Not logged inCSS-Forum
Forum CSS-Online Help Search Login
CSS-Shop Impressum Datenschutz
Up Topic Hauptforen / CSS-Forum / Eröffnungsbuch von Leela-Entwickler
- - By Lothar Jung Date 2020-06-20 07:38 Edited 2020-06-20 07:48
Auf Discord hat Chad folgendes gepostet:

**FYI all testers who use my books**
I have remade my opening books with the latest games added. Additionally, I have taken into account some constructive criticisms about the original books regarding "duplicate" positions where the end position has the same board but not the same board state and history. I have collapsed these "duplicates" into the most frequent opening line by summing up the occurrences. Now there are no "duplicates" but they still contribute to the total ordering of the opening lines. I have not tested these books thoroughly yet but they look reasonable. The following table shows their coverage of the total games played by humans OTB and in correspondence chess:
```
Ply  Openings  Coverage
  2        10    78.70%
  4       100    85.33%
  6       500    84.00%
  8      2000    84.43%
10      5000    84.21%
12     10000    86.76%
14     10000    83.36%
16     15000    86.89%
18     15000    85.58%
20     15000    85.09%
22     15000    85.57%
24     15000    87.38%
```

The purpose of my books is to represent closely human play such that we do not add unintentional bias into our engine testing. My source is 9.7M human games from OTB and CC.

https://discordapp.com/channels/425419482568196106/539960268982059008/723509619762397204
Parent - - By Clemens Keck Date 2020-06-20 11:53
um mit diesen PGN ein Buch zu generieren braucht es viel Arbeit. Alle Züge habe die gleiche Priorität 0 .

C.K.
Parent - By Lothar Jung Date 2020-06-20 13:18
Dieser Test gibt für die ersten Züge eine Orientierung:

** sf-lc0 decisive games from megamatch opening book **
```500 games manually chosen out of 517 decisive games
* megamatch score (500.5-499.5) sf-lc0 in these games
* 47 Sicilian (56.5-37.5)
* 45 King's Gambit (51.5-38.5)
* 41 Queen's Gambit (36.5-45.5)
* 36 Giuoco Piano (34-38) includes Evans gambit, Max Lange attack, and Two Knights Defence
* 34 French Defence (33-35)
* 27 Ruy Lopez (24.5-29.5)
* 24 King's Indian (24.5-23.5)
* 20 Dutch (19-21)
* 19 Queen's Pawn (18.5-19.5)
* 19 English (17-21)
* 19 Alekhine's Defence (20.5-17.5)
* 16 Benoni (12-20)
* 12 Philidor (13-11)
* 11 Scotch (8.5-13.5)
* 10 Vienna (15-5)
* 9 Scandinavian (10.5-7.5)
* 9 Caro-Kann (9.5-8.5)
* 8 Nimzo-Indian (5-11)
* 7 Gruenfeld (7.5-6.5)
* 87 Other (84-90)
* Openings spread out by above groupings
* Openings spread evenly among SF and lc0 wins
* should show no bias throughout the match```
currently testing at <https://www.twitch.tv/wakamex>
Parent - By Lothar Jung Date 2020-06-20 13:21
Das Leela-Killer-Book ist für die zu vermeidenden Varianten nützlich:

Excuse me if this is not the proper channel for this.  I have created a "LeelaKiller Book".  This contains only lines that SF won (i.e. minimatch wins) during the entire course of the "All ECO codes" CCC MegaMatch (1+1).  There are 379 lines.  You might consider it the "ultimate rigged book". I thought someone might find it of use for testing various nets:   https://www.dropbox.com/s/75jady0s8pmsu5b/CCC%20LeelaKiller%20Book.pgn?dl=0
Parent - By Lothar Jung Date 2020-06-20 13:25
Hier ein 50 Positionen Eröffnungsbuch für Lc0 in talkchess:

http://talkchess.com/forum3/viewtopic.php?t=70030
Parent - By Lothar Jung Date 2020-06-20 13:32
Da die Brainfishe das Cerebellum-Leela-Net verwenden, könnte man dessen Varianten mit Nibbler testen.

https://zipproth.de/Brainfish/download/
Up Topic Hauptforen / CSS-Forum / Eröffnungsbuch von Leela-Entwickler

Powered by mwForum 2.29.3 © 1999-2014 Markus Wichitill