take a look at multiple possible variations. MCTS, by contrast, expands out portions of the whole tree making the PV of every move considered by MCTS available. There is therefore no strength hit when enabling MultiPV for MCTS engines, and Komodo MCTS can therefore be much stronger than alpha-beta engines when MultiPV is set to high values.

With regards to the results of the search, MCTS has the advantage of being able to reflect the possibility of the opponent going wrong, as it assumes that all reasonable moves have some chance of being chosen. By contrast, alpha-beta engines always assume that the opponent will play the move that the engine considers best.

As a result, alpha-beta engines can often evaluate situations which are deadly tactical landmines to be dead draws, whereas Komodo MCTS can take this into account, further improving its performance in analysis of human games.

konsolas
  • 3,176
  • 15
  • 32
  • This is interesting and the argument about multiple principle variations makes some sense. However, the reference you provided is to a forum post announcing Komodo 13 by Larry Kaufman who is affiliated with Komodo. As such, one wonders how reliable his claims about his own product actually are. – Qudit May 18 '19 at 22:44
  • I see no particular reason to lie about this point. Unless the use of MCTS is disputed, the ability to use MultiPV at no cost follows logically from what MCTS is. Other MCTS engines like Lc0 are similarly able to use MultiPV at no cost to strength. – konsolas May 19 '19 at 10:12