Chessbase 26 Monte Carlo Analysis Function

[CornfedForever]

I upgraded from CB18 to this the other day. The Monte Carlo (MC) analysis function actually improves over the previous rather standard (although 'buddy' could be useful) quick position analysis options. Has anyone had a chance to check that out? I've had some home issues to deal with but dabbling a bit I found it useful at critical points in some OTB games I played recently. I also had it check out a 'novelty' I invented in a sharp gambit opening and it found my idea in the MC analysis as best to give my opponents more reasonable problems to deal with OTB than they would with the regular Stockfish main move.

Basically, while you have your engine of choice running, you hit a button whick starts separate Monte Carlo analysis. You can set the number of MC threads and and the speed for it to play thousands of games - Quick/Standard/Slow - from the position. Very quick on my AMD processor

[peter]

Has anyone had a chance to check that out?

I gave it only a few trials so far, guessing it's (like "Buddy") only a feature of it's own again without any real connection to the search of the engine- partition doing main search. Shared hash in its true meaning (both partitions of the engine dealing with same hash- tables at all) again (neither is was the case with Buddy, as far as I know) isn't to be expected, is it?

I wrote an email to support already about that question, there isn't much to be read in reference book about it, especially the amount of hash to be given would be good to know.
As quickly as the MC- games are played, minimum of hash would be reasonable, if you wait for some time (not seconds but minutes) you'd give much more hash to the main search SMP, so what, count full time to ponder like with "normal" search as for bigness of hash or otherwise? Could well be, hash for MC- partition is minimum part of fully given amount anyhow, but I don't think so, because with big hash and MC used, GUI gets rather sluggish as for responding to moves made on the board, sometimes it even gets instable, so at the moment I'd say, minimum hash (8 to 64Mb maybe?) is better, at least safer. Threads of main search and of MC- partition can be set separately by user, hash cannot.
Edit: with 64Mb hash only, GUI gets sluggish sometime yet too with biggest part of CPU- threads given to MC and only smallest one to main search (28 to MC and 2 to main search out of 32 threads of CPU existing), so it seems to be rather a question of CPU- load but of hash, or 64Mb hash is still too much for MC- search

Will return, when I have answer from support, regards

[gordonr]

I've not upgraded to CB 26 - still on 18 here. I did recently have a play around with the Monte Carlo analysis in Fritz 19. I wonder how it compares?! I had to use an engine that supports MC and used Rybka - Stockfish didn't work for me in Fritz 19 but I saw the demo video for CB 26 used Stockfish.



I know engines with multiple threads are non-deterministic in their analysis, but for a single thread engine that is deterministic, does anyone know how MC avoids repeating the same game? Is it building a tree based on candidate moves and then evaluating the leaf nodes by quickly playing the game out? And doing one game per new leaf node?

In Fritz 19, the MC analysis highlights what the starting positions are. I've also had a look at the tree it builds but the process still wasn't clear to me.

I like the idea the MC analysis can give an indication of how easy or difficult a win is, even when Stockfish claims something like +3. I'm curious as to how this compares to LC0 WDL analysis with contempt. Any good example positions for us to compare CB 26, Fritz 19 / Rybka and LC0?

[peter]

I like the idea the MC analysis can give an indication of how easy or difficult a win is, even when Stockfish claims something like +3.

Well, yes, on the first glance, and Mr. Wüllenweber says so too, but on the other hand it proves my "suspicion", that the analysis of main search hasn't any form or communication with the one of the MC- partition, in the video (thanks for the link, hadn't seen that till now) even more, because MC is run on a remote machine on the contrary to main search.

So, yes, even if the idea of such "raw" Monte Carlo mode that already Rybka- Fritz had, isn't new, it's one more funny toy of the new GUI, the usefullness, I could imagine of some kind of influence of one part of the search (main one) to the other one (MC), at least as for some kind of communication between the lines of the one and the other one (like cb-GUI "Deep Analysis"- feature with more then one engine is said to have) would give some more sense to me and would raise my interest much more. Matthias Wüllenweber points out, that there are first 3 plies of the output of main search and MC- search in common after a while of pondering, but I guess, that's more or less arbitrarily so. If MC- games start with other one first move(s) of output then given in main search and the ones in main line are more or less forced and more or less single best ones, you can probably not hope for any meaning of the statistics of the MC- search, if any of the first forced but tactically somewhat tricky plies aren't "found" by MC- partition too. And even in the given example in video, an eval of >+3 pawns in main search and an only slightly > 50% WDL- "eval" for White of MC- output doesn't make me really happy as for the "more information" I get by this kind of MC- search other but "maybe won but not very easily""

Shared hash in its literal meaning was to be got so far by Rybka- Fritz only (not the one of F15 but the much sooner one, named Rybka as for the version of the GUI then), there you could start different partitions of Rybka and the hash of the one partition was shared with the other one.
MC- analysis of that kind with Fritz- GUIs and their native engines (starting with Rybka- Fritz way back then) might be more or less the same kind of mode, Rybka- Fritz (and later one F15 ff, being Ryba too) does, only difference that now other engines can be used for it too.
As far as I saw in the video (at most part rather quickly and sometimes fast forwarding), best amount of hash to be set for the new feature isn't mentioned neither, regards

[Odd Gunnar Malin]

I upgraded from CB18 to this the other day. The Monte Carlo (MC) analysis function actually improves over the previous rather standard (although 'buddy' could be useful) quick position analysis options. Has anyone had a chance to check that out? I've had some home issues to deal with but dabbling a bit I found it useful at critical points in some OTB games I played recently. I also had it check out a 'novelty' I invented in a sharp gambit opening and it found my idea in the MC analysis as best to give my opponents more reasonable problems to deal with OTB than they would with the regular Stockfish main move.

Basically, while you have your engine of choice running, you hit a button whick starts separate Monte Carlo analysis. You can set the number of MC threads and and the speed for it to play thousands of games - Quick/Standard/Slow - from the position. Very quick on my AMD processor

I too bought it, even on the first day, together with the Megadatabase.
I've only tried it in even endgame where I agreed to a draw, to check if it was too early. I also checked some games where I went into SLP mode to see if this was too early, or too early to resign in some games.
I don't know which engine it uses for this MC, it looks like it's an engine inside CB26, and none of your installed engines. It seems to do a good job in trying out different variations if you look at how many times it has played the 'best' line.

[gordonr]

I don't know which engine it uses for this MC, it looks like it's an engine inside CB26, and none of your installed engines.

That's an interesting point. For the demo video, although I wasn't assuming any interaction between the Stockfisk main analysis and the MC analysis, I did for some reason assume both were using Stockfish - maybe because I didn't hear any other engine being mentioned. But as you say, maybe MC is using something else.

[peter]

I don't know which engine it uses for this MC, it looks like it's an engine inside CB26, and none of your installed engines.

For the demo video, although I wasn't assuming any interaction between the Stockfisk main analysis and the MC analysis, I did for some reason assume both were using Stockfish - maybe because I didn't hear any other engine being mentioned.

In the video the "remote engine" is used, as for that you can chose between F17.1, Lc0 and F20 offered in cb26 in cloud. But of course you can use any other one engine locally or remotely as well, so I'd say, whichever engine you choose, locally or remotely, the MC- partition (what sense otherwise would make the splitting of the threads as well?) is chipped from the engine loaded for main analysis, and of course you can have several kiebitz- engines splitted that way also. At least that all would be my educated guess

[Odd Gunnar Malin]

I don't know which engine it uses for this MC, it looks like it's an engine inside CB26, and none of your installed engines.

For the demo video, although I wasn't assuming any interaction between the Stockfisk main analysis and the MC analysis, I did for some reason assume both were using Stockfish - maybe because I didn't hear any other engine being mentioned.

In the video the "remote engine" is used, as for that you can chose between F17.1, Lc0 and F20 offered in cb26 in cloud. But of course you can use any other one engine locally or remotely as well, so I'd say, whichever engine you choose, locally or remotely, the MC- partition (what sense otherwise would make the splitting of the threads as well?) is chipped from the engine loaded for main analysis, and of course you can have several kiebitz- engines splitted that way also. At least that all would be my educated guess

Well, with process explorer it shows that all CPU given to MC are in CB26 itself. When using a remote engine for analysis, no other engines are loaded. Using a local engine, the CPU is split between CB26 and the local engine according to your settings for CPU usage.

[peter]

Well, with process explorer it shows that all CPU given to MC are in CB26 itself. When using a remote engine for analysis, no other engines are loaded. Using a local engine, the CPU is split between CB26 and the local engine according to your settings for CPU usage.

What do you mean by "no other engines are loaded"? Using a remote engine, the remote engine is loaded and of course it doesn't show processes in local task manager because these processes run on the remote- machine, whether your own one on a second computer of yours or a rented one at cb-server. Anyhow the MC- analysis (or several MC- analyses, if you start several kiebitzes doing such) is done by the same engine that's showing its own main search also, don't you thinkt so too?

[Odd Gunnar Malin]

Well, with process explorer it shows that all CPU given to MC are in CB26 itself. When using a remote engine for analysis, no other engines are loaded. Using a local engine, the CPU is split between CB26 and the local engine according to your settings for CPU usage.

What do you mean by "no other engines are loaded"? Using a remote engine, the remote engine is loaded and of course it doesn't show processes in local task manager because these processes run on the remote- machine, whether your own one on a second computer of yours or a rented one at cb-server. Anyhow the MC- analysis (or several MC- analyses, if you start several kiebitzes doing such) is done by the same engine that's showing its own main search also, don't you thinkt so too?

Sorry, I can't discuss with people who can't or will not understand what I'm saying.