Life In 19x19

Hi uestccokey, I've been trying out Ah Q Go regular and Lite version now that I've got a new faster phone (Huawei Mate 10 pro). I like the analysis UI, and also the integration with live games from Yike (which saved me the trouble of installing it after failing to find it in the app store) which you can analyse inline and also see the Yike commentary variations too.

A question: which weight file does the Lite version come with?

Ah Q Go Lite 1.4.0 https://play.google.com/store/apps/deta ... id.aq.lite
1.Stability improvements
2.New cartoon and dark themes, currently supporting a total of five themes
3.A new melee mode is added to the play mode, i.e. AI will follow you to play move, which is suitable for practicing combat.

Uberdude, with the light version there are several weight files from 5 to 15 blocks. the strongest of them is modified (file size reduced) c11bc844 http://zero.sjeng.org/networks/c11bc844 ... 484b87c.gz
https://github.com/leela-zero/leela-zero/issues/2192

uestccokey, perhaps a redesigned lm network is right for your program 13x13?

network 13x13 obtained from lm gx89 (https://github.com/leela-zero/leela-zero/issues/2240 by ChangeBoardSizeOfWeight.cpp) against gtp4zen: 89-11.
cpuonly, threads 4, 10 sec/move, komi 7.5, gogui-twogtp 1.5.1
gtp4zen zen7 9dan -r20 -s6000 -n3 -o1.0 -p0.75 (https://github.com/breakwa11/GoAIRating ... /AIcmds.md)

And wrote:uestccokey, perhaps a redesigned lm network is right for your program 13x13?

network 13x13 obtained from lm gx89 (https://github.com/leela-zero/leela-zero/issues/2240 by ChangeBoardSizeOfWeight.cpp) against gtp4zen: 89-11.
cpuonly, threads 4, 10 sec/move, komi 7.5, gogui-twogtp 1.5.1
gtp4zen zen7 9dan -r20 -s6000 -n3 -o1.0 -p0.75 (https://github.com/breakwa11/GoAIRating ... /AIcmds.md)

ok, when I'm done with my work, I'll have a look again.

The 13x13 network received from the network (15 blocks) Ah Q Go Lite is probably better, the gz size is 16.6 MB.
against gtp4zen: 86-14 (on same conditions)

the same network (13x13) against leela 11, cpuonly, 10 sec/move: 98-2

and the same network (13x13) 5 sec/move, 2 threads, against the Crazy Stone Deep Learning Pro (android), 5 dan (~10 sec/move) through the nox emulator, 2 threads: 6-2

KataGo would probably be the best one to try on 13x13, AFAIK it's the strongest net on that board size

uestccokey, could I make a feature request: display the change in winrate from the previous move to the current move somewhere, in the header between the save icon and analyse/play switch perhaps. This would only be available if analysis was done for this move and the previous move. At the moment I work it out by on the previous move looking at yellow move winrate of bot's best suggestion eg 34%, go forward to game move and now yellow move is 70% so this means that move was bad as changed winrate by (100-34) - 70 = -6%. Having the app display this calculation and update whilst analysing would be very useful. Thanks, keep up the good work on my new favourite go app!

iopq wrote:KataGo would probably be the best one to try on 13x13, AFAIK it's the strongest net on that board size

yes, but so far there is no version KataGo cpu-only

2 games 13x13 network received from the network (15 blocks) Ah Q Go Lite 5 sec/move - Crazy Stone Deep Learning Pro 5 dan:

Ah Q Go Lite 1.6.1 with a new, stronger network!

https://play.google.com/store/apps/deta ... id.aq.lite
What's New:
1.Stability improvements
2.It's stronger now!