Twiddle several Moggy playout policy parameters based on CLOP
These are CLOP experiments with pachi-29a24, single-threaded (i7-3770),
15x15 against GNUGo level 10; patterns were not used (?).
The number after [%] is |eigenvalue|.
abstime 500s 1000s 2000s 4000s
nakade 61% 3.39 80% 3.29 53% 7.62 66% 7.92
eyefill 54% 0.52 84% 0.61 49% 0.38 32% 2.45
nlib 29% 0.02 20% 0.09 22% 0.10 27% 0.28
ko 71% 0.08 63% 0.27 60% 0.33 48% 0.57
uElo SD 35 26 22 39 (smaller is more confident)
Just for your interest, mean winrate during tuning:
komi -35.5 -40.5 -50.5 -50.5 (to give GNUGo an advantage)
Elo 140 164 84 101
Nota bene:
- 4000s S.D. single-threaded translates just to 500s S.D. (8 minutes)
8-threaded, which is still faster than a regular KGS blitz.
- In general, 500s and 4000s values are unfortunately of low confidence,
this Elo spread is more than influence of most of the parameters.
Parameter analysis:
- nakade: let's go with the rough mean of 60%, which seems to hold.
- eyefill: we see a clear downward trend; let's go with 40%, more
tuning needed.
- nlib: we see about constant 25% optimum, but the feature is not
very interesting in general.
- ko: steady decrease, but 40% may hit closer to home in realistic
KGS setting.
Overally, after 500 games, the performance changes as follows (komi -55):
-Elo uElo +Elo (+-1SD)
361 377 392 prior=pattern=80,playout=moggy:nakaderate=20:nlibrate=20:korate=20:eyefillrate=60
396 411 427 prior=pattern=160,playout=moggy:nakaderate=20:nlibrate=20:korate=20:eyefillrate=60
384 399 414 prior=pattern=160,playout=moggy:nakaderate=20:nlibrate=25:korate=40:eyefillrate=60
404 420 435 prior=pattern=160,playout=moggy:nakaderate=60:nlibrate=25:korate=40:eyefillrate=40
Not too conclusive, unfortuinately, but it seems we should have a slight
improvement.