New android app "BadukAI"

[akigo]

@And

the kyu-rank bot is not KataGo. It only uses the policy output from KataGo's neural network to value moves. This network is independent of the ruleset. Therefore it can (and does) only exclude moves that are illegal with all rulesets (that is single-stone suicide). But multi-stone suicide moves have a normal (very low) policy value and are therefore sometimes chosen by the kyu-rank bot (but only if there is nothing better to do within the moves he considers, so this should not be a problem).

[And]

KaTrain (kyu-rank) does not commit and does not allow suicidal moves (I tried all the rules that are in the KaTrain), is it possible to do in BadukAI as well?

[And]

As far as I understand, it is now impossible to connect to the KaTrain through gtp. Would it be easier to assess the accuracy of a kyu-rank by playing a match of tens? hundreds? games. maybe this option will appear?

[Amigo]

it may be rational to remove lz 40b and kg 20b in the installation file, and add links to them "download lz 40b" and "download kg 20b" in manage networks. it seems to me not only I delete unnecessary networks after installation. the tablet often runs out of space.
and why unnecessary confusion? it is clear where such values come from in kyu-rank, but maybe it is better to dial "5d" and not "-4"?

[akigo]

it may be rational to remove lz 40b and kg 20b in the installation file, and add links to them "download lz 40b" and "download kg 20b" in manage networks. it seems to me not only I delete unnecessary networks after installation. the tablet often runs out of space.
and why unnecessary confusion? it is clear where such values come from in kyu-rank, but maybe it is better to dial "5d" and not "-4"?

For people with recent devices a few hundred MBs are not an issue, so I am not sure which option would be more convenient for most people. But there are two certain disadvantages of the download option:
1) The buttons for switching the network would not work right after installation
2) It would require internet access permission for the app which most people don't like to grant to an app for security and privacy reasons (especially if it is not really necessary)

The value in the kyu-rank option is more like a strength parameter, it doesn't really denote the rank, as I explained a few posts before. Values like "3d" for this option would further increase this false association, so I don't want to do that.

[akigo]

As far as I understand, it is now impossible to connect to the KaTrain through gtp. Would it be easier to assess the accuracy of a kyu-rank by playing a match of tens? hundreds? games. maybe this option will appear?

You should ask about this in the katrain forum. They must have scripts ready for this, since they run the kyu-rank bots on OGS.

[lightvector]

@And

the kyu-rank bot is not KataGo. It only uses the policy output from KataGo's neural network to value moves. This network is independent of the ruleset. Therefore it can (and does) only exclude moves that are illegal with all rulesets (that is single-stone suicide). But multi-stone suicide moves have a normal (very low) policy value and are therefore sometimes chosen by the kyu-rank bot (but only if there is nothing better to do within the moves he considers, so this should not be a problem).

By the way, if you are playing under known rules where a move is illegal, you can manually remove those moves that are illegal and renormalize the remainder of the probabilities to sum to 1.

This is what KataGo itself does when playing under any ruleset to handle the network putting policy mass on moves that would be illegal in that ruleset (such as superko retakes, or suicide moves). So that even if the neural net suggests the illegal move with a slight chance, the move will absolutely never get searched/played.

[And]

KaTrain 4d - CS Zero 4d 2:0 (B+, W+)
KG BadukAI 4d(-3) - CS Zero 4d 1:1 (W+, W+)

KG BadukAI 5d(-4) - CS Zero 5d 1:1 (B+, B+)

15b network, here too viewtopic.php?p=261143#p261143

KaTrain 4d - CS Zero 4d:

[go4thewin]

@And thanks for games against CS Zero, really very interesting and strong! I think regarding the suicide moves it is better not to fix it. Could the reason be that Chinese rules are not being sent to the net for some reason? So the net is not assigning a policy value of -1 to the suicide moves as it would on a desktop gui. In leela zero i think this would make the bot a liitle weaker because i think lz randomizes moves by likelihood of play, so if move x has 70% of the playouts, it has a 70% chance of being the move played if in "random“ mode. If suicide moves are included, that 70% chance of the top move played would go down just a little bit. i dont know how the katrain/kyu rank bots randomly pick which of the top moves to play. i wonder if allowing the user to select the ruleset from among the config file presets would fix it, though it would not if the kyu_rank bots are not using the policy output

Im really curious if kyu_rank ± 2 is always stronger than katrain‘s corresponding bot. Like -5 vs -3 , 10 vs 12 , etc. I wish I had time to test that out. If so I think it would be mission accomplished.

[And]

KG BadukAI 6d(-5) - CS Zero 6d 0:2 (B+, W+)
(15b network)

KG BadukAI 6d(-5) - CS Zero 6d 2:0 (B+, W+)
KG BadukAI 7d(-6) - CS Zero 7d 0:2 (B+, W+)
KG BadukAI 8d(-7) - CS Zero 8d 0:2 (B+, W+)
(20b network)

[akigo]

@And

thank you very much for conducting all these games.

The results of the high dan bots are according to expectation: Due to its construction the kyu-rank bot can never get better than the policy strength of the network used, and this is around 5d for 15b and 6d for 20b. To get a better opponent you have to run full KataGo with low number of playouts.

The results of the low dan bots suggest that the kyu-rank bot fares better against computer opponents than against humans. After all the 2d-katrain bot (which equals kyu-rank bot with -1) is only 1k on OGS. Elsewhere an explanation for this effect was suggested which I find quite plausible: Humans are much better than computers in adapting to their opponent's style of play. So once they detect that the kyu-rank bot doesn't read at all, they can exploit this by making wild fighting games.

@go4thewin
I understand your point regarding suicide moves, but I have no choice, I must fix that because:
When the kyu-rank bot makes a move, the move is sent to KataGo in order to update KataGo's internal board state so that KataGo can be asked for policy values on subsequent moves. But if that move is multi-suicide, then KataGo won't accept it (as it is playing chinese rules). So KataGo's internal board state would diverge from the game played and everything would go astray. I have to prevent that.

[go4thewin]

@akigo
Thanks! i just started using the 40b ddk kyu bots and really enjoy them. i think maybe they never make suicide moves? That might be a temporary fix. The only thing some may not like is they are very strong in the endgame, but I really enjoy it. For beginner level ddk (like 22-25) they get you out of your comfort zone because they dont usually play any joseki. Its like playing a go player who's playing stones wherever they feel like.
@And
I wonder if the 40b -6 kyu bot can compete with crazystone zero 7 dan? i think i tried the 40b -8 in an older version against crazystone pro 5d and lost.

[And]

KG BadukAI 9d(-8) - CS Zero 9d 0:2 (B+, W+)
KG BadukAI 8d(-7) - CS Zero 7d 0:2 (B+, W+)
(20b network)

KG BadukAI 20b 9d(-8) - CS Zero 7d
after move 225, CS Zero score for black was close to zero until ... move 325!
but there is more to this game too! in response to White's last move - "pass" instead of "passing" and winning, Black makes a suicidal move F2 and loses !!!
so BadukAI need suicidal moves? especially like that?

[go4thewin]

i dont get it though. at 325, it looks like black is trying to make a second eye in a last ditch effort to save the group, which is reasonable. white prevents that. then the f2 move is because there are no good moves left and the game is over. i dont get how it affects the score.

[And]

i dont get it though. at 325, it looks like black is trying to make a second eye in a last ditch effort to save the group, which is reasonable. white prevents that. then the f2 move is because there are no good moves left and the game is over. i dont get how it affects the score.

348 W H2