It is currently Sun May 31, 2020 4:31 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 4 posts ] 
Author Message
Offline
 Post subject: More insight on AlphaGo?
Post #1 Posted: Tue Aug 22, 2017 7:12 am 
Oza

Posts: 2513
Liked others: 16
Was liked: 3641
I was amused by the game below (an Oteai game from the 1930s):



What I found amusing was that there were four potential Go Seigen groups on the board. Better still, such groups eventually appeared and acted the way the theory predicts - devastating the territory around them.

A Go Seigen group was a term I coined as a piece of joke journalism, but it turned out to be at least as useful as any proverb. The reference to Go Seigen came about because in his talks on 21st Century Go he made it plain that there were eight very important areas on the board that were grossly neglected. These were the side areas between the star points. He did not specify why they were important but I did some research of my own and discovered (with some more clues from Go, of course) that if a stable group occupied such a side area it had an unexpectedly large impact on the whole of that quadrant. The typical effect was to make the rest of the quadrant a kind of no man's land, a radiation-polluted area where no territory could grow.

There were various practical benefits from this observation. At the simplest level, it's useful to be able to look at a huge portion of the board and say there is no need to count it! But at the strategic level, getting a single Go Seigen group between two enemy groups and creating a field of desolation means you have made roughly twice as many of his stones ineffective as your own.

For this reason, it is relatively rare to see full Go Seigen groups in pro games - pros know they have to be strangled at birth. But when you do see them, the effect is clear and huge.

In the game shown above, the various side groups are not yet fully fledged GSG groups. For that they need to have two clear eyes (mere occupancy is not enough). But some did acquire them, and as an example of the effect, the whole upper left quadrant (100 points) finished up with just 10 points for Black (and two for White) while the upper right ended up with four points for White and three for Black.

The lower left started off as a potential GSG group but a ko fight meant it got linked to the lower right side and so it ended up differently. The one on the left side got involved in a trade. But, still, two/three GSG groups in one game! Better yet, White had a bigger GSG quotient and won by 8 (no komi).

That got me thinking about GSG groups in AlphaGo games. Several of us amateurs have spotted the similarities between Go and AlphGo, and pros have too. Indeed, the forthcoming book on AlphaGo by Yoda - which I'm hoping will be the book event of the year - is precisely based on that (along with similarities with Dosaku).

So I looked at the 60-game series by AlphaGo in its Master embodiment. I was very surprised.

Apart from the one case where Zhou Junxun played mimic go (game 51) and game 42 where a GSG group resulted from a ko trade rather than a deliberate choice, AlphaGo played a GSG group in only two games. In half a dozen games it has the chance to start making a GSG group but eschews it, and in game 22 it seems studiously to avoid it with a novel centre strategy.

In contrast quite a lot of the humans succeeded in making GSG groups themselves (but still lost!).

When AG has the chance to make a GSG group but doesn't, it opts instead to play out to the centre. But actually that is following the GSG line. Although Go did refer to the interstellar side areas it was me who made something concrete out of that. He himself was on a much higher plane, and was alluding rather to the central portion of the quadrant - the side area was just some sort of Launchpad for activities out in the centre. I never figured out what these were, but it looks as if AG did.

Yet I believe we should still give Go priority over AG. AG's performance in winning 60-0 against top pros is, of course, magnificent. But, as far I can recall, Go at his peak was scoring 85% to 90% (his lifetime score was about 70%) which included giving handicaps. And unlike AG, which never gets tired or upset, Go had a lot of stress in his life and so we'd have to expect more than the odd mistake.

Above all, Go could encapsulate some of what he knew in words we can understand. For the time being, he is surely still the truest Master.


This post by John Fairbairn was liked by 4 people: Bill Spight, gowan, Marcel Grünauer, pookpooi
Top
 Profile  
 
Offline
 Post subject: Re: More insight on AlphaGo?
Post #2 Posted: Tue Aug 22, 2017 8:49 am 
Lives in gote

Posts: 372
Liked others: 347
Was liked: 190
Very interesting, the first time I heard about "Go Seigen groups".

I followed a lecture of Yamashiro Hiroshi 9p recently during the US Go Congress where he talked about AlphaGo.
He did mention that from studying AlphaGo games, it is pretty clear that it avoids developing the sides of the board, as long as it can build either some corner, or the center.

So his conclusion was that while the traditional go theory orders these areas as:
1. corners
2. sides
3.center

, AlphaGo seems to disagree and orders them:

1. corners
2.center
3.sides


This post by sorin was liked by: Bill Spight
Top
 Profile  
 
Offline
 Post subject: Re: More insight on AlphaGo?
Post #3 Posted: Tue Aug 22, 2017 8:57 am 
Honinbo

Posts: 9753
Liked others: 3107
Was liked: 3212
Very interesting observations and commentary. :D

John Fairbairn wrote:
But at the strategic level, getting a single Go Seigen group between two enemy groups and creating a field of desolation means you have made roughly twice as many of his stones ineffective as your own.

For this reason, it is relatively rare to see full Go Seigen groups in pro games - pros know they have to be strangled at birth. But when you do see them, the effect is clear and huge.

{snip}

That got me thinking about GSG groups in AlphaGo games. Several of us amateurs have spotted the similarities between Go and AlphGo, and pros have too. Indeed, the forthcoming book on AlphaGo by Yoda - which I'm hoping will be the book event of the year - is precisely based on that (along with similarities with Dosaku).

So I looked at the 60-game series by AlphaGo in its Master embodiment. I was very surprised.

Apart from the one case where Zhou Junxun played mimic go (game 51) and game 42 where a GSG group resulted from a ko trade rather than a deliberate choice, AlphaGo played a GSG group in only two games. In half a dozen games it has the chance to start making a GSG group but eschews it, and in game 22 it seems studiously to avoid it with a novel centre strategy.


In AlphaGo's published self play games I think that it rarely makes GSG groups. It's play is fluid. But against humans it often plays quite solidly, so I am also surprised.

Quote:
In contrast quite a lot of the humans succeeded in making GSG groups themselves (but still lost!).


wineandgolover wrote:
AlphaGo - Whatever You Do Is Wrong

;)

Quote:
When AG has the chance to make a GSG group but doesn't, it opts instead to play out to the centre. But actually that is following the GSG line. Although Go did refer to the interstellar side areas it was me who made something concrete out of that. He himself was on a much higher plane, and was alluding rather to the central portion of the quadrant - the side area was just some sort of Launchpad for activities out in the centre. I never figured out what these were, but it looks as if AG did.


That reminds me of Korshelt's rule of thumb: one eye and access to the center, which, I believe, he learned from Murase Shuho. (Curiously, I haven't seen it anywhere else than Korshelt.) And yes, AlphaGo seems to "understand" the center better than any other player of this time.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: More insight on AlphaGo?
Post #4 Posted: Wed Aug 23, 2017 12:03 pm 
Gosei

Posts: 1435
Liked others: 392
Was liked: 407
Rank: 5d
GD Posts: 1000
I just received Yoda's book from amazon.co.jp. In this book, titled AlphaGo Kenkyuu, Go Seigen Dousaku Yomigaeru ISBN 978-4-8399-6299, Yoda studies the ways Master's play resembles that of Go Seigen and Dosaku, as John has mentioned above.

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 4 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: No registered users and 1 guest


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group