Life In 19x19

yoyoma wrote:They aren't exactly straightforward about it. Yes earlier Aja Hwang had said that deepmind was his own personal account. But later there was a Korean article with a picture of Demis Hassibis and the CEO of Tygem. The title was something like "Thanks Tygem!". Even in that article it's sorta like Demis is saying it's Aja's account, but he sorta "indirectly" confirmed it. It's like they are saying it's Aja's personal account, but they are no longer directly denying that he used it for AlphaGo anymore.

I want to be careful about this too but at this point is just seems overwhelmingly likely that AlphaGo is playing on that account, just for some reason Deepmind just doesn't want to publicize this. That's my take at least.

I think you're probably right, looking at that article (I assume it's the same one). Probably Aja Hwang created the account, and they are playing AlphaGo's moves on it (at least) now.

I'd guess they are probably not going all out to win every game on Tygem, but probably, they are using it as a testing ground to get new ideas from humans, and to iron out bugs (like AlphaGo's response to Lee Sedol's move 78 in Game 4, for example).

uPWarrior wrote:I think that graph is believable is you treat "9d" as "average 9d". In that case, Lee Sedol himself would not be 3500 but probably 3800+.

Remember that these scales are arbitrary anyway.

I assume that the scale they are using is the one of goratings.org, where Lee Sedol is 3518.

I also checked goratings before posting, but as far as I know there is no indication that they are using it. I just offered a possible explanation.

uPWarrior wrote:I also checked goratings before posting, but as far as I know there is no indication that they are using it. I just offered a possible explanation.

They're using it, it stated in nature paper, and the nature AlphaGo rating in nature paper is the same as this AlphaGo v13 in this graph.

If you see a slide that says "CAUTION", it's good to take that into account.

The word covered by the speaker's head is "self". Bot results in self-play are always massively exaggerated. It's not uncommon to see a 75% self-play winrate in selfplay to translate to 52% winrate against a third-party reference opponent. c.f. fig 7&8 in http://pasky.or.cz/go/pachi-tr.pdf . Intuitively, I'd expect the effect to be less pronounced with very strong programs, but we don't know anything about the mechanics here and experiments are difficult.

It's no doubt today's AlphaGo is much stronger than the Nature version. But how much? We'll have a better idea when they pit it in more matches with humans, and ideally when other programs catch up further. Without knowing more (like the rest of the slides or a statement by someone from Deepmind), I wouldn't read much into this graph.

pass

djhbrown wrote:you have all been hoaxed; the photo is a fake.

On what basis do you say this?

djhbrown wrote:you have all been hoaxed; the photo is a fake.

fake? Aja Huang is already comment about this graph

"Note that recently for Crazy Stone and Zen improvements in self-play also
transferred to playing strength against human players. According to Remi
and Hideki, Crazy Stone and Zen are both >=80% stronger with a policy net
and they both reach 7d on KGS (1 stone stronger).

But generally I agree that we should be cautious about self-play results.

Aja"

This photo is taken by Thai student that attend UCL lecture by David Silver