Komodo 12 and MCTS

[lkaufman]

We do it the normal MCTS way (non-min-max) except for one small case in the tree.

You don't average mates or draws-by-rule, I presume

they use a neural network, which Komodo doesn't. MCTS has nothing to do with it.

Well, yes and no. There's reasons why people tried MCTS with playouts and then later neural networks, over alpha-beta. You can't entirely decouple those concepts. If it was so easy, they would not lose 330 Elo.

But that doesn't mean mixing them up won't work. For sure a neural network evaluation in an alpha-beta searcher works fine. As for the other way around, that's up to the Komodo guys to prove, right.

My best guess is that Komodo MCTS won't surpass normal Komodo for engine vs engine play, but that it will surpass normal Komodo when both are asked to display the best five moves and evals, and might also surpass normal Komodo vs. humans. But this will talk time.

[Gian-Carlo Pascutto]

MCTS or Neural Network, I couldn't care less ; if Komodo CAN'T use the GPU (or TPUs), its simply not upto snuff, and that's all there is to it.

So Stockfish isn't up to snuff? Despite beating all NN engines handily?

(Aside from the one that was never published and whose results we can't verify anyway)

[Gian-Carlo Pascutto]

Yes, MTCS does average draws and mates. Draws have a 0.5, and giving mate has a 1.0 chance of winning, being mated a 0.0 chance.

So if your MCTS engine finds a position with one move that loses a pawn (0.4) and one that gives mate (1.0), you back up 0.7 as the score?

Can I suggest an improvement to you? Do not do that.

[Milton]

In the one game I played with mcts checked, the engine did not seem to use syzygy tablebases. Is that a feature or a bug?

[Daniel Shawul]

Yes, MTCS does average draws and mates. Draws have a 0.5, and giving mate has a 1.0 chance of winning, being mated a 0.0 chance.

So if your MCTS engine finds a position with one move that loses a pawn (0.4) and one that gives mate (1.0), you back up 0.7 as the score?

Can I suggest an improvement to you? Do not do that.

I did this with _raw_ score where mate was 20000 centi-pawns and it was horrible but "different" .
Doing it with scaled scores (0-1) alleviated the problem a bit but the problem is still there because a 0.99999 score will translate back to a 5000 centi-pawns score, so in positions where you have a single escape move the averaging sucked. It would extremly exaggerate its
attacking chances in such cases, and suddely looses it in the next move . On the other hand a minimax backup MCTS reports
steady scores that are not exaggerated from move to move just like alpha-beta.

Daniel

[Werewolf]

My best guess is that Komodo MCTS won't surpass normal Komodo for engine vs engine play, but that it will surpass normal Komodo when both are asked to display the best five moves and evals, and might also surpass normal Komodo vs. humans.

I hate to say it but...why bother then? Engines already kill humans easily. There's not much to gain by seeing if the engine can do it with an even larger handicap.
Much more interesting is if a stronger entity can be produced, either outright, or by combining with a regular engine. One way forward would be if MCTS Komodo could handle closed positions better and the algorithm could automatically switch in those positions to MCTS.

[mjlef]

Yes, MTCS does average draws and mates. Draws have a 0.5, and giving mate has a 1.0 chance of winning, being mated a 0.0 chance.

So if your MCTS engine finds a position with one move that loses a pawn (0.4) and one that gives mate (1.0), you back up 0.7 as the score?

Can I suggest an improvement to you? Do not do that.

LOL well technically the 1-0.4 = 0.6 gets backed up, then 1.0 - 1.0 = 0 gets backed up, which makes an average for the parent node of 0.3. But the way nodes get selected for expansion, the 1.0 node would get selected most of the time, meaning the average backup up win percentage for the parent node would be rather close to 0.0. MCTS looks silly with a few expansions, but much smarted with thousands of them. But certainly a refinement would be if a certain win was found for one node, cheat a little and select it even more.

[shrapnel]

My best guess is that Komodo MCTS won't surpass normal Komodo for engine vs engine play, but that it will surpass normal Komodo when both are asked to display the best five moves and evals, and might also surpass normal Komodo vs. humans.

I hate to say it but...why bother then?

My thoughts EXACTLY...heh heh.
As I wrote earlier, the whole thing is a marketing gimmick with no real substance and with the above Statements Kaufman practically admits it !

[mjlef]

In the one game I played with mcts checked, the engine did not seem to use syzygy tablebases. Is that a feature or a bug?

We have not decided how to incorporate the Syzygy scores into the MCTS in the best way yet. In regualr searches, Syzygy wins just get a single number, which is not accurate for a win probability search scheme. I am likely to change that with a future update. Note the Syzygy probes are used in some of the searches used in calculating win probabilities for the MCTS tree. So they do effect it, just in a more indirect way. And Syzygy is used at the root to decide the nodes worth searching.

[Daniel Shawul]

In the one game I played with mcts checked, the engine did not seem to use syzygy tablebases. Is that a feature or a bug?

We have not decided how to incorporate the Syzygy scores into the MCTS in the best way yet. In regualr searches, Syzygy wins just get a single number, which is not accurate for a win probability search scheme. I am likely to change that with a future update. Note the Syzygy probes are used in some of the searches used in calculating win probabilities for the MCTS tree. So they do effect it, just in a more indirect way. And Syzygy is used at the root to decide the nodes worth searching.

What is wrong with reporting a 0, 0.5 or 1 TB score in MCTS search?? Infact that is the best kind of score you can give to
and MCTS search because one visist to a TB would be enough, while you would have to sample the node many times in other
cases to narrow down the winning probability with enough certainity. I use it TB cutoffs in my MCTS engine just fine.

So much bullshit in this thread.