BrainLearn

[amchess]

Finally, again public!

https://github.com/amchess/BrainLearn

This is an article explaining our idea:
https://github.com/amchess/BrainLearn/b ... A%20RE.pdf

Have fun!

Kelly and Andrea

[zullil]

Finally, again public!

https://github.com/amchess/BrainLearn

This is an article explaining our idea:
https://github.com/amchess/BrainLearn/b ... A%20RE.pdf

Have fun!

Kelly and Andrea

Forgive me, but it seems all you've done is add a "permanent" hash table. This has been done many times before, even for Stockfish. How is this different? I mean, who would want to publish this?

[amchess]

Corrected a bug on opening variety
https://github.com/amchess/BrainLearn/r ... /tag/4.0.1

Andrea

[Leo]

How do you download the engine? I cant find it. Its not in the src.

[Graham Banks]

How do you download the engine? I cant find it. Its not in the src.

https://github.com/amchess/BrainLearn/r ... Windows.7z

[Leo]

Thanks Graham.

[Michael Sherwin]

Finally, again public!

https://github.com/amchess/BrainLearn

This is an article explaining our idea:
https://github.com/amchess/BrainLearn/b ... A%20RE.pdf

Have fun!

Kelly and Andrea

Forgive me, but it seems all you've done is add a "permanent" hash table. This has been done many times before, even for Stockfish. How is this different? I mean, who would want to publish this?

I agree!

I see no evidence of RL. It is okay to store game moves (not search moves) in a hash table on disk instead of storing them in a linked tree like I did in RomiChess. It probably would have been better if I had used a hash instead. However, for there to be RL as is claimed there has to be a small bonus or penalty added to each game position in the hash after every game.

[peter]

Have fun!

Thanks, Andrea!

Works without move- history (opening moves leading to position of interest) now too, that's fine!

[peter]

Yet as for difficult tactical positions time to solution doesn't get as much better, as one is used to see with old-fashioned learning files like the ones, e.g. SF-PA had.
Especially for Zugzwang- studies it's still some kind of luck, if and how much hash- entries are stored usefully or not, depending much on computing time, on depth and amount of hash used.
And it seems to me, positions stored with move history (full .pgn with moves leading to the given position from opening onwards) still work better.

[amchess]

Version 4.1 released
Corrected bug on learning table's update
https://github.com/amchess/BrainLearn/releases/tag/4.1