alphago self-analysis by fan hui

dhu163 · Post by **dhu163** » Sun Jul 09, 2017 12:24 pm

https://www.youtube.com/watch?v=nljNZT4UP3U

"Liu: Though all our games have had some intense fighting, but it's still not beyond comprehension/imagination

and today we invited Lian Xiao, who is known for his fighting strength.

Fan: today's game, it would be hard to find a game more complicated than this. The fighting is very messy, from start to end. All sorts of exchanges, often in places you can't imagine."

I'm up to move 22, at 20 mins out of 133. I'll continue working on it tonight. The intro raised our expectations high for an exciting game. It's probably particularly worth it to go through this one slowly. (19:24)

edit: I'm up to move 108, at 60 mins. Fan Hui has finally said the opening ends here, and the complications start (22:56)
edit: 101 mins move 206, sorry I haven't found the time/energy to continue this last week
edit: all done

aeb · Post by **aeb** » Fri Jul 14, 2017 4:46 pm

dhu163 wrote:I'm up to move 108.

I added the five SGFs to my AlphaGo page http://homepages.cwi.nl/~aeb/go/games/games/AlphaGo/. Corrections and comments are welcome.

luigi · Post by **luigi** » Sun Jul 16, 2017 2:54 pm

I'm surprised that DeepMind doesn't mention these five new games on its official site. Have they publicly announced them to a Western audience at all? Or have they just showed them to the Chinese pros that commented them?

dhu163 · Post by **dhu163** » Wed Jul 19, 2017 3:47 pm

sorry for taking so long, I've had orchestra practice, and got too tired. I've completed the game 5 review above.

Yes, sad they only put it out in Chinese.

Baywa · Post by **Baywa** » Thu Jul 20, 2017 4:21 am

That was like in one of these martial arts movies where the ninjas / ronin are swirling through the air and running up the walls while simultaneously kicking, slashing and punching in all directions... :rambo:

White won by 1/2 point although Black had 13 more prisoners (43:30). So, was making the seki in the lower right corner the decisive move then?

dhu163 · Post by **dhu163** » Wed Dec 20, 2017 1:01 pm

I thought that the AGA were meant to be translating these, but I haven't heard anything except video 2 on the youtube channel.

I make no predictions about when I will spend time on this, given study, but I do hope to do the whole set of 5.

So far 32 mins of 135 on 1. Chang Hao 9p talks a lot ... (edit: up to 52 mins)(edit: 2 hours to get up to 80 mins)(edit: 1 hour to get up to 94 mins, a fascinating discussion here)(spent 2 more hours to get up to the full 134 mins)

dhu163 · Post by **dhu163** » Fri Mar 23, 2018 9:25 am

just a note that I've been working on the above translation today, a present for you all. It is shocking/amazing how much Chang Hao talks, but the section 80-90 mins is particularly fascinating for fans of alphago and its sense of the centre.

For future bits, I'll try to cut how much I note down for efficiency. I anticipate another update within 3 months.

John Fairbairn · Post by **John Fairbairn** » Fri Mar 23, 2018 11:12 am

Thank you. Great stuff.

Just a couple of points to get a discussion going:

Chang: it isn't so concerned with local profit and loss

Fan: yes, normally, as long as you play towards an open area, it can't be too bad a move. Wherever is wide and open it will play there.

I have often mentioned the Japanese word "hiroi" as a go term. Ordinarily it just means wide, but the strategic idea in go is always "wide and open" and pros often mention that these are the places to play (and actually that overlaps with what is meant also by "big point" which is "big area" rather than point. Although there have been a couple of books on the topic, it has long been my impression that pros give this advice but don't take it often. They prefer territory. Perhaps AG is saying the pros must heed their own advice more.

Liu: efficiency

Fan: perhaps it thinks that stone efficiency is more important.

I have just been browsing again through a book by Takemiya on Positional Evaluation without counting, subtitled "Endeavouring not to surround territory". His claim is that if you look at a position in terms of good shape/efficiency you have all you need to evaluate the position. If one player's moves are all efficient and the other's aren't, we know at a glance who is ahead. From this it follows that all you need to do is make sure all your stones are efficient and not to worry about territory. It will come naturally. This is what he means when he tries to reject the cosmological labels attached to his style and asks people to use "Natural" as the epithet instead. (He mentions this so often, he presumably finds it intensely irritating.) In taht connection, he also says (often) that one of the worst things taught to beginners is that "go is about who has the most territory." If you have to talk about territory you must stress that go is about who has the most territory at the end of the game. Anything before that is about efficiency.

So Takemiya's ideas seem in synch with what AG is telling us. Yet how many pro players follow ex-world champion Takemiya? Precisely! They nearly all prefer to take territory.

Move 8 concerns the attachment against the small knight's shimari from 4-4. Chang Hao remarks that he was taught that this was either a handicap-game play or an expedient to make sabaki. But there are over 1,000 games in the GoGoD database with that play. It was introduced by Hashimoto Utaro in the 1930s and he played it very often for a long time afterwards.

As I have mentioned probably too often before, many other AI moves have been prefigured in Shin Fuseki (i.e. the 1930s). It seems to me that modern pros neglect the lessons of the human past too much. WE seem to get the same phenomenon in chess, but top players there hire weaker players to do analysis of past games for them.

As well as listening more to their own advice, maybe go players should likewise be going back to the archives?

Tryss · Post by **Tryss** » Fri Mar 23, 2018 1:09 pm

John Fairbairn wrote:As I have mentioned probably too often before, many other AI moves have been prefigured in Shin Fuseki (i.e. the 1930s). It seems to me that modern pros neglect the lessons of the human past too much. WE seem to get the same phenomenon in chess, but top players there hire weaker players to do analysis of past games for them.

And how many ideas in Shin Fuseki had been discarded (for now)? How do we know which lessons of the past are correct and which are not?

Uberdude · Post by **Uberdude** » Fri Mar 23, 2018 3:19 pm

Thanks dhu, that building the centre thickness and a fair sized territory box to sacrifice the hard to convert to territory lower moyo is indeed interesting.

John Fairbairn wrote:Move 8 concerns the attachment against the small knight's shimari from 4-4. Chang Hao remarks that he was taught that this was either a handicap-game play or an expedient to make sabaki. But there are over 1,000 games in the GoGoD database with that play. It was introduced by Hashimoto Utaro in the 1930s and he played it very often for a long time afterwards.

What are you classifying as "that play". The AlphaGo-ism is to play the attachment when there are no other stones in that quadrant (so generally quiet positions e.g. this move 35), rather than in response to a pincer (e.g. 48 in same game) as Chang Hao mentioned (doing so in response to a 2-space pincer is particularly common in recent human pro games); are there 1000 of those? Doing a mostly GoGoD database search of an otherwise empty quadrant plus one line (because a stone a line from the midpoint would still serve as a pincer in a kick and 2-space extension shape) with white next (AG also likes to tenuki, not a new idea either, but probably it is more fond of tenuki here than most humans were) I get ~14k hits: most common is the slide at a (6k), with side extensions of b and mini-Chinese style c ~2k each. The attachment is down at letter n and off the kombilo chart, with 57 hits. So the new idea of AlphaGo is that this attachment is basically the default move in this position, rather than a rare one, down at #14 and played under 1 time in 200. So yes 57 is quite a lot more than 0, but a lot less than 1000, and going from the 1-in-200 choice to the #1 choice is a big change in opinion. I suppose another part of this is not just liking the attachment, but AG's generally dismal assessment of the slide, which was previously seen as an often decent move.

Click Here To Show Diagram Code: [go]$$B $$ ----------------------+ $$ . . . . . . . . . . . | $$ . . . . . . . . . . . | $$ . . . . . X . . . . . | $$ . , . . . . . X . a . | $$ . . . . . . . . . . . | $$ . . . . . . . . O . . | $$ . . . . . . . . . . . | $$ . . . . . . . . . . . | $$ . . . . . . . . . . . | $$ . , . . . . . b . . . | $$ . . . . . . . . c . . |[/go]

P.S. Here is the Chang Hao vs Wang Yao game mentioned with Chang playing a 5th line move: http://ps.waltheri.net/database/game/43174/

sorin · Post by **sorin** » Fri Mar 23, 2018 5:25 pm

I wonder how big is the difference between the traditional slide at a compared to AG's preferred tsuke.
It may be so small that it does not matter for practical purposes in human play.

I noticed that FineArt prefers the slide at a in fact: I watched many of its 4 handicap games on FoxGo server and it seemed to almost always start with approach, slide and 2 space jump (when black replies the slide at 3x3). Which by the way seems a very dull/slow way for white to start in a 4 handicap game, but FineArt somehow managed to still win many such games against Fox 9d amateurs.

John Fairbairn · Post by **John Fairbairn** » Sat Mar 24, 2018 1:58 am

What are you classifying as "that play". The AlphaGo-ism is to play the attachment when there are no other stones in that quadrant (so generally quiet positions e.g. this move 35), rather than in response to a pincer (e.g. 48 in same game) as Chang Hao mentioned (doing so in response to a 2-space pincer is particularly common in recent human pro games); are there 1000 of those?

I don't entirely follow you and you don't seem to follow me, so I infer we are on different wavelengths

My fault, no doubt.

I did use a different quadrant from you, but my purpose was not to show one thing is better/bigger than another. All I want to show is that there is a lot of human data out there that can be mined. At the risk of inducing groans, let me say again that what was special about the 1930S and the Shin Fuseki period was not just the similarity of ideas with the AIs but that pros then commented copiously on their moves. Their new ideas were never really repudiated. Although the onset of komi had a major effect, they basically gave up because it was too hard. Since the AIs have shown that it is not, however, impossible, my suggestion is that we revisit what the old pros said in the light of what the AIs now show.

In passing, I might add that another special aspect of the 1930s theorising was that it was abstract, almost geometric. It was not concerned (as much of the discussion of the AIs moves seems to be) with josekis. There are some current pros who are trying to see the big picture, of course, and Chang Hao seems to be one of them. A very good example is Ohashi Hirofumi who has a brand new book on the AIs just out which doesn't really touch on josekis at all. The sort of thing he is pursuing is reclassifying shoulder hits not as erasures but as an attacking tool. (I think Go Seigen made the same point but in perhaps a better way by focusing more on the follow-up move.)

At some point, if modern pros are going to benefit from AIs they are going to have to learn how it plays the middle game, not the josekis, surely? The josekis at best are just signposts on a map. The AIs' win-rate numbers may give us grid points and contour lines but what the humans have said before in words might tell us more, in a Rough Guide or Lonely Planet way, about where to start and what is worth seeing.

To revert to a specific example in the current discussion, I recall the slide being criticised by a human pro (I think it was Takagawa). His point was that if the opponent ignores it (which is usually easy to do because there is room to extend along the other side), the only sensible follow up for the slider is to take the 3-3 point. But then the slide itself, in tewari terms, is not on a point the player would have chosen if he had the kakari and the 3-3 in place first. The point being made wasn't about the joseki per se but about efficiency.

Uberdude · Post by **Uberdude** » Sat Mar 24, 2018 4:29 am

What I mean is playing the attachment in this small shaded region is not remarkable, loads of humans did this before AlphaGo. Often in response to a black pincer around a.

Click Here To Show Diagram Code: [go]$$B $$ ----------------------+ $$ . . . . ? ? ? ? ? ? ? | $$ . . . . ? ? ? ? ? ? ? | $$ . . . . ? X ? ? ? ? ? | $$ . , . . ? ? ? X O ? ? | $$ . . . . ? ? ? ? ? ? ? | $$ . . . . ? ? ? ? O ? ? | $$ . . . . ? ? ? ? ? ? ? | $$ . . . . . . . . . . . | $$ . . . . . . . . a . . | $$ . , . . . . . , . . . | $$ . . . . . . . . . . . |[/go]

But the new idea from AlphaGo is that when looking at this larger shaded region with no additional black or white stones in it, the attachment is the default best move if white is going to play here. When I do a human game database search of this region, the attachment is the 14th most popular choice, with 57 examples in 13790 games, so not unheard of, but pretty rare. I didn't actually check just on AG games (and we don't have a huge number), but I think it's the #1 choice for AG.

Click Here To Show Diagram Code: [go]$$B (imagine the shading includes star points, that's just diagram limitation) $$ ----------------------+ $$ ? ? ? ? ? ? ? ? ? ? ? | $$ ? ? ? ? ? ? ? ? ? ? ? | $$ ? ? ? ? ? X ? ? ? ? ? | $$ ? , ? ? ? ? ? X O ? ? | $$ ? ? ? ? ? ? ? ? ? ? ? | $$ ? ? ? ? ? ? ? ? O ? ? | $$ ? ? ? ? ? ? ? ? ? ? ? | $$ ? ? ? ? ? ? ? ? ? ? ? | $$ ? ? ? ? ? ? ? ? ? ? ? | $$ ? , ? ? ? ? ? , ? ? ? | $$ ? ? ? ? ? ? ? ? ? ? ? |[/go]

John Fairbairn · Post by **John Fairbairn** » Sat Mar 24, 2018 4:45 am

Yes, I understood that part of what you were saying.

My point is that it is definitely not new even in a big empty quadrant. Great players such as Hashimoto, Kajiwara, Ishida, Fujisawa Hideyuki and many others have played it, even in empty quadrant areas as big as 14x14.

What you appear to be saying, as best as I can work out, is that what is new is that AG regards that as the default move. OK, up to a point, Lord Cropper. But surely when these great humans played the move (and very consistently in cases such as Hashimoto Utaro), they thought the attachment should be the default, just like AG. So let's listen to why they thought that. They can talk. AG can't.

Uberdude · Post by **Uberdude** » Sat Mar 24, 2018 5:10 am

John Fairbairn wrote:What you appear to be saying, as best as I can work out, is that what is new is that AG regards that as the default move. OK, up to a point, Lord Cropper. But surely when these great humans played the move (and very consistently in cases such as Hashimoto Utaro), they thought the attachment should be the default, just like AG. So let's listen to why they thought that. They can talk. AG can't.

Yes. Averaging over many humans the attachment is a rare (in % terms, 57 is still quite a few examples) move in the large empty space, but maybe some particular humans liked it. We've already seen how Go Seigen liked many of the moves AG does too (but I'm not sure how many you would find in his games if he was having these ideas in later life in study groups when he was no longer so active in tournaments). Doing the same quarter board plus 1 line search on just Hashimoto's games (1500, wow! didn't expect so many) with him as white playing next in this area I get 22 hits, with him sliding 8 times (#1 choice) and attaching twice (joint #3-6). So he still liked to slide as default, but attach was occasional rather than rare.

Next I'll try to do the same search on some AG games, maybe the 50 Master self play, though I have a feeling AG Zero likes attach more.

Update: An important part of science is to state your hypothesis before testing and publish results regardless

. This is new install of kombilo with the 50 Master self-play and I seem to have the next colour setting messed up, but with white (or black if colours are swapped) to play next in the big quadrant I get 15 games in total (4 with black 4-4, 11 with white, is this Master preferring 4-4s over 3-4s more when white?). Slide has 0 hits, as expected. But most popular is extend to hoshi (5), one point jump up (3), mini-Chinese extend (3), and attach is only 4th with 2.

Update 2: Doing the search on the 20 AG Zero vs AG Master games, there are 7 hits, now attach is #1 with 5, and 1 each of one point jump out and a stone appearing on top side as part of a group there. That's the result I was thinking of, though we only have 20 games so not a great sample (and the openings are all pretty much the same for a while).

Life In 19x19

alphago self-analysis by fan hui

Re: alphago self-analysis by fan hui

Re: alphago self-analysis by fan hui

Re: alphago self-analysis by fan hui

Re: alphago self-analysis by fan hui

Re: alphago self-analysis by fan hui

Re: alphago self-analysis by fan hui

Re: alphago self-analysis by fan hui

Re: alphago self-analysis by fan hui

Re: alphago self-analysis by fan hui

Re: alphago self-analysis by fan hui

Re: alphago self-analysis by fan hui

Re: alphago self-analysis by fan hui

Re: alphago self-analysis by fan hui

Re: alphago self-analysis by fan hui

Re: alphago self-analysis by fan hui