Life In 19x19

Another game : H4, komi 7.5, Exynos 990

W: BadukAI 0.3 (KG 20b, 5sec/move)
v
B: CrazyStone Deep Learning 5d

W+1
PS at move 302, B passed, considering the game was over with W+1. W wanted to continue, I stopped the game, to avoid 50+ useless moves.

Lazy Baduk is installed but your program is not - incompatibility. tablet on MediaTek MT8321, 4x1.3 GHz ARM Cortex-A7. Android 7.0. will not work?
limitation on android version? is the limitation in the program?

@Vargo

thanks for conducting and sharing the games.

As far as KataGo's behaviour at the end of the second game is concerned, the reason is as follows:

I configured KataGo to use Tromp-Taylor rules (for the sake of consistency within the app since LeelaZero always uses Tromp-Taylor rules, it doesn't know any other ruleset). With these rules your final score is "stones on the board" + "territory". They have the advantage that they can be formulated very concisely (ideal for AI). But a "disadvantage" is that you can continue playing until the board is full without hurting your score: Playing inside your safe territory doesn't cost (the stone will live, so +1 for the additional stone and -1 for the lost territory) and playing inside your opponent's safe territory doesn't cost either (the stone will die, so no plus here, but also no minus for lost territory). So when there is nothing productive left to do, all moves and the pass are of equal value. I assume that in this case KataGo just plays one of these candidates randomly which is mostly an additional move.

@And

The reason for the incompatibility is that the tablet you mentioned has 32 bit operating system whereas BadukAI-APK currently only supplies 64 bit binaries.

I will try to provide a 32 bit build too. But for this to succeed all the 3rd party software that I use must be available/buildable for 32 bit, so I can't promise anything at the moment. But I will try ...

it would be great! users have many more 32-bit devices, many will say thank you!

Vargo wrote:W wanted to continue, I stopped the game, to avoid 50+ useless moves.

akigo wrote: I configured KataGo to use Tromp-Taylor rules (for the sake of consistency within the app since LeelaZero always uses Tromp-Taylor rules, it doesn't know any other ruleset). With these rules your final score is "stones on the board" + "territory". They have the advantage that they can be formulated very concisely (ideal for AI). But a "disadvantage" is that you can continue playing until the board is full without hurting your score: Playing inside your safe territory doesn't cost (the stone will live, so +1 for the additional stone and -1 for the lost territory) and playing inside your opponent's safe territory doesn't cost either (the stone will die, so no plus here, but also no minus for lost territory). So when there is nothing productive left to do, all moves and the pass are of equal value. I assume that in this case KataGo just plays one of these candidates randomly which is mostly an additional move.

@Vargo @akigo - Just so you know, it should be fairly rare that KataGo prolongs the game by a lot. Usually play some solid connecting moves to make some of its groups absolutely connected, and then proceed to focusedly capture all the dead groups. After that, it will pass, and will usually *not* attempt to play any further moves that would waste time. For example, you can take a look at games like this for typical behavior: https://online-go.com/game/27067104
Or this one, which has more dead stones: https://online-go.com/game/27066163 - after the last dame is finished, kata spends almost all moves to cleanly capture the dead stones (required by tromp-taylor) and interposes one or two solid connection moves besides that, but no more, and then readily passes.

The reason KataGo behaves like this is because I added a *very* slight bias to prefer "orderly cleanup and then pass without wasting time" moves as a way to discriminate between moves at the end of the game that otherwise have no intrinsic preference or difference. It's only a very slight bias, but enough to produce this kind of behavior when all the moves still lead to the same outcome, rather than cleanup and passing being at the whim of the neural net's random noise preferences among all the possible moves. Unless you've disabled or changed it, the default config also has KataGo play faster if the opponent is passing a lot, so this cleanup also often doesn't take long.

50+ moves would be very atypical, although maybe it could happen in some case. But very, very atypical.

@akigo @lightvector
Very clear explanations, thanks.


Two more games (BadukAI03_KG20b is White, 5sec/move, komi 7.5, Exynos990):

H9 : Gnugo 3.8 is Black. Result W+R H6 : CrazyStoneDL 5d is Black. Result B+1.5

KG init was almost instantly here. Huawei Mate 20 X

KataGo 10 blocks are stronger than Leelaz 15 blocks: KataGo g170e 10 block d204M - Leelaz 15b 0c4ade79 2 sec/move (GT 610) 152:48 (76-24 KataGo Black, also White)

Wow that means that katago 10b is stronger than some of the 40b Leela nets from last year when Leelazero .17 was first released (time parity)? That's crazy. If I remember right, bubblesid's 15b was ahead of the 40b nets for quite some time on consumer machines. That's ridiculous how strong katago is. I think the 10b with 1 playout was stronger than zen6 on breakawa's 5d setting.

GT 610 is a low-power graphics card that is close enough in performance to tablets. but it seems to me that KataGo 10b and on other settings (time parity) is stronger than Leelaz 15b

I added a link to a separate 32 bit APK to https://aki65.github.io/. No guarantees since I don't have a 32 bit device to test. But anybody, who had the "incompatible with your device"-problem up to now, might have another try with the 32 bit APK ...

on the tablet while it works! thank!

can you make it possible to replace networks? can put networks in an accessible folder where you can remove or add a network? for example, I don't need Leelaz networks, but I need KataGo networks 6 and 15 blocks

Why is this app using internet connection? I was not able to find any feature that would need it.