Komodo MCTS

[mjlef]

Hi Mark!

Any plans to implement position- learning to komodo (MCTS)?

Works fine for me in StockLearn, which derived from an idea of Kelly Kinyama, as far as I know, has something to do with kind of MCTS too.

Marco Zerbinati (SugaR_MCTS) and Andrea Manzo (ShashChess and StockLearn) gave GPL- Code of own SF- branches to be found on github easily.

ShashChess seems to just save each final search result as a hash entry that goes into a disk file. Well, at least two disk files (one is used just for K&P endings). The "learning" I see is just saving previous search results, so it would only help should those positions be found in future searches. Basically it appears to be like my old NOW scheme, although split into two files based on material. I only spent a few minutes looking so there may be other things there also I did not notice. These schemes can influence future move choices, but are not general learning like NN MCTS programs can do.

[peter]

Hi Mark!

Any plans to implement position- learning to komodo (MCTS)?

Works fine for me in StockLearn, which derived from an idea of Kelly Kinyama, as far as I know, has something to do with kind of MCTS too.

Marco Zerbinati (SugaR_MCTS) and Andrea Manzo (ShashChess and StockLearn) gave GPL- Code of own SF- branches to be found on github easily.

ShashChess seems to just save each final search result as a hash entry that goes into a disk file. Well, at least two disk files (one is used just for K&P endings). The "learning" I see is just saving previous search results, so it would only help should those positions be found in future searches. Basically it appears to be like my old NOW scheme, although split into two files based on material. I only spent a few minutes looking so there may be other things there also I did not notice. These schemes can influence future move choices, but are not genral learning like NN MCTS programs can do.

Thanks for looking into code. Did you notice, the engine has to have the full move history of positions to be "learned"? If opening moves leading to a position are to be read by engine, there should be 8 .bin-files written to disk for every opening shorter than about 7 moves.
Experience.bin itself does contain useful infos only, if other .bin- files appear at disk also. As far as I understand the idea, there are connections between positions along the lines from opening to endgame evaluated and stored. Kelly Kinyamas ideas behind the principal method are to be found at github and fishcooking too.
BTW, of course I do know, that any engine- learning like this does't have anything in common with the one using neuronal networks, yet it seems a lot less hardware-time consuming as for certain supervised learning.

[marsell]

The engine BrainLearn from Kelly Kinyama produces significantly more files than 8. Nearly 900
What am I doing wrong??

[peter]

The engine BrainLearn from Kelly Kinyama produces significantly more files than 8. Nearly 900
What am I doing wrong??

You're not doing anything wrong. If you read my post carefully, I wrote, 8 files per single different opening- line. As much different openings, as much .bin- files (except experience.bin remaining a single one) x7 + 1 experience.bin.
If you stick to a single line of opening- moves of >7 moves length, there will be 8 files only but constantly growing with the amount of games played with the same one opening.

That's my way of doing it, having separate folders of .bin- files for each and every single opening line (e.g. corr.- game) of interest of mine.

[marsell]

okay, thanks for the clarification.

[sound67]

I played an engine torunament between Komodo 13, Komodo 13 MCTS and Stockfish 10 - on a P5-8400, 2GB Hash, 5 Cores, 40/2h+20/1h+30 mode with the 2017perfect book. 90 games, took almost a month.

Komodo 13 MCTS finished last. It was beaten by its own "regular" version (+0, -5, =25) as well as by Stockfish (+0, -6, =24).

Regular Komodo 13 lost to Stockfish by +0, -9, =21.

From the individual games it looks as if neither version of the new Komodo has any hope to beat Stockfish, not even once.

[Ovyron]

I played an engine torunament between Komodo 13, Komodo 13 MCTS and Stockfish 10 - on a P5-8400, 2GB Hash, 5 Cores, 40/2h+20/1h+30 mode with the 2017perfect book. 90 games, took almost a month.

Komodo 13 MCTS finished last. It was beaten by its own "regular" version (+0, -5, =25) as well as by Stockfish (+0, -6, =24).

Regular Komodo 13 lost to Stockfish by +0, -9, =21.

From the individual games it looks as if neither version of the new Komodo has any hope to beat Stockfish, not even once.

Now try the same match at MultiPV=4 (as some people use in analysis)

[Modern Times]

Now try the same match at MultiPV=4 (as some people use in analysis)

I've tried it, no noticeable difference in performance.

[sound67]

Now try the same match at MultiPV=4 (as some people use in analysis)

I used the maximum number of cores for the engines that my P5 processor allowed, MultiPV=5. Ponder=Off.

[lkaufman]

Now try the same match at MultiPV=4 (as some people use in analysis)

I used the maximum number of cores for the engines that my P5 processor allowed, MultiPV=5. Ponder=Off.

Did you set MultiPV = 5 for ALL engines, not just Komodo? With that setting, Komodo 13.02 MCTS is MUCH stronger than normal Komodo, and about the same strength as Stockfish 10 (also using MultiPV). If you are getting results with Komodo MCTS below normal Komodo, something is wrong.