Plenty of weak networks on CGOS 9x9

[Rémi Coulom]

As I told earlier, I connected plenty of weak networks to CGOS:

http://www.yss-aya.com/cgos/9x9/standings.html

These are the weak networks I use for the Crazy Stone app. They were trained to imitate KGS game records. The kyu level is the KGS rating of human players of the training set. I plan to let them run on the server.

If they are not weak enough, I will add more.

Folkert, we will watch you climb the ladder. I will be glad to advise and help.

Have fun.

[Hiroshi Yamashita]

Thanks!
Is it one playout/move?
I think human kyu players play blunder sometimes(eg. mistake of one liberty stone escape or capture),
but network may be not able to imitate them because they happen rare. Can your network re-produce them?

BTW, I added time on cross-table. It is not useful for your weak fast bots though.

[Rémi Coulom]

Yes, it is one playout / move, with randomness added in the opening for variety.

You are right. Supervised learning from weak human does not play like the games being imitated because of this "averaging" effect. I noticed it very clearly in chess: beginners will leave pieces hanging from time to time, but choosing the greedy policy move will not play such mistakes.

I also tried selecting moves at random according to the policy distribution instead of choosing the most probable move, but it produces an unpleasantly inconsistent style, mixing good moves and bad moves. Maybe something in between might be best.

[flok]

As I told earlier, I connected plenty of weak networks to CGOS:

http://www.yss-aya.com/cgos/9x9/standings.html

These are the weak networks I use for the Crazy Stone app. They were trained to imitate KGS game records. The kyu level is the KGS rating of human players of the training set. I plan to let them run on the server.

If they are not weak enough, I will add more.

Folkert, we will watch you climb the ladder. I will be glad to advise and help.

Have fun.

Cool thanks!

Have been working on an other project last week, time for Go again now.

[flok]

Something goes wrong:

Code: Select all

2023-02-27 23:08:15,119: Starting game against net_12k(1661). Local engine ("DellaBaduckce07ab0", rated 1726?) is playing black.
2023-02-27 23:08:15,119: This is a restart. Catching up 1 moves
2023-02-27 23:09:11,291: Game over. Result: B+Time
2023-02-27 23:09:11,291: Local engine won :-)

(please ignore the restart part)

net_12k did not do any moves at all (EDIT: and it looks like net_2k is not going to do any either)

[Hiroshi Yamashita]

http://www.yss-aya.com/cgos/9x9/cross/D ... 55ae0.html
Strange. Only net_*k lost by time.
From server log, net_8k who is playing with DellaBaduck0e55ae0 disconnected again and again.
It may be server bug because many matching is rare.

2023-02-28 02:24:27 | match-> CrazyStone-0006-99(2934) Gnugo-3.7.10-a1(1800)
2023-02-28 02:24:27 | match-> net_15k(1574) net_13k(1636)
2023-02-28 02:24:27 | match-> fuego1604_10k(2272) pureUctRave_1k(1274)
2023-02-28 02:24:27 | match-> net_2k(1838) net_3k(1825)
2023-02-28 02:24:27 | match-> net_5k(1726) DCNN_AyaF128a523x1(2041)
2023-02-28 02:24:27 | match-> net_18k(1416) net_4k(1772)
2023-02-28 02:24:27 | match-> net_12k(1661) net_10k(1707)
2023-02-28 02:24:27 | match-> net_6k(1766) net_7k(1784)
2023-02-28 02:24:27 | match-> DellaBaduck0e55ae0(1923?) net_8k(1762)
2023-02-28 02:24:27 | match-> net_9k(1728) net_14k(1563)
2023-02-28 02:24:27 | match-> net_11k(1659) Aya783a_50(1711)
2023-02-28 02:24:29 | [net_8k] disconnected
2023-02-28 02:24:32 | gameover: 1401544 net_9k net_14k B+4.0
2023-02-28 02:24:32 | gameover: 1401538 net_2k net_3k W+10.0
2023-02-28 02:24:33 | gameover: 1401541 net_12k net_10k W+2.0
2023-02-28 02:24:33 | gameover: 1401536 net_15k net_13k B+20.0
2023-02-28 02:24:33 | gameover: 1401542 net_6k net_7k B+24.0
2023-02-28 02:24:34 | gameover: 1401539 net_18k net_4k B+22.0
2023-02-28 02:24:42 | Games in progress: 5
2023-02-28 02:24:59 | client: e1 cgosGtp 0.98 alpha - engine client for CGOS Linux-x86_64 by Don Dailey
2023-02-28 02:24:59 | [net_8k] logged on
2023-02-28 02:24:59 | is net_8k currently playing a game?
2023-02-28 02:24:59 | testing 1401535 CrazyStone-0006-99 Gnugo-3.7.10-a1
2023-02-28 02:24:59 | testing 1401537 fuego1604_10k pureUctRave_1k
2023-02-28 02:24:59 | testing 1401540 net_5k DCNN_AyaF128a523x1
2023-02-28 02:24:59 | testing 1401543 DellaBaduck0e55ae0 net_8k
2023-02-28 02:24:59 | YES!
2023-02-28 02:24:59 | setup 1401543 9 7.0 300000 DellaBaduck0e55ae0(1923?) net_8k(1762) F4 300000 f3 298009
2023-02-28 02:24:59 | [net_8k] disconnected

[flok]

I also run test-matches at home with my homegrown matcher (stats generated with ordo):

Code: Select all

   # PLAYER                     :  RATING   POINTS  PLAYED   (%)
   1 GnuGO level 0              :  3879.9  31492.0   31684    99
   2 AmiGo                      :  3271.6  25484.5   31661    80
   3 Stop                       :  2293.4  17685.5   31705    56
   4 Daffy Baduck               :  1997.6  11963.0   31657    38
   5 dellabaduck_0e55ae0_1th    :  1950.8    282.5    1000    28
   6 dellabaduck_acbe918        :  1705.1   1462.5    9333    16
   7 dellabaduck_d3155f8        :  1660.8   2905.5   21202    14
   8 donaldbaduck_ffcb9be       :  1640.8   3685.5   31680    12

'Stop' is the stop-0.9-002 version, dellabaduck_0e55ae0_1th is the one-thread version of the one I ran last night on cgos and Daffy Baduck is a random mover. donaldbaduck should be the c++ version of stop, but for some reason plays very bad. Have to figure that out.

[Rémi Coulom]

Something goes wrong:

Code: Select all

2023-02-27 23:08:15,119: Starting game against net_12k(1661). Local engine ("DellaBaduckce07ab0", rated 1726?) is playing black.
2023-02-27 23:08:15,119: This is a restart. Catching up 1 moves
2023-02-27 23:09:11,291: Game over. Result: B+Time
2023-02-27 23:09:11,291: Local engine won :-)

(please ignore the restart part)

net_12k did not do any moves at all (EDIT: and it looks like net_2k is not going to do any either)

It may be because my gtp engine for the weak networks does not understand lower-case moves (a1 instead of A1). I am stopping them, and will connect them again later when it is fixed. I thought the server would always send the moves with the same case, but if it is just copying the string of players, then my program will not work against opponent who use lower case.

[Rémi Coulom]

I also run test-matches at home with my homegrown matcher (stats generated with ordo):

Code: Select all

   # PLAYER                     :  RATING   POINTS  PLAYED   (%)
   1 GnuGO level 0              :  3879.9  31492.0   31684    99
   2 AmiGo                      :  3271.6  25484.5   31661    80
   3 Stop                       :  2293.4  17685.5   31705    56
   4 Daffy Baduck               :  1997.6  11963.0   31657    38
   5 dellabaduck_0e55ae0_1th    :  1950.8    282.5    1000    28
   6 dellabaduck_acbe918        :  1705.1   1462.5    9333    16
   7 dellabaduck_d3155f8        :  1660.8   2905.5   21202    14
   8 donaldbaduck_ffcb9be       :  1640.8   3685.5   31680    12

'Stop' is the stop-0.9-002 version, dellabaduck_0e55ae0_1th is the one-thread version of the one I ran last night on cgos and Daffy Baduck is a random mover. donaldbaduck should be the c++ version of stop, but for some reason plays very bad. Have to figure that out.

How does your program work? Weaker than random should not be very difficult to improve. Simple Monte Carlo evaluation should help you.

[flok]

How does your program work? Weaker than random should not be very difficult to improve. Simple Monte Carlo evaluation should help you.

Oh I'm experimenting with all kinds of algorithms. DellaBaduck is currently a simple a/b search (reaching depth 4 on a threadripper...) with some things bolted on like:

* extend chains (at their 'ends')
* random field
* kill chains (by filling their liberties)