The Deep Zen team said in KGS kibitz that Zen19K now had a Value Net -- AlphaGo's breakthrough innovation where a neural net evaluates the board position. Zen19K was already 8d on KGS back in July, presumably running with a Policy Net and Monte Carlo playouts. But now version 12.1 with the addition of a Value Net has become even stronger.

I'm really happy to see they are on track to reproduce AlphaGo's success. Back in March the Zen team announced they were teaming up with Dwango, a Japanese company, to get the hardware needed to train neural networks. (see viewtopic.php?f=18&t=12750 ). I'm hopeful they will be able to improve their neural nets quickly and we'll see the first KGS 9d bot soon!

Zen19K's info shows the hardware they're running on:

Honestly, I feel like I need a bot just to be able to get a game with one of these bots. There ought to be a queue system in place.

_________________
Ko is the best solution.
With Ko, I can keep eating and drinking until I am full.

Visit >>>Koosh's Study Journal<<<

Where can we buy this 12.1 version?

zen19k is very popular, a lot of people competing in clicking fastest to play with it.
Imagine the popularity when AlphaGoBot play, you already won something if you click fast enough to get to play it

No idea, I don't think it’s possible to purchase it … but meanwhile it’s version 12.2, allegedly a bugfix

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

Zen19k is dominating at 8d, but 2 games against Ilya Shikshin 1p EGF caught my eye. First of all Zen won with a pretty convincing powerful attacking style on Ilya's weak groups which it almost teasingly allowed to live because it still won:


But then Ilya had his revenge with a magical win: he died in an approach ko early on and Zen looked good, but it then seemed to get confused when Ilya activated the ko and proceeded to waste a load of threats and fall behind and go bonkers. Is this weakness with approach ko something humans will be able to exploit?

Sinan Djepov aka YoungPro 6d on KGS posted some impressions of the program on his blog:

https://explorebaduk.wordpress.com/2016 ... en-zen19k/
https://explorebaduk.wordpress.com/2016 ... 9k-part-2/
https://explorebaduk.wordpress.com/2016 ... 9k-part-3/

Is this really going to work on an average laptop, with just one average GPU? When they they sell commercial version in the future, maybe it is going to be sold separately - CPU only vs CPU&GPU version?

Zen made a new account -- Zen19K2 -- and it is now 9d on KGS! Congratulations to the DeepZen team, well done.

It's the exact same software and hardware as Zen19K, but it limits handicap to 3 or less (Zen19K limits to 4 or less).

ETA: Zen is now #4 on the Top 100 list. Rating graph shows it's over 10d.

Zen19K2 is dominating! Beat cheater (Pavol Lisy 1p EGF) a few times, HitGirl7 (Mark Lee super strong korean insei who won US open beating quite a few low dan Chinese pros there), breakfast (Alex Dinerstein 3p)..., Tungpo (a strong player who plays Zen a lot).







Not quite invincible though: it did lose against a mystery guest, probably a Chinese 1p? Interesting that it lost when the human made a huge moyo.

I wonder why this page sometimes shows the games and sometimes doesn’t: http://www.gokgs.com/gameArchives.jsp?user=Zen19K2

(link edited, thanks to ez4u!)

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

It has to do with the extra characters at the end of the link:"%EF%BB%BF". It should only have "user=Zen19K2".

_________________
Dave Sigaty
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21

I'm more interest in Zen19L (long time setting), there's also Zen19M registered but I'm not sure if M is for medium time setting.

For the curious, the %EF%BB%BF on the end of Bonobo's link was the URI encoding of the UTF-8 encoding of the unicode character U+FEFF, a non-printable character used as a byte order mark or as a marker in text streams. It probably got there from some weird copy-paste.

Thanks to ez4u for the correction,

and thanks, Uberdude, for the explanation:
Yes, I’m curious indeed, and yes, I copied this and pasted (of course I didn’t type it), and I either didn't see or didn’t notice those characters, sorry for the confusion

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

This is a good site to use to keep track of this bot's stats.
90% win rate (across the board) so far...

_________________
Ko is the best solution.
With Ko, I can keep eating and drinking until I am full.

Visit >>>Koosh's Study Journal<<<

Could someone place zen18k on goratings list somehow?

It doesn't have to be 100% precise.

I estimated it to be ranked aroud 400th in blitz time setting (that it currently play in kgs, 15sec/move) and 600th in normal time setting (30 minutes main time) and 800th in long time setting (2 hours main time or more).

When comparing to goratings.org -- a list of human professional go players -- I feel obligated to put disclaimers: Zen is playing blitz time controls which historically favors the bot. Also we don't know how seriously the humans are taking the games, in particular cloudnotes lost a game where Zen gave him 2 stones, but cloudnotes appeared to be experimenting, or even trying to half give back the handicap by playing 2-2 as the first two moves. Also another problem although goratings is better than Elo and other systems, it still suffers the difficulty of separate player populations. These western pros listed are probably overrated by goratings. There are some crazy rumors about how strong some of the guests playing Zen are. But this is just KGS kibitz so I don't give them much credit for now.

KGS name, real name, goratings rank, goratings rating

HitGirl7, Mark Lee, N/A but should be stronger than those listed below -- Zen won 1-0
cheater, Pavol Lisy, #779, 2800 -- Zen won 3-0
breakfast, Alexander Dinerchtein, #815, 2748 -- Zen won 1-0

For reference Fan Hui is currently #533, 2990.

Also cloudnotes is rumored to be Andy Liu. He is #263, 3152.


For these blitz games it seems clear that it's pro strength. No telling how much that changes if the time controls are longer like the Fan Hui games.

I guess in current blitz settings it is probably on low eastern professional level if not mediocre professional level.

But I guess with the improvement we can assume that in the near future it will reach average professional level even on LTC, which is actually great.