Komodo 12.2: Regular vs MCTS

[leavenfish]

This has probably already been noted somewhere here - and if anyone knows of the proper thread, I think that might suffice, can't find it in a search - but for the current version of Komodo (my Sub ended with 11.3) which would be better for this use:

Using Chessbase (or Chess Position Trainer - great for storing a repertoire!!) to go thru your openings and letting the computer 'chew on a position' for a while to come up with what it considers the best move...preferably in say multi-pv = 2 or 3...and usually I carry that out a few moves, look at the new evaluation and and only then decide on updating my repertoire(s).

I say that because it is my standard use, but also notice the website says (without fleshing this out): While the MCTS version is not yet as strong as normal Komodo (unless you are using MultiPV = 3 or more)

Hardware, pretty normal for your average enthusiast:

CPU: Intel i7-6700 @ 3.4 GHz
RAM: 16 GB

[mjlef]

This has probably already been noted somewhere here - and if anyone knows of the proper thread, I think that might suffice, can't find it in a search - but for the current version of Komodo (my Sub ended with 11.3) which would be better for this use:

Using Chessbase (or Chess Position Trainer - great for storing a repertoire!!) to go thru your openings and letting the computer 'chew on a position' for a while to come up with what it considers the best move...preferably in say multi-pv = 2 or 3...and usually I carry that out a few moves, look at the new evaluation and and only then decide on updating my repertoire(s).

I say that because it is my standard use, but also notice the website says (without fleshing this out): While the MCTS version is not yet as strong as normal Komodo (unless you are using MultiPV = 3 or more)

Hardware, pretty normal for your average enthusiast:

CPU: Intel i7-6700 @ 3.4 GHz
RAM: 16 GB

In regular search mode, when you set say MultiPV = 3, the engine has to effectively search 3 times (one for each move). Of course, the second and third searches should take less time since they progressively have fewer moves to consider. But it does take a lot longer than finding the best move. In MCTS mode, since the program is creating a search tree in memory, with the nodes linked, it does not take extra time to find the second and third best move (well other than figuring out what they are based on visit count and win percentage). So MultiPV is "free" in MCTS mode. Of course, the way MCTS grows the search tree is to visit and expand the best performing moves/nodes more, so lower scoring moves will not have as large a tree under them which you can see by looking at the PVs. I suppose humans do the same thing, search the lower chance of winning moves less deeply than the more promising ones.

[lkaufman]

This has probably already been noted somewhere here - and if anyone knows of the proper thread, I think that might suffice, can't find it in a search - but for the current version of Komodo (my Sub ended with 11.3) which would be better for this use:

Using Chessbase (or Chess Position Trainer - great for storing a repertoire!!) to go thru your openings and letting the computer 'chew on a position' for a while to come up with what it considers the best move...preferably in say multi-pv = 2 or 3...and usually I carry that out a few moves, look at the new evaluation and and only then decide on updating my repertoire(s).

I say that because it is my standard use, but also notice the website says (without fleshing this out): While the MCTS version is not yet as strong as normal Komodo (unless you are using MultiPV = 3 or more)

Hardware, pretty normal for your average enthusiast:

CPU: Intel i7-6700 @ 3.4 GHz
RAM: 16 GB

Actual tests at fast time limits suggest that regular Komodo is still stronger than MCTS with MultiPV=2, but MCTS is stronger with MultiPV=3 or more. But this is not the whole story. In general, regular Komodo is likely to be better for correspondence players, while MCTS is likely to be better for over-the-board players. The reason is that regular Komodo assumes that the opponent is also using Komodo to choose the best move, while MCTS "thinks" that the opponent may well choose a move it considers slightly inferior at any point, which is surely more realistic for humans not using computers themselves.