It is currently Wed Jul 17, 2019 3:17 am

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 10 posts ] 
Author Message
Offline
 Post subject: Yoda's book on AlphaGo
Post #1 Posted: Thu Sep 07, 2017 2:51 am 
Oza

Posts: 2334
Liked others: 15
Was liked: 3352
I had been looking forward intently to Yoda Norimoto's book on AlphaGo.* I was a little disappointed, to be honest. That was simply because it was not the book I was expecting. I was expecting a book on AlphaGo. After all, the main title is "Studying AlphaGo." But I was looking forward to the book mainly because it promised to shed new light on Go Seigen through the prism of AG. It does that (and on Dosaku, as bonus) - the subtitle tells us the book is about bringing GSG and Dosaku back to life. On top of that, the biggest segment of the book is a detailed commentary of Kamakura Game 1, and since I wrote a detailed book about the Kamakura match that is sheer heaven. So how can I possibly say I was disappointed?

I do expect that disappointment to disappear, quickly, but while I've still got a bit of residual grumpiness about it, I suppose it's because I expected rather more about AG - as per the main title.

Let's look at the structure of the book, by way of explanation.

In his preface Yoda welcomes AI to the go world. He thinks it will spread the game and increase its charm. He is impressed by AG/Master (and interestingly doesn't mention Zen), above all by the fact it has good suji, which is is high praise from the man who invented the theory of sujiba, I suppose, and innovativeness. It already surpasses humans (but, as he makes plain later, still makes mistakes, he believes).

He explains that his main focus is to shine new light on old games by using his analysis of AG's style. He says that, "From childhood I have played over Go Seigen's games in particular countless times." (And if you think about it, that might be the most important message in the book.) He plans to illustrate GSG's inventiveness and Dosaku's mastery of sabaki. He mentions Showa Go (which usually means New Fuseki) just once, but doesn't touch on it again, except indirectly through GSG's games. This is all a matter of familiarity, no doubt: it is plain from his commentaries that his way of working was to play over an AlphaGo/master game and say "Oh, that reminds me of GSG's game against ...." (In contrast, very, very, very few pros are now familiar with the run-of-the-mill New Fuseki games, but I can tell you that the correspondences with AG are astounding.)

Chapter 1, about a quarter of the 300-page book, is a survey of some games by Master, lightly annotated to adumbrate its main features. These, according to Yoda, are: (1) the two-space shimaris, (2) the shoulder hits, and (3) the presses - kake. Some surprises, then. No mention of the early 3-3 invasion, or the contact plays. And a lot of emphasis on kake.

Actually the contact plays are dealt with mostly in the Dosaku commentaries (part of the sabaki theme there) and the kakes are also a feature of Dosaku's games. The 3-3 invasions are also covered sporadically in the commentaries and there is one 3-3 game specially added at the end to show Yoda playing Ohashi Hirofumi (a colleague with a new book on AG himself) where they are both influenced by the new AG style. So you could say it's all there (excluding any discussion endgame of AG's endgame) but not packaged in quite the most convenient way.

Chapter 2, about a third of the book, is a set of 12 games by GSG, and by far the biggest commentary is on Kamakura Game 1 vs Kitani. This is the one ("The One with the Nosebleed") where GSG famously started with a provocative three-stone wall in the centre and had Kitani going through contortions to attack it efficiently. The idea is to revisit GSG's games through the prism of AG, but rather than producing a band of colours, Yoda prefers to focus a bright white light on GSG's inventiveness. Several of us on L19 have commented on the similarities between GSG and AG, and I think they all stand, but Yoda has found even more, and more subtle ones, so this is a very valuable contribution.

Incidentally, it is not all hero-worship by Yoda. While admitting that sometimes he doesn't know what is going on in GSG's games, he is not afraid to criticise. There is one point where he criticises a forcing move GSG played early as a probe. GSG had commented that a probe cannot be bad, but Yoda disagreed. Ironically, although Yoda appears to have missed the connection, the predilection of AG similarly to make early forcing moves has also been remarked upon here.

Chapter 3 presents 10 games by Dosaku and 2 by Yoda, much more lightly annotated than the GSG games. The focus here is on shoulder hits, contact plays and kakes. In other words, Dosaku also foreshadowed AG.

The final short chapter covers 4 games by AlphaGo as opposed to Master (three being self-play games), and the Yoda-Ohashi game already mentioned. Interestingly, one of the so-called innovations of AlphaGo, the White peep at P5 against Black's sanrensei before sliding into the corner, is something Yoda said he had previously considered playing himself. Perhaps on different grounds? Yoda liked the logic of getting it in before the slide because otherwise Black will not answer it as he will have a stone at the 3-3 point by answering the slide. But AG seems to just like early centre-facing forcing moves in general.

There is no index - a standard failing in Japanese books of any ilk - but there is a "sidebar" page. This may seem to be specially for Robert :), being entitled "Don't think. FEEL!" But it could apply to Yoda and the eerly force P5, could it not? Yoda after all is borrowing the advice from Bruce Lee. He is trying to convey how we should be reacting to AG's self-play games. In these he senses the kind of awe summoned up by Tetsuka Osamu's manga "Phoenix." Obviously this is meant to convey a feeling of viewing the game from the empyrean - literally, as space firebirds! - but it also hints at his theme of "resurrection" of the games of past masters.

So, to summarise, what you get out of this book is a dish made of perfect organic ingredients but cooked in the nouvelle cuisine way. A masterpiece on the palate, but don't expect the chef to explain in detail how he cooked it.


* 依田流アルファ碁研究 (よみがえる、呉清源·道策)ISBN 978-4-8399-6299-9.


This post by John Fairbairn was liked by 8 people: belikewater, Bonobo, Drew, gowan, jeromie, jptavan, Marcel Grünauer, Shenoute
Top
 Profile  
 
Offline
 Post subject: Re: Yoda's book on AlphaGo
Post #2 Posted: Thu Sep 07, 2017 3:48 am 
Dies with sente

Posts: 99
Liked others: 0
Was liked: 8
Thank you for this wonderful review. I honestly think the best commentaries about AG should come from Fan hui (some other pros of course) and deepmind. I'm actualmy reading invisible and I'm a little disappointed of this book. Not enough variations inside.
Im going to look this book (yoda's book) in my next travel, bur not gonna buy it.
Thank you ^^

Top
 Profile  
 
Offline
 Post subject: Re: Yoda's book on AlphaGo
Post #3 Posted: Thu Sep 07, 2017 8:00 am 
Lives in sente

Posts: 838
Location: Littleton, CO
Liked others: 284
Was liked: 262
Rank: KGS 3k
Universal go server handle: jeromie
Thanks, John. I'm getting a new set of go stones for my birthday in a couple weeks, and I plan to play through the games in your Kamakura book as soon as they're in my hands. It will be interesting to keep the AlphaGo games in mind as I do so! Since I don't have the ability to read Japanese I don't plan to get Yoda's book, but even being aware of the major themes will make my experience all the more enjoyable.

Top
 Profile  
 
Offline
 Post subject: Re: Yoda's book on AlphaGo
Post #4 Posted: Thu Sep 07, 2017 8:21 am 
Gosei

Posts: 1365
Liked others: 348
Was liked: 386
Rank: 5d
GD Posts: 1000
I pointed out some of these aspects of the Yoda book without as much detail in another thread. One thing I particularly like about the book is that it points out that even super strong players still make/made mistakes. In particular it is clear that AlphaGo/Master is far from playing perfect go, whatever that might mean. I think Yoda is strong enough to be able to criticize these great players, including AlphaGo/Master as a "player". It is easier to spot mistakes when one is in the peace and quiet of one's study than in the pressure cooker of playing a game.

Top
 Profile  
 
Offline
 Post subject: Re: Yoda's book on AlphaGo
Post #5 Posted: Thu Sep 07, 2017 9:22 am 
Tengen

Posts: 4784
Liked others: 0
Was liked: 648
AlphaGo is known to make score-optimisation mistakes of various kinds (wasting ko threats, suboptimal endgame, wrongly playing a ko etc.) when ahead because it does not make win-orientated mistakes. I have hardly looked into the self-play games yet; maybe they reveal win-orientated mistakes when behind? In current play against humans, we cannot really say that AlphaGo was behind so we can only try to find win-orientated mistakes when the win is still unclear. Some things have been pointed out as "mistakes" but, during the opening or middle game, we hardly can prove that they would be mistakes. From a human perspective, if relying on traditional, informal go theory, we would call some things "mistakes", such as early elimination of aji. However, there have not been proofs that would at least resemble formal (mathematical) go theory. AlphaGo likes simplification because it eases recognising larger win probability. We have to work much harder to prove win-orientated mistakes in the post Lee Sedol match era.

Top
 Profile  
 
Offline
 Post subject: Re: Yoda's book on AlphaGo
Post #6 Posted: Thu Sep 07, 2017 10:45 am 
Honinbo

Posts: 8591
Liked others: 2511
Was liked: 2962
RobertJasiek wrote:
AlphaGo is known to make score-optimisation mistakes of various kinds (wasting ko threats, suboptimal endgame, wrongly playing a ko etc.) when ahead because it does not make win-orientated mistakes.


Of course AlphaGo makes win-oriented mistakes. First, it makes evaluations based upon probabilities, which assume mistakes. (Perfect play will produce probabilities of 1 or 0.) Second, we cannot always tell its mistakes because 1) it is better than we are, and 2) its evaluations do not have publicly well defined meanings. To say that a certain play has a 60% probability of winning does not mean that if AlphaGo made that play and played the position out against itself 100 times, the player who made the play would win around 60 times, which is what the probability of winning ought to mean. We -- including its programmers -- do not know what the numbers mean, and it is not telling. There is nothing to argue about except numbers, and the ultimate defense: AlphaGo does not think like humans do, blah, blah, blah.

Quote:
I have hardly looked into the self-play games yet; maybe they reveal win-orientated mistakes when behind?


A look at the self-play games reveals that AlphaGo often ends play with a fairly large ko. Presumably one side or the other thought that it was behind and managed to produce the ko. Given that propensity, I find it hard to believe that squandering a ko threat improves the actual probability of winning, as a rule. (By squandering a ko threat I mean playing a sente before its normal time. Also, choosing between equally large local plays in a way that gives the opponent more ko threats or yourself fewer ko threats seems like it would, as a rule, reduce the probability of winning, regardless of how you calculate the probability.)

Quote:
AlphaGo likes simplification because it eases recognising larger win probability.


AlphaGo does not play for simplification in self-play, only against humans. My guess is that it tends to simplify when it thinks it is ahead, just like humans do. Anyway, versus humans its simplification does not seem to hurt.

_________________
There is one human race.
----------------------------------------------------

The Adkins Principle:

At some point, doesn't thinking have to go on?

— Winona Adkins

Top
 Profile  
 
Offline
 Post subject: Re: Yoda's book on AlphaGo
Post #7 Posted: Thu Sep 07, 2017 11:05 am 
Lives in sente
User avatar

Posts: 709
Liked others: 7
Was liked: 92
Rank: German 1 Kyu
Bill Spight wrote:
AlphaGo does not play for simplification in self-play, only against humans. My guess is that it tends to simplify when it thinks it is ahead, just like humans do. Anyway, versus humans its simplification does not seem to hurt.

I remember that Michael Redmond 9p often mentioned (may be supposed ?) in his commentaries on AlphaGo vs. human games that (he "feels" that) AlphaGo plays "very aggressive" / "unusual" moves, because she evaluated to be behind.

In AlphaGo vs. AlphaGo games, it is kind of natural that one of these will (most of the time) evaluate to be behind.
This might be the main reasoning for her moves that cannot be easily understood by humans.

_________________
The really most difficult Go problem ever: http://igohatsuyoron120.de/index.htm
Igo Hatsuyoron #120 (still unresolved by professionals, maybe solved by three amateurs)

Top
 Profile  
 
Offline
 Post subject: Re: Yoda's book on AlphaGo
Post #8 Posted: Thu Sep 07, 2017 12:21 pm 
Dies in gote

Posts: 29
Liked others: 7
Was liked: 14
Quote:
AlphaGo does not play for simplification in self-play, only against humans. My guess is that it tends to simplify when it thinks it is ahead, just like humans do. Anyway, versus humans its simplification does not seem to hurt.


Michael Redmond said something similar to what Bill is saying in the first video where he reviews AG vs AG self-play. He said AG seems to like complicated middle games now. He speculated that it had to do with Fan Hui feeding it complicated middle game problems during its training after the melt down in game 4 against Lee Sedol.

Top
 Profile  
 
Offline
 Post subject: Re: Yoda's book on AlphaGo
Post #9 Posted: Thu Sep 07, 2017 6:02 pm 
Lives in gote
User avatar

Posts: 392
Liked others: 23
Was liked: 43
Rank: NR
From John's description of it, Yoda's book sounds like a pretty decent attempt to make sense of Alfie, which is a pretty tall order, because she doesn't have any sense in the ordinary sense of the word.

But she is a phenomenon worthy of study from a Go perspective as well as a stimulus-response learning one, because it is quite likely that she does have something to teach pros, if for no other reason than that she beats them.

From the human perspective, Go is a game of global strategy (based on a structured mental image) supported by local tactics (reading) of up to around 20 moves, but to Alfie it is a game of local strategy (policy net) supported by global statistics (Monte-Carlo with bells on) all the way to game end. It is hard to imagine two approaches that could be further apart.

Yoda mentions a few of Alfie's idiosyncratic superficial local patterns that have been generally recognised (shoulder hit + tenuki, 4th line nikken tobi, etc), plus a fondness for centre control (something she shares with Zen) and it may be that these memes will be added to the set of human heuristics (i've noticed many of my opps using them, albeit in a context-free way!).

To me, Alfie's success is testament that Go isn't quite the game it used to be thought to be; rather than a metaphor for life philosophy, it's more like a Rubick's Cube puzzle, more susceptible to autistic savant brute-force "end-to-end" combinatorial search than judicious empathetic semantic reasoning. As hardware technology becomes more affordable, we can expect the modus operandi gap between human Go and computer Go to widen, rather than coalesce. On the other hand, Alfie and her successors surely have something to contribute to the science of Go, even if they have little to contribute to the science of understanding.

... and on the third hand, there may be a third way, a "middle way" to swim with both reasoning and search... there is some indication that i am not alone in thinking of this possibility

_________________
i shrink, therefore i swarm

Top
 Profile  
 
Offline
 Post subject: Re: Yoda's book on AlphaGo
Post #10 Posted: Fri Sep 08, 2017 2:42 am 
Lives in sente
User avatar

Posts: 1306
Location: Ghent, Belgium
Liked others: 184
Was liked: 582
Rank: Bel 2d KGS 3d TG 4d
KGS: Artevelde
Tygem: Knotwilg
Even cosmology is eventually susceptible to brute force explanation and not a philosophical effort per se. The more we can compute the less we have to mystify.

So your revelation about Go being more like a Rubic's cube is not a revelation to me. It's a logical consequence of computation becoming more powerful.

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 10 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: No registered users and 0 guests


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