TalkChess.com

abgursu wrote: ↑Fri Oct 16, 2020 10:05 pm

amchess wrote: ↑Fri Oct 16, 2020 10:02 pm https://github.com/amchess/BrainLearn/releases/tag/12.1

Why removing montecarlo?

Not well working in some guis and, in general, not fully mature.
Moreover, as demonstrated by the latest tcec, even on a gpu, MCTS is slower than classical alpha/beta.

amchess wrote: ↑Sat Oct 17, 2020 1:03 am

abgursu wrote: ↑Fri Oct 16, 2020 10:05 pm

amchess wrote: ↑Fri Oct 16, 2020 10:02 pm https://github.com/amchess/BrainLearn/releases/tag/12.1

Why removing montecarlo?

Not well working in some guis and, in general, not fully mature.
Moreover, as demonstrated by the latest tcec, even on a gpu, MCTS is slower than classical alpha/beta.

But that was only an option. It was not the default. Some users like me wants to play with mcts, even if it's not the best. I believe you shouldn't remove montecarlo just because someone can't use it or lc's lost against stockfish.

abgursu wrote: ↑Sat Oct 17, 2020 6:57 am

amchess wrote: ↑Sat Oct 17, 2020 1:03 am

abgursu wrote: ↑Fri Oct 16, 2020 10:05 pm

amchess wrote: ↑Fri Oct 16, 2020 10:02 pm https://github.com/amchess/BrainLearn/releases/tag/12.1

Why removing montecarlo?

Not well working in some guis and, in general, not fully mature.
Moreover, as demonstrated by the latest tcec, even on a gpu, MCTS is slower than classical alpha/beta.

But that was only an option. It was not the default. Some users like me wants to play with mcts, even if it's not the best. I believe you shouldn't remove montecarlo just because someone can't use it or lc's lost against stockfish.

+1 I totally agree. Many of us use the engine for analysis and don't always care about the most Elo.

Has anyone had good experiences with this engine learning and improving itself? I haven't.

https://github.com/amchess/BrainLearn/wiki/Tests


An interesting result:
BrainLearnPro is BrainLearn opensource file + a bin experience file improved by a private application.
Stockfish uses probably the best ctg currently available.
The match has been run on a quadcore.
The demonstration, imho, nnue engines + learning can also ignore a book.

Interesting indeed. What was the size of the experience.bin ?

About 25MB for now.
The best would be to analyze all eco codes with Q-learning and standard learning + bin experience file improvement.
A complete rewrite of all opening theory...

amchess wrote: ↑Sat Dec 05, 2020 10:23 pm About 25MB for now.

Wow that's not much for a so good result ! I already managed to have an experience bin of 40MB with only 3300+ top elo engine games PGNs

A friend of mine that likes Brainlearn engine a lot wants to ask if there will be soon some update of it or none is planned ?

Thanks