Heute auf Discord die beiden Test mit jeweils starker Hardware:

1. Test:

**Match**: J92/J94 vs SF-12 16 Threads
**Hardware**: RTX 3080, i9-10900kf
**Time control**: 5s + 0.5s (time forfeit disabled)
**Openings**: TCEC SuFi 10-19 (500 pos.)
**Lc0 26.3 cuda-fp16**: threads=2, minibatch-size=204, move-overhead=0
**MLH tcec-19**: moves-left-max-effect=0.2, moves-left-threshold=0.00000, moves-left-slope=0.004, moves-left-scaled-factor=1, moves-left-quadratic-factor=0, moves-left-constant-factor=0
**Stockfish 12 bmi2**: threads=16, hash=256, move overhead=0
**TBs and adj.**: syzygy 5-men, draw 5 moves 5cp move 30, resign (two-sided) 5 moves 600cp
**Comments**: Excellent performance by J94-40. J92-190 and SF-12 16T on par in this setup.
**Games**: https://gofile.io/d/gV3Vrk
```# PLAYER            :  RATING  ERROR  PLAYED  (%)   CFS   W    D    L    D(%)
1 lc0-263-j94-40        :    23.6   11.8   1000  53.20   90  210  644  146  64.40
2 lc0-263-j92-330  b   :    11.8   12.5   1000  51.60   97  202  628  170  62.80
3 stockfish-12-16T.      :     0.0   ----   3000  48.40   50  489 1926  585  64.20
4 lc0-263-j92-190       :     0.0   11.7   1000  50.00  ---  173  654  173  65.40

Engine             Depth   MIDG   EARLY   ENDG   LATE
lc0-263-j92-190    9.70   10.59 | 11.15 | 8.95 | 7.25
lc0-263-j94-40     9.82   10.68 | 11.08 | 9.10 | 7.28
lc0-263-j92-330    9.90   10.81 | 11.25 | 9.00 | 7.66
stockfish-12-16T   26.82  22.65 | 23.06 | 28.22 | 43.27```

2. Test:

**SPRT-Matches** of 384 nets
**LC0 version:** v0.26.3
**LC0 options:** Threads=1, MinibatchSize=64, MaxPrefetch=32, MaxCollisionEvents=32, MaxOutOfOrderEvalsFactor=2.0, Backend=cuda-fp16, ScoreType=centipawn_2019, MoveOverheadMs=0,  SmartPruningMinimumBatches=2, RootHasOwnCpuctParams=false, CPuct=1.350, FpuValue=0.228, PolicyTemperature=1.497, CPuctFactor=3.468, CPuctBase=20253, MLH-S18-SuFi-Settings: MovesLeftMaxEffect=0.2, MovesLeftThreshold=0.0, MovesLeftSlope=0.007, MovesLeftScaledFactor=1.0, MovesLeftQuadraticFactor=0.0, MovesLeftConstantFactor=0.0
**Hardware:** Ryzen 9 3950X (3.5GHz) + RTX 3080
**Time control:** 1.5s/game+0.025s/move
**Speed:** ≈1000 nodes/move
**Book:** 3-moves unbalanced book, order=random
**Tablebases:** 6-man
**Adjudication:** 6-man TBs + -resign movecount=3 score=500, -draw movenumber=20 movecount=5 score=10
**Software:** Cutechess-CLI, time forfeits disabled
```diff
   # ENGINE             :  RATING  ERROR  CFS(%)   GAMES    W     D    L  DRAWS(%)
+  1 lc0.net.J94-40     :    15.6   10.2    99.9    2336  659  1104  573      47.3
   2 lc0.net.J92-320    :     0.0   ----     ---    2336  573  1104  659      47.3
White advantage = 155.5 +/- 5.1
Draw rate (equal opponents) = 59.5 % +/- 1.3```BayesElo bounds [0.0;5.0] scaled to logistic bounds based on drawrate:
`sprt.py --elo0 0.0 --elo1 3.883 --results 15 212 635 284 22`:
```
[Elo0,Elo1]                 :  {0.000,3.883}
Elo                            :  12.1
Confidence interval         :  [4.6,19.7] (95.0%)
LLR [u,l]                    :  2.944  [-2.944,2.944]```

Hier die Einzelheiten und Neuigkeiten zum J94-40 Netz auf Github:

https://github.com/jhorthos/lczero-training/wiki/Leela-Training

Hier die Beschreibung der neu eingeführten „Value Repair Method“:

https://github.com/jhorthos/lczero-training/wiki/Value-Repair-method

Lothar