Hier ist eine detaillierte Beschreibung des neuen Zeitmanagements, das als Grundlage die Verteilung/Abschätzung der Bedenkzeit nach nodes/sec. vornimmt:
http://lczero.org/dev/docs/timemgr/<@!450371839680708608> it tries to equalize total nodes per move, and not total time.
It's not the same as there is tree reuse. It also tries to account for earlier stop due to smart pruning.
Also the search is extended when max Q is not max N.
In diesem Test hat das neue Zeitmanagement schlechter abgeschnitten:
**Match:** Gauntlet 63700 vs SF11
**Hardware:** Geforce RTX 2070S & i7-5960X (8 cores/16 threads)
**Lc0 version (legacy):** v0.25.1
**Lc0 version (smooth):** 0.26.dev with PR1310
**Lc0 options:** Backend=cudnn-fp16 MovesLeftMaxEffect=0.2 MovesLeftThreshold=0.0 MovesLeftSlope=0.007 MovesLeftQuadraticFactor=0.85 MovesLeftScaledFactor=0.15 MovesLeftConstantFactor=0.0
**Additional Lc0 options (smooth):** --time-manager=smooth(init-piggybank=0.0,per-move-piggybank=0.18,max-piggybank-use=0.95,max-piggybank-moves=27.0)
**SF11 options:** Threads=14 Hash=1024
**Speed:** Lc0:11.4 kN/s, SF11:12554 kN/s (from benchmarks)
**Time control:** 15s+0.5s
**Book:** openings-8ply-10k.pgn, sequential, color reversed
**Tablebases:** 6-man syzygy
**Adjudication:** 6-man TB, -draw movenumber=50 movecount=5 score=8 -resign movecount=5 score=1000
**Software:** cutechess-cli
```diff
# PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L
1 lc0.net.63600-legacy : 54.4 13.5 460.5 800 57.6 100 190 541 69
2 lc0.net.63600-smooth : 29.1 13.6 432.5 800 54.1 100 172 521 107
3 stockfish_11 : 0.0 ---- 707.0 1600 44.2 --- 176 1062 362
White advantage = 47.88 +/- 5.04
Draw rate (equal opponents) = 69.93 % +/- 1.24
```
Alles schwierig und theoretisch, aber so ist das bei Leela.
Lothar