12.80 needed retuning after change to root prior mix.
Retuned pattern prior after switching to normalized probabilities
Good as these haven't been retuned since pre-mm days ! (pachi < 12.00)
Until pachi 12.80 pattern probs were scaled based on max (not total), which is nice because it boosts lower ranked moves but probabilities aren't really probabilities then (don't add up to 1). Switching to normalized probs has advantages: can measure probability bias in t-predict stats (maybe even correct it), and makes pattern prior logic basically the same as dcnn so can meaningfully compare the two.
Also simplified negative eqex prior values, so -400 is 400 on 19x19 (not -2000 -> 400 as before).
Dcnn threads:
Made OpenBLAS follows uct threads param, so setting OMP_NUM_THREADS env variable not necessary anymore.
Running with threads=1 will also evaluate dcnn on one thread now.
Results:
pattern prior: 160 -> 300
57% winrate against pachi-12.80 (nodcnn).
250 was the optimal value for 12.80, and an increase of 50 more or less compensates normalized probs.
So sweet spot is around 300 now.
dcnn prior: 1300 -> progressive 1300 down to 900 by move 150
Sweet spot for fixed dcnn prior is in [600 - 900] range now (55% winrate against 12.80).
However low prior isn't great early game: playouts are pretty clueless and mcts sometimes doesn't follow dcnn when it should, leading to early game blunders. On the other hand it's good in middle/end game to escape dcnn blind spots. So use progressive dcnn prior : 1300 down to 900 by move 150 (55% winrate against 12.80 also).
Progressive dcnn prior can be changed with dcnn_high, dcnn_low prior params.
Using dcnn prior param reverts to fixed dcnn prior.
Retuning release:
Good as these haven't been retuned since pre-mm days ! (pachi < 12.00)
Until pachi 12.80 pattern probs were scaled based on max (not total), which is nice because it boosts lower ranked moves but probabilities aren't really probabilities then (don't add up to 1). Switching to normalized probs has advantages: can measure probability bias in t-predict stats (maybe even correct it), and makes pattern prior logic basically the same as dcnn so can meaningfully compare the two.
Also simplified negative eqex prior values, so -400 is 400 on 19x19 (not -2000 -> 400 as before).
Dcnn threads:
Made OpenBLAS follows uct
threadsparam, so settingOMP_NUM_THREADSenv variable not necessary anymore.Running with
threads=1will also evaluate dcnn on one thread now.Results:
pattern prior: 160 -> 300 57% winrate against pachi-12.80 (nodcnn). 250 was the optimal value for 12.80, and an increase of 50 more or less compensates normalized probs. So sweet spot is around 300 now.
dcnn prior: 1300 -> progressive 1300 down to 900 by move 150 Sweet spot for fixed dcnn prior is in [600 - 900] range now (55% winrate against 12.80).
However low prior isn't great early game: playouts are pretty clueless and mcts sometimes doesn't follow dcnn when it should, leading to early game blunders. On the other hand it's good in middle/end game to escape dcnn blind spots. So use progressive dcnn prior : 1300 down to 900 by move 150 (55% winrate against 12.80 also).
Progressive dcnn prior can be changed with
dcnn_high,dcnn_lowprior params. Usingdcnnprior param reverts to fixed dcnn prior.