By Lothar Jung
Date 2020-10-27 08:54
Edited 2020-10-27 08:59
In dem neuen Discord Thread „tune-results“ fand ich folgenden Eintrag:
**Tune of CPuct/FpuValue/PolicyTemperature**
**LC0 version:** v0.26.3
**LC0 options:** Network: J92-190, Threads=1, MinibatchSize=30, MaxPrefetch=15, MaxCollisionEvents=15, MaxOutOfOrderEvalsFactor=2.0, Backend=cuda-fp16, MoveOverheadMs=0, ScoreType=centipawn_2019, MLH: TCEC-SuFi S18
**SF options:** SF-dev, Threads=1, Hash=2, Contempt=0, "Move Overhead"=0, SyzygyProbeDepth=10
**Tuning ranges:** CPuct: [0;5], FpuValue [-0.5;1.5], PolicyTemperature [0.5;3.0]
**Tuning configuration:** 4 separate tunings with different acq functions: mes/pvrs/ts/ei, rounds=5 and 1 tuning with CLOP: 10 games/iteration (equivalent to 5 rounds/iteration), 10000 games/tuning for each tuning session
**Hardware:** Ryzen 9 3950X (3.5GHz) + RTX 2060
**Time control:** 1.2s/game+0.02s/move (LC0/SF)
**Speed:** ≈200 nodes/move
**Book:** balanced 3-move book
**Tablebases:** 6-man
**Adjudication:** 6-man TBs + -resign movecount=3 score=500, -draw movenumber=20 movecount=5 score=10
**Software:** chess-tuning-tools / CLOP
**Optimum found:**
``` CTT.EI default CTT.TS CTT.MES CTT.PVRS CLOP
CPuct 2.084 2.147 1.524 1.543 1.700 2.422
FpuValue 0.296 0.443 0.261 0.360 0.190 0.612
PolicyTemperature 1.602 1.607 1.866 1.729 1.303 1.711```
**Comment:** Comparison between different optimiziation strategies in chess-tuning-tools and CLOP. Some of the setting configurations might be more suitable for specific tasks like pure exploration and are probably not designed to be used alone by itself. Nevertheless I wanted to see what works best in a 10000 games tuning.
**Test gauntlet:**
``` # ENGINE : RATING ERROR CFS(%) GAMES DRAWS(%) OppN
1 lc0.net.J92-190.CTT.EI : 0.4 7.0 54.3 10000 47.1 1
2 lc0.net.J92-190.default : 0.0 ---- 79.6 10000 46.7 1
3 lc0.net.J92-190.CTT.TS : -2.9 7.0 60.2 10000 47.0 1
4 lc0.net.J92-190.CTT.MES : -3.9 7.0 66.0 10000 46.0 1
5 lc0.net.J92-190.CTT.PVRS : -5.4 7.2 55.8 10000 47.9 1
6 lc0.net.J92-190.CLOP : -5.9 7.0 77.4 10000 46.5 1
7 stockfish-dev : -7.8 5.0 --- 60000 46.9 6
White advantage = 3.6 +/- 1.1
Draw rate (equal opponents) = 46.9 % +/- 0.2```
Lothar