Zen 15.0 on CGOS very strong new version

[kris]

Until now only 10 games, but all of them won!
http://www.yss-aya.com/cgos/19x19/cross ... -2c1g.html
Maybe new hero is born?

[pookpooi]

Hard to compare because it's 2c1g (2 cpu + 1 gpu). The strongest and latest 1c0g is still 14.8. But I get that the developer want to test if it works well on better hardware because Zen has to participate in MLily Cup next week.

[kris]

So you think it's exactly this version with only 2CPU & 1GPU that will play on MLily Cup as DeepZenGo?

[pookpooi]

So you think it's exactly this version with only 2CPU & 1GPU that will play on MLily Cup as DeepZenGo?

It'll be this version but on a much better hardware. They just try out if multi CPU and GPU work correctly. If it's not exceed Zen-14.7-2c1g then something might gone wrong. They'll fix it, or simply revert back to using Zen-14.7-2c1g version.

[kris]

After 20 games still unbeatable!
http://www.yss-aya.com/cgos/19x19/standings.html

[Uberdude]

Not lost yet, here is a game vs Zen 14.6

[pookpooi]

Zen-15.0-1c0g is on the way. This'll be easier for comparison

[Friday9i]

Zen-15.0-2c1g will probably be significantly above 4000 Elo on CGOS: it is 13/13 against Zen-14.6-1c1g, which is around 3815 Elo!
Impressive...

[Friday9i]

Their rating is quite conservative (and I'd say a bit strange): Zen15.0 has a 100% score against all programs (one being 3580 Elo) excepted Zen14.6 where it has 90% win rate (18 wins over 20 games), and Zen14.6 is around 3815 Elo.
Normalyy, a 90% win rate is around 380 Elo difference (as they state it here : http://www.yss-aya.com/cgos/), so Zen15.0 should be around 4200... Maybe a bit less, because the 90% is not well established and the true % may be only around 80%, but that is still a 250 Elo difference, hence Zen15.0 is very probably 4050 or more Elo, but it's currently rating is "only" 3968. Hence, in a few days, it should continue to go up and stabilize around 4100 / 4200, we'll see.

[pookpooi]

In the next thirteen hours check this page
http://www.yss-aya.com/cgos/19x19/bayes.html
For more accurate rating and ranking.

[Friday9i]

In the next thirteen hours check this page
http://www.yss-aya.com/cgos/19x19/bayes.html
For more accurate rating and ranking.

Nice, didn't see that before, thanks!

[lightvector]

Auuhhh, when are they going to stop making it stronger and instead prepare a commercial release?

Anyways, this is really cool, thanks!

[goame]

Auuhhh, when are they going to stop making it stronger and instead prepare a commercial release?

Anyways, this is really cool, thanks!

Yes it's now time to buy Zen 15. or it will be useless to develop it, when no one can buy it and use it.

[pookpooi]

From http://www.nihonkiin.or.jp/news/release ... ngo_1.html
Japan Go Association will use DeepZenGo to strengthening Japanese National Team until the end of this year
So Zen may exclusively used within professional body for a while.

[John Fairbairn]

Japan Go Association will use DeepZenGo to strengthening Japanese National Team until the end of this year
So Zen may exclusively used within professional body for a while.

I would be interested in hearing from someone, such as uberdude, who is actively trying t get strong in actual play, what they think of the training method used here.

The national training squad is essentially an ongoing tournament of quickplay games at 1 minute per move and the DeepZen games will be same (via the internet). Since DeepZen is not yet demonstrably much stronger than the human opponents/coaches, I'm not sure what extra the players will get out of the DeepZen games, beyond perhaps a nebulous familiarity with computer-style play.

Given that the purpose of this exercise is to boost the strength of the squad, would it not be more logical to allow them to use DeepZen to suggest moves in test positions input by the pros? I can understand that amateurs might find it boring to do this with databases, but when national pride is at stake.... And why no tests at longer time limits? After all, pros will be playing DeepZen at longer time limits in the Mlily.

Or do I underestimate the value of actual play and overestimate thinking about go theory?