minigo/notes.txt at minigui · artasparks/minigo

Repository navigation

1. 3/4 of the way through v125, change squash from 0.95 => 0.98, change temp cutoff

to move 30 (from 31, because odd number = bias against black)

1. v192 -- learning rate erroneously cut to 0.001. v207 returned to 0.01.

(v206 trained with 0.001, v207 @0.01) (Later rolled back completely)

1. v231 -- moved #readouts to 900 from 800 as game length was too short

1. v 295-6 -- move shuffle buffer to 1M from 200k

1. v347 -- changed filter amount to 0.02. (first present for v347)

1. v348 reverted #4

1. v352 change filter to 0.03 and move shuffle buffer back to 200k

1. during v354 (for 355) change steps per generation to 1M from 2M, shuffle

buffer down to 100k

1. v360ish -- entered experimental mode; freely adjusted learning rate up and

down, adjusted batch size, etc, to see if valnet divergence could be fixed.