Not logged inCSS-Forum
Forum CSS-Online Help Search Login
CSS-Shop Impressum Datenschutz
Up Topic Hauptforen / CSS-Forum / Weshalb Lc0 40b in Tests mittelmäßig abschneidet
- - By Lothar Jung Date 2022-01-30 10:49 Upvotes 2
… liegt an folgendem:

1. Es wird Lc0 0.28.2 anstatt 0.29 dev. eingesetzt.

2. Es wird nicht das stärkste Netz 780800 eingesetzt.

3. Es wird default und nicht die optimalen Parameter eingesetzt.

Diese sind:

CPunct 1.75

CPuctBase 45669

CPuctFactor 3.973

FPuValue 0.4645

Minibatchsize 52

PolicyTemperature 1.4582

Lothar
Parent - - By Walter Eigenmann Date 2022-01-30 11:58 Upvotes 1
Lothar Jung schrieb:
1. Es wird Lc0 0.28.2 anstatt 0.29 dev. eingesetzt.
Aber "0.29 dev." ist gar nicht downloadbar ?!
https://lczero.org/play/download/

.
Parent - - By Lothar Jung Date 2022-01-30 12:26 Edited 2022-01-30 12:31 Upvotes 3
Hallo Walter,

ich hatte den Link gepostet:

https://ci.appveyor.com/project/LeelaChessZero/lc0/builds/42166691

Erst das Backend und dann Artifacts auswählen.

Lothar
Parent - - By Walter Eigenmann Date 2022-01-30 13:30
Lothar Jung schrieb:
Hallo Walter, ich hatte den Link gepostet:
Dank und Gruss

.
Parent - By Lothar Jung Date 2022-01-30 16:59 Upvotes 1
Hier die endgültigen Parameter heute auf Discord:

**Comment:** Optimum was unstable and frequently changing even at the end of the tuning. Maybe kiudee can merge the games with the other T78 tuning running in parallel and give us more useful information about the optimum. So for now I don't recommend to run tests with the values below.
Statistics of games in tuning:```   # ENGINE     :  RATING  ERROR  CFS(%)   GAMES   SCORE(%)  DRAWS(%)
   1 engine2    :     9.6    3.3   100.0   12110      51.3      70.3
   2 engine1    :     0.0   ----     ---   12110      48.7      70.3

White advantage = 85.3 +/- 1.7
Draw rate (equal opponents) = 80.0 % +/- 0.5```I hope the statistics are not different with the other tuning for meaningfull merging.
**Optimum found:**
```
                  tuned                default
CPuct                1.722          1.745
CPuctFactor       3.973 (fixed)  3.894
CPuctBase         45669 (fixed)  38739
FpuValue           0.298            0.330 1.237          1.359``````
Parameter          Lower bound  Upper bound
-------------------------------------------
CPuct                      1.1          2.2
FpuValue                  0.09         0.87
PolicyTemperature         1.01         1.67```

Lothar
Up Topic Hauptforen / CSS-Forum / Weshalb Lc0 40b in Tests mittelmäßig abschneidet

Powered by mwForum 2.29.3 © 1999-2014 Markus Wichitill