Handicap go with LZ variable komi

[Vargo]

10 game match LZ-variable komi with network #9006c708 v GnuGo (5K) , Handicap 9, komi 0.5
twogtp 1.5.1, no error, no duplicate game.

#9006c708 wins 70 % !!!

Code: Select all

gogui-twogtp -white "C:\path to LZVK\leelaz.exe --gtp --weights=path to weights\15x192_9006c708c.gz --noponder -r -1 -v 1600 -m 20" -black "C:\Users\jm\gogui151\gnugo38\gnugo.exe --mode gtp" -games 10 -handicap 9 -sgffile lzVK9006c_gnuH9 -auto -komi 0.5

I don't know if #9006c708 (15x192, bjiyxo, 2019-05-12) is the best network for high handicap, but I find it very impressive to win 70% at 9 stones.
On my computer -v 1600 corresponds to less than 3 minutes for the whole game, or less than 2 seconds per move. Maybe, with more visits, LZ would win near 100% at H9...
-m 20 to add some variety and try to avoid duplicate games.
-r -1 to avoid too early resignation.
Gnugo and LZVK always agree on the winner, but not always on the score...




LZ_Dynamic Komi v0.42 : HERE

The games :

lzVK9006c_gnuH9.zip

(9.1 KiB) Downloaded 483 times

The stats :

h9.gif (42.19 KiB) Viewed 5030 times

Example settings with Sabaki :

15x192.jpg (39.79 KiB) Viewed 5030 times

[Vargo]

HELP....

I wanted to try the same H9 10 game match, with -v 6400 (instead of -v 1600)

Code: Select all

gogui-twogtp -white "C:\Users\jm\Desktop\gogui151\leelazKOMIVARIABLE042\leelaz.exe --gtp --gpu 0 --gpu 1 --weights=C:\Users\jm\Desktop\LZ_networks\15x192_9006c708c.gz --noponder -r -1 -v 6400 -m 20" -black "C:\Users\jm\Desktop\gogui151\gnugo38\gnugo.exe --mode gtp" -games 10 -handicap 9 -sgffile lzVK9006c_gnuH9 -auto -komi 0.5

In the gogui 1.5.1 stats, the scores are always different when they're seen from B's side or from W's side.

For the game "lzVK9006c_gnuH9-1.sgf" , there are 4 results (see stats)

B+0.5 in twogtp (seen from B's side)
W+9.5 in twogtp (W)
B+5.5 in Sabaki (area)
B+15.5 in Sabaki (territory)


Is it a bug ? Is it me doing something wrong ?
What's the correct score?

Any idea, please ?


stats and score in Sabaki :

x.jpg (309.46 KiB) Viewed 4944 times

Sabaki territoy :

xx.jpg (145.23 KiB) Viewed 4940 times

The game with the strange result :

lzVK9006c_gnuH9-1.sgf

(1.96 KiB) Downloaded 456 times

[And]

1. Sabaki is often mistaken with the definition of dead stones. in your example, the error in the upper right corner - is a white group of 6 stones is alive.
2. gnugo uses Japanese rules by default. add --chinese-rules.
3. area and territory values are not always equal, but sometimes coincide. the territory probably is Japanese rules, and the area is Chinese

therefore, the correct result: W + 9.5 by the Chinese and B + 0.5 by the Japanese rules

[Vargo]

1. Sabaki is often mistaken with the definition of dead stones. in your example, the error in the upper right corner - is a white group of 6 stones is alive.
2. gnugo uses Japanese rules by default. add --chinese-rules.
3. area and territory values are not always equal, but sometimes coincide. the territory probably is Japanese rules, and the area is Chinese

therefore, the correct result: W + 9.5 by the Chinese and B + 0.5 by the Japanese rules

Thx , @And

__________________________________________________________

Another 10 game match at H9 : LZVK_9006 v gnugo

For LZVK:

Code: Select all

--handicap --noponder -r -1 -v 6400 -m 20

and for gnugo:

Code: Select all

--chinese-rules --play-out-aftermath --capture-all-dead --komi 0.5 

And this time, both agree on the winner, and almost agree on the score (gnugo adds almost every time 9 points to W, probably the 9 handicap stones)

The score is once again 7-3 for LZVK
The stats :

an.gif (170.1 KiB) Viewed 4813 times

The games :

lzVK9006c708_gnuH9.zip

(10.24 KiB) Downloaded 468 times

[And]

yes, differences between the results of gnugo and leela, probably because leela ignores the handicap. but what is with lzVK9006c708_gnuH9-1? you know?
crazy stone dl shows 26.5, (I did not see that crazy stone incorrectly counting). where is the mistake? lz or gogui?