TalkChess.com

I debugged the certainty propagation (essentially alpha/beta backup of terminals) in MCTS. Now that it's working, it seems to make a huge difference. No oops! checkmates or won endgames slipping away into a draw. This is with Bad Gyal 8 and Little Ender at 3+2. Will add tree reuse tomorrow. Stay saf...

Using a endgame_net branch, I combined bad Gyal 8 and little Ender, switching at 16 pieces. (Need to update the badgyal package for this.) Getting closer. Next is basic certainty propagation. Will put the a0lite games up on lichess. TC 3 min + 2 sec # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W...

Rather than jumping through hoops with an old lczero engine, simply try the small nets: 11258-16x2 distilled and tiny gyal (also 16x2). They can be used with the latest lc0 on cpu (or android). https://github.com/dkappe/leela-chess-weights/wiki/Distilled-Networks https://github.com/dkappe/leela-ches...

Well, a description of the Bad Gyal nets can be found here: https://github.com/dkappe/leela-chess-weights/wiki/Bad-Gyal Bad Gyal data can be found here: https://github.com/dkappe/leela-chess-weights/wiki/Bad-Gyal-Data Training is a bit more challenging. Daniel Shawul’s Keras training code on github ...

I added another branch, 'batching-prune,' that has smart pruning. Seems to make a fair bit of difference. [Event "?"] [Site "?"] [Date "2020.04.01"] [Round "5"] [White "A0Lite-Pruning"] [Black "A0Lite-Batch"] [Result "1-0"] [ECO "B06"] [GameDuration "00:02:33"] [GameEndTime "2020-04-01T16:49:34.165 ...

I decided to test the virtual loss and tree reuse branch at 30 min + 5 sec against Crafty on 1 cpu. Using Noomen 3. # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L D(%) 1 Crafty : 0 ---- 26.5 40 66.2 100 14 25 1 62.5 2 A0Lite : -126 67 13.5 40 33.8 --- 1 25 14 62.5 Not quite there yet. :-)

Question: Once I make the changes, they take effect immediately and there is nothing to compile? All of this is done at runtime? So if I make a local branch with some changes and want to test it against master should I just copy to another folder and run from there? Yes, just make a new folder. Als...

Tried out a0lite with batching and tree reuse at a tc of 30 m + 5 sec. Donna 4.1 lost on time, but it was a foregone conclusion. [Event "?"] [Site "?"] [Date "2020.03.30"] [Round "2"] [White "Donna"] [Black "A0Lite"] [Result "0-1"] [ECO "B00"] [GameDuration "01:03:06"] [GameEndTime "2020-03-30T21:52...

Added in a branch that combines batching with tree reuse.

https://github.com/dkappe/a0lite/tree/b ... tree-reuse

I will add in smart pruning in a day or two. Then I will need to think about simplifying the code.

Ah, you'll need to reinstall badgyal, as it has some bulk eval support added.