New android app "BadukAI"

[And]

does anyone know how to make sure that LZ does not give up, for example in Sabaki? and what is the range of the -r parameter?

[And]

the new board and stones are certainly more interesting, thanks!

[go4thewin]

BadukAI 15b 4 playout (1 thread) vs Golaxy 4d 1:0 baduk ai won easily as black with 7.5 komi

15b4poVSgolaxy4d.sgf

(674 Bytes) Downloaded 618 times

order to play katago bots ddk - 4d+ by playout and net size below

Katago Mobile Strength.xlsx

(5.88 KiB) Downloaded 658 times

BadukAI 15b 4 playout (1 thread) vs Golaxy 5d 1:0 but very close. (misclick so middle game appears as variation) http://eidogo.com/#2SJaN1Tt

BadukAi 20b 1 playout vs Golaxy 6d 1:0

20b1poVSgolaxy6d.sgf

(990 Bytes) Downloaded 627 times

20b katago is stronger than the 6d bot I think
BadukAi 20b 2 playout vs Golaxy 7d 1:0
http://eidogo.com/#2mYoaKgPz:0,0
I think with 2po katago is stronger than the 7d bot

BadukAi 40b 1 playout vs Golaxy 8d 1:0
http://eidogo.com/#I4UY0h6J:0,0
15b 1 playout vs cspro 5d 1:0 http://eidogo.com/#vVunLVxD

[akigo]

I made a new version (0.14) responding to feature requests from various sources:

• "edit config"-dialog now has a fixed height
• settings (like think time) are now persistent, that is BadukAI remembers them on next app start
• the current position is saved automatically and you can reload it after next app start with the new button "load last game" in the "load sgf"-dialog
• restriction of pondering to a rectangle is now also available for KataGo
• the "resign percentage" for LeelaZero is now editable in "edit config"-dialog (value between 0 and 99, LZ resigns if winrate percentage is lower than this value)

[Amigo]

what a great app! I am delighted!!!

[ispolin]

Terrific! Thank you, Andreas!

[And]

KG B, resign. after saving sgf: screenshot. I pressed "ponder" twice, it seemed to be normal, I watched the game, but after a couple of minutes it was the same again.
EDIT the winrate is okay, just such a game

[akigo]

@And

The space bar on a physical keyboard toggles pondering too. This is normally irrelevant on real devices, so I never mentioned it. But if you used an emulator, this might explain the surprising pondering.

I will remove that in the next release, since it bypasses the UI and leaves the "ponder"-button in an inconsistent state.

[And]

@akigo

thanks! I tried to save the sgf on the tablet, when I typed the name, when touch the "space" can see in the background how pondering turns on, when touch it again, it turns off

[go4thewin]

@akigo
thanks so much, it feels like sabaki is now on android. Feel free to ignore this suggestion to increase kyu rank strength, it would probably be a huge task or maybe not even possible. I was wondering, the onlinego app on github found a way to specify different visits in katago for "hint“ (30) “territory“ (10) and play (20). I was wondering if you want to use his code to make the higher kyu ranks stronger. The 15b net is easily stronger than the golaxy's 7d bot at 20 visits. Since most devices are octacore or less, maybe kyu rank -4 can be the policy network full strength (katrain has a policy bot) kyu rank -5 can be 8 visits, -6 12 visits, -7 16 visits, and -8 20 visits. Using the 20b net instead would give a true 9 dan bot if the user has the cpu for that.

also, if it is of interest to anyone, i just realized you can use both analyze all and ponder together. So it will give you the winrate over the whole game and you can go and ponder a specific move and see the optimal variation played out by clicking the coordinate icon at the bottom

[akigo]

@go4thewin

Currently the label "kyu-rank" is a little misleading: After all, what this setting does is providing a katrain bot with the strength parameter you enter. But as your experiments (and also the online rankings of the katrain bots) show, the real rank of the resulting bot is several levels off this strength parameter. So a more accurate label would be "katrain bot with strength parameter:", but this doesn't fit on the screen .

But as the setting isn't a real rank right now, I don't want to mix it up with other methods providing ranked bots.

I think that the current approach in BadukAI for generating opponents of appropriate strength works quite well:

• BadukAI provides the "raw" settings to tune bot strength:
  • choosing a network
    limiting playouts and/or visits via "edit config"
    reducing strength below policy strength via "kyu-rank"
  Skilled users provide recommendations which "raw" settings to choose to get a certain bot strength (like you did in earlier posts)

This is much more flexible than any hard-coded approach.

[And]

how to connect KataGo to Sabaki, with the KaTrain parameter? for example "kyu-rank" -8

[go4thewin]

@And good idea i dont think we can yet, maybe they will be inspired by akigo!
@akigo you are right. this current setup is the best, the program is already so strong. i did one game of the 15b 1 playout policy vs crazystone pro highest level “5d“ baduk ai won by 90!

15b1poVScs5d.sgf

baduk ai - black
cs pro - white
komi 7.5

(1.37 KiB) Downloaded 606 times

If anyone needs a bot weaker than the 6b net 1 playout policy, 20b kyu_rank 4 or 5 and above are weaker as are all of the golaxy kyu bots. 20b kyu rank -15 is good for someone learning the rules. On the high end, the 40b net at 3 visits is probably atleast kgs 7d. it beats golaxy's top 9d+++ free bot.

40b3visitVSgolaxypro3star.sgf

(647 Bytes) Downloaded 618 times

I tested the golaxy bot and won on kgs against the kgs 6d Hirabot by 50 playing as black. last updated mobile strength suggestions 30 kyu - 7d: https://docs.google.com/spreadsheets/d/ ... sp=sharing

[And]

KataGo makes a suicidal move again

[And]

KaTrain 3d - CS Zero 3d 2:0 (B+, W+)
KG BadukAI 3d(-2) - CS Zero 3d 2:0 (B+, W+)
KG BadukAI 3d on emulator ~2sec/move