Page 3 of 3
Re: Zen19K 8d -- now with Value Net
Posted: Mon Sep 26, 2016 4:55 am
by johnsmith
yoyoma wrote:Here is the server people are talking about:
http://www.yss-aya.com/cgos/19x19/standings.html
Zen 12.5's results:
http://www.yss-aya.com/cgos/19x19/cross ... -1c0g.html
12.5 is 17 wins 16 losses to 12.4, so I don't think you can say it's weaker. 12.4 uses the same naming "1c0g", so I guess they are using the same hardware. If you're going by ratings, 12.5 is a little below 12.4 but it's just because 12.5 hasn't played enough games yet. Probably it's rating will get closer to 12.4 after some more games.
Maybe it was wrong to say weaker, but their policy nets are of a very similar strength. When I was reviewing their games, all those moves seemed very good. But I guess, as you said, they need to play many more games in order to have some statistically relevant analysis. As of this moment v12.5 has less than 50% winrate against v12.4, though it's "just" 1c0g hardware.
Re: Zen19K 8d -- now with Value Net
Posted: Tue Sep 27, 2016 2:45 am
by mschribr
Krama wrote:Why is zen not playing on KGS anymore?
Why should Zen play on KGS? Zen is already number 1 on KGS. Zen should play stronger players. Which servers have stronger players? Can Zen play in human tournaments?
Re: Zen19K 8d -- now with Value Net
Posted: Thu Sep 29, 2016 7:54 am
by pookpooi
In Bayesian Elo page that updated daily instead of real time, 12.5 is now 'slicely' better than 12.4 from 398 games (as I'm writing this 12.5 has played 402 games)
http://www.yss-aya.com/cgos/19x19/bayes.html
Anyway, it's interesting that Zen seems struck at the level around AlphaGo v13. Maybe what DeepMind did to AlphaGo from v.13 to v.18 (to now) is like a secret sauce (main dish is Nature paper). Surprise me, Yamato
Re: Zen19K 8d -- now with Value Net
Posted: Thu Sep 29, 2016 8:13 am
by Uberdude
I wouldn't say "stuck" when it doesn't get stronger over 2 weeks and a minor version release... Yes I expect diminishing returns, but I see no reason for pessimism over Zen's remarkable rate of improvement yet.
Re: Zen19K 8d -- now with Value Net
Posted: Thu Sep 29, 2016 2:41 pm
by ez4u
pookpooi wrote:In Bayesian Elo page that updated daily instead of real time, 12.5 is now 'slicely' better than 12.4 from 398 games (as I'm writing this 12.5 has played 402 games)
http://www.yss-aya.com/cgos/19x19/bayes.html
Anyway, it's interesting that Zen seems struck at the level around AlphaGo v13. Maybe what DeepMind did to AlphaGo from v.13 to v.18 (to now) is like a secret sauce (main dish is Nature paper). Surprise me, Yamato
I am not up to date on everything that has been published. Is the Zen = AlphaGo v13 claim based on anything other than hand waving by the AlphaGo team?
Re: Zen19K 8d -- now with Value Net
Posted: Thu Sep 29, 2016 3:37 pm
by Uberdude
It's probably somewhat based on the way it beat up a load of KGS 9 dans (who are probably around or even above Fan Hui's level), though this does come with the online fast game caveat.