Skip to content

Home

Jump to bottom
Videodr0me edited this page Jun 17, 2018 · 55 revisions

Certainty Propagation and Autoextending

more here: https://github.com/Videodr0me/leela-chess-experimental/wiki/MCTS-Solver---Certainty-Propagation-and-Autoextending

P1: +181 -152 =667 Win: 51.45% Elo: 10.08 LOS: 94.40% P1-W: +102 -69 =328 P1-B: +79 -83 =339

Backpropagation: Q-Moving-Average

https://github.com/Videodr0me/leela-chess-experimental/wiki/Backpropagation:-Q-Moving-Average Works well at low visit searchs (< 800 visits per move) but fails at high visit searches - needs more investigation.

Selection

Do not trust q from initial visits:

More here: https://github.com/Videodr0me/leela-chess-experimental/wiki/Selection:-Don't-trust-initial-visits Did not yield any elo gains.

Optimal-Selection (--optimal-select=1)

This is not yet final - and just a playground for various selection strategies, currently it does well at tactics (170/200 WAC Silvertestsuite) but suffers somewhat in selfplay, even though results against different opponents (non leela) are better. Needs more work. Might be useful for long analysis as it restores MCTS convergence properties (under some circumstances leela would never find moves no matter how many nodes visited.) Not recommended at the moment.

Note: Test parameters were old default cpuct=1.2, fpu-reduction=0.0, and NN 5d46d9c438a6901e7cd8ebfde15ec00117119cabfcd528d4ce5f568243ded5ee

For test positions threads=1 and batchsize=1 were used for reproducability.

Clone this wiki locally