alphago self-analysis by fan hui

[dhu163]

a translation of https://www.youtube.com/watch?v=n06tQJiQ680, a 2 hour video, with a lot of variations

the latest is that weiqitv will look at 5 games of alphago self-play, with longer time controls than the 50 self-play published at 2 mins/move, or 8-10 hours/game

this is the 2nd of those 5

[pookpooi]

Is this Japanese rule?
So we get even more AlphaGo games?

[lichigo]

Wouahhhhhh, thank you so much. I tried to watch but i couldn't understand. Apparently they released 5 games, so I guess 5 videos.

[John Fairbairn]

Is this Japanese rule?
So we get even more AlphaGo games?

In China this game was given as B+0.75 with Chinese rules. As far as I know only two games have been released so far.

FineArt has also been in recent action again, beating a trio led by Gu Li.

In Japan, DeepZen (immediately after its Mlily debut) began a long series of games against the Japanese National Training Squad. I have about 50 games and DeepZen lost 6, from a quick count. What is striking about these games is how many of them end so early - and maybe only two were counted. In one, the human won by 0.5 (Japanese rules were used here). I think one of the most satisfying things on earth must be to beat a Monte Carlo computer by 0.5!. I believe Ueno Asami can now also claim the distinction of being the first woman to beat a computer in an even game, which must have some significance considering the huge growth in women's go in Japan at present (all down to some pop singer whose name I've forgotten).

[dhu163]

I don't know about rule set but fan said 7.5 komi and B+0.5, I.e. 0.25 stones

[Baywa]

I played that game to the end and got the following end position:



In that "roll-out" (OHNE GEWÄHR! ) white filled the last dame-point and the result is B184:W177 (B66:W59), so B+0.5 under both rule-sets. Under the old komi settings of the first 50 games, Black would have lost - but both Black and White would have played differently then. I wonder if these games no. >50 are really played under Japanese rules.

[Uberdude]

Baywa: there's a (double) problem at f2.

[Baywa]

Oh yeah! White can't fill in that liberty. But then, when Black does, White has to cover inside. That way Black would win by 1.5 points... There must be something wrong with my sequence.
Edit: There was that note that White has to play B9 eventually. Right now, I don't see that. Black would have to play inside Whites territory to force that move but that wouldn't gain anything. Or am I wrong?

[moha]

There was that note that White has to play B9 eventually. Right now, I don't see that.

B B8 after dame fill?

[Baywa]

OK, moha, got it! So White has to play B9 before all the dame are filled in. But then (with Uberdude's remark) I'm back to my B+1.5 score. I'll figure it out (sooner or later).
Edit: Argh! I didn't read dhu163's post. Then it's Chinese rule, 7.5 point komi and B+0.5. So everything's fine.
Edit2: I'm still at it (sigh!). Now I'm "convinced" that B+9 on the board (Chinese) / B+8 (Japanese). But never mind!

[Simba]

Thanks for the translation . Some interesting stuff in here, especially the new 3-3 invasion at the top left, with AlphaGo directly attaching under rather than just bumping as we're used to. We might have to switch to that one in our practice games now . The comment about how AlphaGo's games with itself are a complete mess whereas in games against humans it plays just solid and slowly wins without doing anything special was interesting too.

[shoryuu]

the new 3-3 invasion at the top left, with AlphaGo directly attaching under rather than just bumping as we're used to. We might have to switch to that one in our practice games now

You're giving it too much credit. The attach under has been around for a long time now. As far back as the 90s at least.

[Simba]

the new 3-3 invasion at the top left, with AlphaGo directly attaching under rather than just bumping as we're used to. We might have to switch to that one in our practice games now

You're giving it too much credit. The attach under has been around for a long time now. As far back as the 90s at least.

lol, I'm aware of this I'm making the point that AlphaGo until now preferred to bump rather than attach underneath, and as such dhu and I have been playing the bump exclusively in our practice games. So now we'll have to go back to the usual way instead.

[Dontbtme]

Just passing by to tell dhu a big THANKS for translating the video! It's been a real pleasure to go through the sgf and I know it must have been hours of work

[dhu163]

Thanks, Yeah, it probably takes me double the time of the video to make this sgf. (edit: it's probably more like 6 hours ...)

https://www.youtube.com/watch?v=vD-Z1-V1tTQ

game 4


My feeling is that games 1 and 2 were long games and still have the most interesting variations.



I'm pretty sure I've mistranslated the title of this video