Zen 15.0 on CGOS very strong new version

[lightvector]

Some thoughts:

* If Zen's policy and value nets are strong enough to start to have some AlphaGo-like qualities in the opening and early-middle-game and to produce some innovations in joseki, then even if Zen overall is not clearly stronger, playing it a bunch could have plenty of value.

* Given that new-generation Go bots are strongest in whole-board judgment and weaker at complex fights or races, I could also see value in playing the first 100 moves with it a whole bunch to practice just the opening and early midgame, where Zen is most likely to be super-human and have things to teach, particularly if you also get to see Zen's evaluation (getting very rapid feedback about Zen's opinions on different positions).

* Exclusively only using Zen as an opponent would be strange. If Zen is at least close in strength to the players, then there should be a lot of value in also playing games with each other rather than Zen (so as to practice against humans and not just bots), and renting a bunch of servers and running Zen overnight with long-time-control searches on critical parts of those games for analysis for moves where Zen might have useful suggestions to study. The long searches should improve any weakness in Zen a bit and make use of a plentiful resource that would otherwise go unused - overnight computation time.

* Having Zen suggest moves "one by one" in hand-crafted positions would probably be a pretty inefficient way to use a bot. It's the 2%-5% of moves where you make a notable mistake or where the bot would play something innovative and interesting that you want to focus your study, as most of the time it will consider and suggest the same moves you would. Playing games whether with Zen or other players, and having Zen sweep over all the moves in the whole game and finding points where it disagrees is a good way to get around this.

[Krama]

I don't understand the point of zen software when they don't update it for 2 years.
Why make such advancements but then refuse to update software for people who bought it.

[kris]

After 200 games Zen 15.0 is number one. Over 100 ELO better then previous version:
http://www.yss-aya.com/cgos/19x19/bayes.html

[kris]

Mlily Cup 1st Round DeepZenGo won with Shin Minjun 5p:

https://lifein19x19.com/download/file.php?id=8608

[pookpooi]

After 200 games Zen 15.0 is number one. Over 100 ELO better then previous version:
http://www.yss-aya.com/cgos/19x19/bayes.html

Click refresh, see Zen-15.0-1c0g
It's a little bit weaker than Zen-14.8-1c0g
So 15.0 is not the best that Zen could offer
Plus, they begin to test on 14.9, so something might gone wrong in 15.0?
But don't get me wrong, all of these little tiny version perform very similar that the difference is not statistically significant

[Satorian]

After 200 games Zen 15.0 is number one. Over 100 ELO better then previous version:
http://www.yss-aya.com/cgos/19x19/bayes.html

Click refresh, see Zen-15.0-1c0g
It's a little bit weaker than Zen-14.8-1c0g
So 15.0 is not the best that Zen could offer
Plus, they begin to test on 14.9, so something might gone wrong in 15.0?
But don't get me wrong, all of these little tiny version perform very similar that the difference is not statistically significant

Perhaps 15.0 has a more GPU-dependent computation bias that makes it ultimately stronger if given a GPU. And all those incremental improvements add up, if for example compared to something like 13.3, which I would consider to have been a benchmark for a comparatively long time.

[pookpooi]

Zen lost today shows that it's still unreliable even with Chinese rule.
Though FineArt which is stronger than Zen is still not on the 'unbeatable' level.
Update: lost today due to horizontal line effect (basically a popular excuse for strong programs now)

https://twitter.com/webigojp/status/877453072129064961

[kris]

Mlily Cup 2nd Round DeepZenGo lost with Wang Haoyang 6p (of 0.5 point):

https://lifein19x19.com/download/file.php?id=8609

[pookpooi]

Update on DeepZenGo

“DeepZenGo has improved by one stone compared with where it was in March, and our goal this time is to get to the top spot,” said Hataki Kato, the AI’s creator. “Our team wants to make the best Go software to help people.”

LOL at 'Hataki Kato', anyway this is interview just after Zen won in first game and before its lost on second game, so Hideki is being overly confident.

https://www.yicaiglobal.com/news/japane ... tournament

This interview is after it lost on second game

"Everything looked fine at the beginning of the contest until the machine made mistakes in the middle of the game.

DeepZenGo's performance "did not meet my expectations," said Kato, who believes a program bug was caused by technical problems, adding that a misjudgment made the machine confused whether to win or lose.

World top-ranked player Ke Jie sent his congratulations to Wang after the game. The 19-year-old talent insisted on figuring out the weaknesses of DeepMind's Go-playing program AlphaGo and has tried to defeat the machine after he lost a battle in Wuzhen last month.

Wang said he had been defeated by DeepZenGo on Internet games before. But these games were fast, and the contest was slow. Human players normally do better in slow games."

from http://www.ecns.cn/cns-wire/2017/06-22/262555.shtml

and

After the victory, Ke Jie, a Chinese player who is the world’s number one and was whitewashed three to nothing by Google DeepMind’s AlphaGo last month, congratulated Wang, describing him as “the hope of humanity.”

DeepZenGo didn’t play as well as it should have, Hataki Kato, its developer, said after the game, blaming a technical problem within the program. It mistook a dead group on the left side of the board as an alive one, Kato added, saying that such a misjudgment is obviously a major bug in the program.

from https://www.yicaiglobal.com/news/chines ... tournament