Komodo 12 and MCTS

[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.

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.

I know TB scores can be used. But we want the search to use scores likely to lead to the fastest mates. Syzygy does not contain that information. It just has Win/Loss/Draw and distance to conversion, so I have not decided the best way to integrate it in the MCTS nodes. If distance to win or distance to draw was available like in some other TB formats, it would be simpler. We plan on investigating this more shortly. We think a bunch of things can be improved, and this is on the todo list. I would be surprised if this leads to a huge elo gain, given how MCTS node selection works. But it would be nice to be surprised.

Why the hostility?

[Daniel Shawul]

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 !

There is so much bullshit here. The authors are throwing around ideas to benefit from the A0 hype. E.g.
a) We use averaging => style is much better like A0's
b) MCTS better against humans. I bet standard komodo is much better against computers, humans and cockroaches
c) MCTS better in multi-pv mode, I highly doubt it
d) "Although tactical strength is lower with MCTS, positional play and judgment may well be better in many positions". Pure bullshit as they
don't have NN evaluation which account for A0's positional strength. In general, MCTS maybe better strategically but not at the level they are suggesting without the NN
e) "Evals don't fluctuate so suddenly or wildly." So what? That is infact bad since you will have to spend many more simulations to realize
you are screwed with some shallow tactics.

Daniel

[Daniel Shawul]

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.

I know TB scores can be used. But we want the search to use scores likely to lead to the fastest mates. Syzygy does not contain that information. It just has Win/Loss/Draw and distance to conversion, so I have not decided the best way to integrate it in the MCTS nodes. If distance to win or distance to draw was available like in some other TB formats, it would be simpler. We plan on investigating this more shortly. We think a bunch of things can be improved, and this is on the todo list. I would be surprised if this leads to a huge elo gain, given how MCTS node selection works. But it would be nice to be surprised.

Another bullshit. You can't get a mate-in-x scores with MCTS search, so why put such restriction on the TB. Even if you had the perfect Nalimov TB, you won't be able to do mate searches like alpha-beta. I use Scorpio bitbases WDL just fine in MCTS search, and for that matter LCZero with its MCTS search played with Syzygy in TCEC.

Why the hostility?

You don't give straightforward answers but imply here and there that you are doing something non-trivial -- while at the same time
replying with a calculated message inline with the A0's style. I understand now why commerical guys are so annoying to someone just interested in the mechanics of things. If you have no intention of benefiting from the MCTS as you so claimed, why don't you stop being "nebulous" and answer questions as to your implementation. You don't plan to profit from the MCTS implementation right?, so you might as well explain what you did well to contribute to the MCTS discussion.

[lkaufman]

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.

Thanks. The situation you describe would never happen. But it could happen for example that we have nine instances of finding a long mate (1.0), and one instance of winning a queen (say 0.99). Should we back up 1.0 or 0.999? If you back up 1.0 that's switching to minimax, maybe it is justified but not in the spirit of allowing for fallibility. It probably makes no elo difference, and it is a matter of opinion which way is preferable.

[lkaufman]

[quote=Milton post_id=762253 time=1526323211 user_id=206]
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?
[/quote

Syzygy and several other UCI options don't work or don't work well with MCTS. This is neither a bug nor a feature, but a consequence of the fact that Komodo MCTS is only about a month old. Over time I expect more of these options will work with it. At the moment, top priority is expanding the three thread limit.

[lkaufman]

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.

Well, the multiPV point is a big one for many users. Also, I imagine that eventually Komodo MCTS will be stronger than normal Komodo on some large number of cores, maybe 32 or 48. But I don't know that yet. I do hope that Komodo MCTS will surpass normal Komodo even on say four cores, but I'm not predicting that or betting on it yet.

[Geonerd]

Pointless having MCTS option when its unable to use the power of the GPU or of Tensor Processing Units.
Sorry, but the simple truth is that this is just a Sales Gimmick by the Komodo team trying to take advantage of the current interest in MCTS and Neural Networks.

Christ, dude, we heard you the first time...

WTF IS it with this forum? Why is a civil conversation so damn difficult?

[AdminX]

Pointless having MCTS option when its unable to use the power of the GPU or of Tensor Processing Units.
Sorry, but the simple truth is that this is just a Sales Gimmick by the Komodo team trying to take advantage of the current interest in MCTS and Neural Networks.

Christ, dude, we heard you the first time...

WTF IS it with this forum? Why is a civil conversation so damn difficult?

Well you can now just use the "add foe" option to the profile of the user that does not know how to be civil. Then all post from that user will be hidden from you.

[mjlef]

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 !

There is so much bullshit here. The authors are throwing around ideas to benefit from the A0 hype. E.g.
a) We use averaging => style is much better like A0's
b) MCTS better against humans. I bet standard komodo is much better against computers, humans and cockroaches
c) MCTS better in multi-pv mode, I highly doubt it
d) "Although tactical strength is lower with MCTS, positional play and judgment may well be better in many positions". Pure bullshit as they
don't have NN evaluation which account for A0's positional strength. In general, MCTS maybe better strategically but not at the level they are suggesting without the NN
e) "Evals don't fluctuate so suddenly or wildly." So what? That is infact bad since you will have to spend many more simulations to realize
you are screwed with some shallow tactics.

Daniel

Wow. Just wow. One by one. Larry giving his opinion as a GM about most of these. We never claimed what you said in a. We have just said the resulting move choices are different, not necessarily . The fact that the MCTS mode is 330 elo weaker than the standard mode shows that so far it is not "better". b is simply Larry's opinion, and I admit it is untested. Having known Larry for a long time and following up on his hunches, I would bet on Larry, but time will tell. In all of these we are telling what we know so far and I think it is clear when we are giving opinions. c was tested. In MP mode showing all the moves and PVs at the root, the strength in MCTS mode surprised us by being very close to regular search mode showing all PVs. Perhaps not so surprising since to get the PVs for say 32 moves, you need to run 32 searches. MCTS epands all root moves, although by different amounts based on win performance and visit counts. As for d, where are you showing data saying AlphaZero's NN accounts for its positional strength? You may be right, but is there data or speculation? In any case, Larry speculates a MCTS search improves positional strength, and given that the search depth in MCTS mode is so much lower than in traditional search, I am guessing he is right because a standard search engine seeing so less deeply would be crushed tactically, so we speculate the difference is positional eval.

If you want to know more of what we are doing, do you think calling it a "Marketing Gimmick" is going to be helpful? Our marketing budget is exactly $0. We just have posted what we have, admitting we were time limited, and intend to improve it wit time. How are we going to open up and give you more if you complain that me sticking with a pure MCTS backup scheme on the first release about mates is "BS"? I am kind of surprised you are acting this way, and find it unlike you.

[Tdunbug]

I don't speak much here but I do want to say that MCTS is a great idea and look forward to it being improved. At the very least, it is a fresh new idea that can be tested.