Life In 19x19

Sander:
"Calibrated rank is broadly unreasonable below 10k or so. But there are really no good bots below that level anyway."
https://discord.com/channels/4170221623 ... 5688365067

several games 15 kyu Zen - CS Zero
(Zen - gtp4zen, Zen7, -t4 -r3 -s100 -n0 -o5.0 -p0.2
Zen's settings were picked up in the match against CS Zero):

rate these games. do they look like human games? interesting?

So ugly!

jlt wrote:So ugly!

Whose game is it? Zen or CS Zero?

I just looked quickly at the first game below your question "do they look like human games? interesting?"
I won't bother to comment every move.

is unnatural for a human, a human of that level wouldn't tenuki.
doesn't make sense, it's like a pass, only a human 30k would play like that.
is ridiculous
White chose earlier not to capture the cutting stone, why not if the intention was to attack the group J15, but now why give up that group?
Black 101 horrible
White 108 ugly shape. A human would play R11.
White 110 what's that shape??
White 130: what was White thinking???

thanks for the comment! in your opinion, there are no better bots than these?

I haven't played much with weak bots recently, but so far I have always been disappointed by bots below 10k. Games against humans are so much more interesting.

of course, playing with a person is completely different! but if you compare these games - you can find logic in them, with the games of kyu-rank bot-in them some moves are generally incomprehensible

Gnu Go, Many Faces of Go, Turbo Go,... they are all good around 10 kyu and have many different versions.

Rémi wrote:As longshanks wrote, Unbalance is in charge of selling the program, making new releases, customer support, etc. I'd rather not interfere with their communication.

As far as I know, no new version of the English Crazy Stone is planned for 2015. They made a new Japanese version. The main difference in the new Japanese version is that I managed to make weak levels that are much more human-like, in order to make the program more pleasant for beginners. Strength improved since 2013, but not enough to be worth releasing a new English version. They will release a new English version when there is a big strength improvement. I don't know yet when this will happen, but I am working on it.

https://lifein19x19.com/forum/viewtopic ... 03#p194303
It's probably about Crazy Stone Meijin

Rémi wrote:

Kirby wrote:How are weaker levels calibrated? Is it purely a function of time spent per move?

No. Weaker levels play instantly.

All the levels up to 1k are pure neural networks. They play almost instantly. I collected game records of weak players from KGS, and produced neural networks that imitate them. So the weak levels are considerably more human-like than before. Strength was calibrated by playing games on KGS.

As you can see on the attached sgf, it produces a very human-like way of being weak. I am really happy with the result.

It is interesting that the 13k neural network was produced by learning from game records of 15k players. I suppose it may have become stronger because it plays instantly, and because it can read ladders better.

Crazy Stone Deep Learning
https://lifein19x19.com/forum/viewtopic ... 89#p204189

kvasir wrote:Gnu Go, Many Faces of Go, Turbo Go,... they are all good around 10 kyu and have many different versions.

Gnu Go and Many Faces of Go play well at 15k level? is it better than katrain kyu-bot?

Amigo wrote:Gnu Go and Many Faces of Go play well at 15k level? is it better than katrain kyu-bot?

Gnu Go is good

Some version was too knowledgeable about monkey jumps compared to its actual level of play.

Better in what sense? Also this changed from 10 kyu to 15 kyu, a huge jump really for beginners. At least 10 kyu in EGF rating is not so shabby.

To answer the question. The calibrated rank bots as available in katrain are hard to play against and are only whatever rank because they force bad results. Something like Gnu Go plays at whatever level for a similar reason as human player, it has mistaken and simplistic concepts and is similarly stubborn as any human player. OK, maybe a bit of an exaggeration but this calibrated rank stuff often plays really well and then just drops a bad move, but the old stuff for example mostly doesn't start winning capturing races to then just drop the last move. That said it is sometimes possible to win with wild handicaps against Gnu Go and Manyfaces, I think because they often just do their thing which the opponent can easily predict, but they are pretty good at this level, and I wonder how the other stuff fares with wild handicaps.

thank you, it's interesting! and if you choose between Gnu Go and Crazy Stone Deep Learning for 11-15 kyu?

Rémi wrote:

Satorian wrote:It's been updated today. Couldn't find a change log though.

I'd be interested to know whether the strength ratings differ from the desktop version, especially if they refer to the strength of data set used for that particular playing strength, or whether they are ratings the engine actually achieved on KGS at that setting.

The change is a message that indicates how to open external sgf files.

The weak levels are identical to the desktop version, and are calibrated by playing on KGS. The strongest levels are time-controlled, and strength may vary depending on hardware.

https://lifein19x19.com/viewtopic.php?p=211321#p211321

you can download the free android version 15k - 2d.
https://play.google.com/store/apps/deta ... gocsdllite
on the computer it works through an emulator