Review: Invisible - The Games of AlphaGo

GENERAL SPECIFICATION

* Title: Invisible - The Games of AlphaGo
* Author: Antti Törmänen
* Publisher: Hebsacker Verlag
* Edition: 2017
* Language: English
* Price: EUR 34
* Contents: games
* ISBN: 978-3-937499-06-2
* Printing: almost good
* Layout: intermediate
* Editing: almost good
* Pages: 272
* Size: 176mm x 250mm
* Diagrams per Page on Average: 3
* Method of Teaching: examples
* Read when EGF: 10k - 7d
* Subjective Rank Improvement: -
* Subjective Topic Coverage: o
* Subjective Aims' Achievement: ++

**************************************************************************

INTRODUCTION

Invisible contains games of AlphaGo - 73 commented games (5 of the Lee Sedol match, 3 training games, 60 online games against human opponents, 5 of the Ke Jie event) and 5 uncommented games (against Fan Hui). Recent self-play games of AlphaGo did not make it into the book.

PRINTING, LAYOUT, EDITING

The digital printing of text and diagrams is very good, and the paper pleasant for reading. The binding is bit weak: after one reading, it begins to divide the pages into two blocks. The cover paper is too thin for the large size of the book, but the real problem is the coating, which comes unstuck on most edges, looking ugly. It would have been better to omit the coating with its too weak fixation or use a better coating. Blue is a risky colour for ordinary coating.

The layout is designed to be visually pleasing. This is achieved by generous paragraph spaces, bold font starting most parapgraphs with for example its first move number, a generous font size and layout lines accompanying headings. There is, however, a compromise between sufficient size of the game diagrams and a slightly narrow space between the two columns per page. The large book format is not used for more contents but enables the generous layout.

However, another aspect of the layout is weak: half of the commented games conclude with so much white space that there could have been 1, 2 or 3 additional game sequence diagrams per game. This is very unfortunate because it makes following the game sequence hard especially during the endgame. Just three hours of additional editing would have avoided this problem.

The editing has eliminated most accidental mistakes and just a few remain, as expected for every book. On the other hand, the proofreaders should have identified the circa three go terms, which are partly consistently, partly occasionally used instead of more appropriate terms. The consistent mistake is 'area counting' used with the meaning of 'area scoring'. These errors do not inhibit understanding of the contents though.

It may be amusing that a fight is often called a melee. I am more concerned about the frequent remarks that AlphaGo would have played something 'novel'. This is correct only half of the times. For the other half, Antti Törmänen appears not to have been aware of earlier occurrences in human play.

THE COMMENTARY

Studying AlphaGo's games from a book can be sufficient reason to buy it. However, the book also contains commentary so the question arises whether it creates additional value. By far most games are commented by Antti Törmänen. Some other games are commented by a few other professional players. Every game sequence diagram shows circa 50 moves.

The three major parts of the book have entertaining introductions with interesting background information. A few of the game commentaries contain sound speculation about how the AI might think, but we cannot know for sure and one can come up with alternative explanations. Nevertheless, these parts generate some enjoyable reading.

73 commented games etc. on 272 pages - you can easily do the maths: we have from 2 to 4, infrequently a few more pages of commentary per game. The emphasis has been put on including as many games rather than providing as detailed commentary as possible. Accordingly, there are also only a few (for example, 4) additional commentary diagrams per game.

The text contains relevant commentary for intermediate to stronger players (often for circa 5 kyu to 3 dan), elementary commentary useful for 7 to 10 kyu (such as pointing out moves building influence) as well as trivial comments and references (such as "the move starts the endgame" or "also appears in game x").

With few commenting diagrams and little space available or used for relevant commentary, necessarily the commentary has to be highly selective. A few aspects of the game are commented but most aspects and moves are left uncommented. That said, the diagrams and relevant commentary are chosen well for the purpose of creating the reader's illusion of understanding what was going on in every game. The diligent student of AlphaGo's play will be, however, disappointed and is left to his own study of its strategy and tactics.

As far as the relevant commentary is offered, it almost always shows correct moves and comments. Antti Törmänen shows his principle ability to teach. I found only a few suboptimal moves and pieces of advice throughout the book.

With the exception of positional judgement. AlphaGo's strategy heavily relies on accurate positional judgement. Most of its judgements are not discussed in the book. Most of the discussed positional judgement is weak: inaccurate (incorrect, rough statements or rounding to multiples of 10), partial (e.g., only territorial), speculative (reflecting what AlphaGo might have thought instead of analysing the position). Only some of the endgame or territory judgements of the other professional commentators are reasonable.

Antti Törmänen writes to have improved significantly from studying AlphaGo's games. Unfortunately, he does not convey related insight to the reader. The game commentaries read like your ordinary GoWorld commentary. Apart from a few special insights, such as those related to AlphaGo's 3-3 invasions, nothing points to a broad knowledge for improving one's playing strength due to understanding AlphaGo's play.

AlphaGo's games could have been analysed deeply and systematically for its strategies and tactics. With a few exceptions mostly related to its opening and joseki preferences, the book does not do it.

Did you expect this new book on AlphaGo's games to summarise, comment on and go beyond earlier books, video commentaries and online commentaries? Although some of the variations are chosen particularly well, you will be disappointed. The format of this book does not offer the space for an exhaustive analysis.

CONCLUSION

The price reflects the large format. Buy the book if you want to study AlphaGo's games in printing or the rough guidance through the games already makes you happy. Although the commentary is mostly correct (except for weak positional judgement), it adds less additional value than the short historical background information. The commentary might be good enough if you expect no more than GoWorld style. It is disappointing if you expect the promised significant improvement of playing strength or deep, systematic insight in AlphaGo's play.

Books based on the AlphaGo (and Master) games are starting to appear. This one is the only one in English so far, I guess. One Japanese book related to AlphaGo's play just appeared, by Yoda Norimoto. That book is mostly devoted to discussing how some of the typical AlphaGo moves and tactics appeared in historical players' games, e.g. Dosaku and Go Seigen. It seems to be the case that AlphaGo/Master has been retired so the games we have are all we are going to get. We are not going to be able to "ask" AlphaGo about anything. In view of this we have to be grateful for any accessible commented games of AlphaGo and look forward to how the top players make use of the innovations introduced by AlphaGo which may be explained in more detail in the future.

I think it has been pointed out (by John Fairbairn?) that, thanks to AlphaGo, we are in a time like that of the "new" fuseki in Japan in the 1930's. The innovations back then had a lasting effect on professional go but at the time there were many experiments that were later found to be unsound. In his book Yoda comments on eight games involving AlphaGo or Master. I wonder whether human opponents playing Master might improve their results once intensive study of these innovations and experiments with these moves have been done. As we know, even pros can be confounded when faced by unexpected moves or patterns. This also happened back in the '30s, too.

AlphaGo vs Lee Sedol: The Match that Changed the World of Go
https://www.amazon.com/AlphaGo-Lee-Sedo ... 542800714/

AlphaGo vs. Ke Jie 9P
https://www.amazon.com/AlphaGo-vs-Ke-Ji ... 974674584/

_________________
http://tchan001.wordpress.com
A blog on Asian go books, go sightings, and interesting tidbits
Go is such a beautiful game.

Thank you for your review, Robert.

It is inevitable that this book be compared to Invincible.

But I do not think we will ever see such a comprehensive analysis of AlphaGo's games in book form. The reason for that is how much the world has changed since Invincible was produced.

John Power's book was first published in 1982! Pre-YouTube, Pre-SGF, Pre-Lifein19x19, and basically pre-Internet as we know it today. Back then an exhaustive book was the only effective way to convey this sort of information.

Now, of course, things are quite different. Whether other presentations are better for delivering go analysis is not for me to judge, but I believe it absolutely has weakened the demand for and potential impact of an "Invincible 2" type work.

The initial printing of Invincible used hard cover. Nowadays, it is hard to sell English hard cover go books because few want to pay the extra price for extra production quality.

Invincible puts a bit more emphasis on commentary per commented game on average than Invisible, I think. Neither goes in a deep and systematic analysis of the discussed player's playing style with an approach of relating strategy and tactics of all games. Instead, both books are similar in including as many games as possible.

IMO, there is demand for deep and systematic analysis in book form and potential for much more thorough discussion than in videos or online commentaries. Naturally, I have considered writing on AlphaGo but have been more than busy with other projects. It would not be the type of book that could be written quickly but would require much time for careful study. Compare Relentless, which took its time.

Surely, AlphaGo deserves much more study than so far.

Is not AlphaGo already the most studied player in history?

What is needed is for more exceptional players who can write to share their insights with Go Neanderthals like me

Nowhere near, I'd say. And if we look at chess, we don't see chess pros systematically studying games by much superior chess programs (or at least publishing books of computer games).

I think we can allow ourselves to be cynical here. A very large part of the so-called "study" of AlphaGo is professionals jumping on the bandwagon to make a quick buck, or lesser players to get a reputation. Of course there's genuine fascination with AlphaGo in the mix, but even top pros are not really capable of getting or giving any insight into AlphaGo, simply because it "thinks" in a completely different way from humans. We already "understand" how it uses Monte Carlo algorithms, for example, but that understanding is completely useless to us in our own games.

So far the so-called study of AI is just assuming that AlphaGo is better than humans at fuseki. We don't even know whether that's true. Its real strength may lie later in the game, but either way we can assume that machines will become even stronger in all aspects, by using machine-like thinking, and so will become even more distant from our way of thinking. The pace of progress will probably continue to be astounding. Two months ago we saw DeepZen described as a better model for humans because it makes more human-like moves than AlphGo. This month, in Go World, Ohashi Hirofumi says DeepZen has already moved to a new level of making "creative" moves like AlphaGo that a human couldn't make. In the same issue, O Meien says AI is already permanently beyond the reach of humans in go.

Whether you think this spoils go is a personal matter. I get the impression that a lot of established go players have been depressed by AlphaGo (despite a little early fascination) and some have already given up. But if we look at chess, we can say that while AI will change the game in many negative ways (e.g. cheating and creating disdain for players of the past whose wondrous plays will be shown to be flawed), it can also have a positive effect - chess seems to be played by many, many more serious players than ever before and more pros seem to make a good living nowadays. I suspect the real difference between go and chess that is preventing many people from absorbing the lessons of chess is simply the sheer pace with which AlphaGo came up on the rails.
Chess players had years and years to get used to the robots. They have even found time to reduce them to an aspect of chess entertainment! (Humans 1, Computers 0)

We need to take a deep breath and accept that not only will none of us truly understand AlphaGo, 99.999% of us don't even understand yet how a human pro thinks and no-one has found any insights, bar one, which lead us anywhere near go's Everest. The one exception is, of course, that you have to work very, very, very hard.

But for those of us who don't want to work that hard and just want to be entertained as fans, AlphaGo has already opened the door, and not just to pot-boiler books. For me and some others here it has cast Go Seigen in a new (and even better) light. It's easy to fall into the trap of being overawed at what machines can do. Yet it's also possible, and more satisfying, to look in amazement at what humans can do.

Is Invincible really that in-depth analysis? It's been a while since I read it (and only parts of, not cover-to-cover), but class it more as a generous medium-length analysis; Robert mentioned Relentless from GoGameGuru on the Lee Sedol vs Gu Li match which I think is more thorough (both deeper and broader, e.g. going off on tangents of related fuseki theory with many diagrams) but is also more spaced-out and generous with diagrams so its higher pages per game could exaggerate the difference. Lee Sedol's book on his games is even more detailed (the most I've seen in English books) with about 100 pages per game. Redmond's 1 hour plus video reviews of the AlphaGo self-play games are probably of a similar detail to Lee's book. I very much enjoy these (and joined the AGA to support them) and like the format. I could also be interested in a book but is it really practical? All 50 games would probably need a book a foot or two thick (so multiple volumes) and would be a monumental writing/editing/typesetting effort (video is easier to produce). I suppose one key point is the strength and quality of the analyser: I like Lee Sedol's book because the Go content clearly comes from the super strong player on the cover (unlike some other ghost written books). A book based on AlphaGo analysis from the top Chinese pros (e.g. via WeiqiTV commentaries) would be much more attractive to me than from Robert Jasiek. Redmond is not a top pro but I think does a good job of analysis and is a great communicator; I wonder how his analysis differs from others, e.g. does he make outright mistakes or just miss deeper things: it will be interesting to compare his analysis of the WeiqiTV games with theirs with the Chinese pros and AlphaGo via Fan Hui. He does admit reasonably often when he doesn't understand something or can only say "if this other choice happens we get a trade which probably ends in a really close game 150 moves later and maybe a different colour wins". Perhaps some others would make a decent stab of trying to answer that question, but given how much stronger AlphaGo is than humans I think it'll be hard to trust a conclusion that AlphaGo Black should have played a different way and would have then won not lost by half a point 200 moves later.

If this book is similar to Invicible : the games of Shusaku, I would definitely not recommand it, it would be a waste of time.

Also I think this book won't bring more much than "Redmond's Reviews" on YouTube about AlphaGo which is a better format imo.

Which book is this?

That's because you have not seen some such game commentary book from me yet.

https://senseis.xmp.net/?CommentedGamesByLeeSedol. I've got just volume 1. Looks like the German but not English version is discounted at https://www.schaakengo.nl/goshop-keima/ ... -volume-1/. There's a sample pdf on gogameguru's shop but they don't sell anymore: https://shop.gogameguru.com/commented-g ... e-sedol-1/.

Recently Nie Weiping said that AlphaGo is like Feng Qing Yang while Zen is Mei Chaofeng but I don't understand what he means though.

A quick Google says they are characters in martial arts novels by bestselling Chinese author Jin Yong, so presumably some kind of wise old master and apprentice allusion.

Whilst I do prefer to judge go commentary/books/analysis by their content rather than appeal to authority of the author, at certain points I think it does come down to that: notably with things like positional judgement or how confident one can be that the author hasn't missed better alternative moves (that I as a weak reader won't see either). And given your level I struggle to see how you can confidently analyse such high-level games, but maybe I'd be pleasantly surprised. Some things though I would expect you to do better than some others, such as thoroughly finding previous occurrences of apparently novel AlphaGo plays (John says Yoda's book is doing this back to Dosaku/Go, Michael Redmond has made a similar comparison): it does irk me a bit when they say "Wow, look at this new move no human plays!" when I, following along with GoGoD+Kombilo or ps.waltheri, can find it in a few seconds.

Uberdude, for positional judgement, I have described the necessary theory, and done so in greater, and sometimes much greater, detail than I have seen elsewhere. I am working on more theory that will come in suitably for game commentaries. In fact, it is a reason why I have not written some yet: first I want to build the foundation.

Professionals have some great advantage: their reading is much faster. Since I do not compete on reading speed, I compensate that by greater amount of time spent on reading before I comment on reading-dependent aspects, such as aji or life status. You can verify the quality of my reading in writing by possibly spending even more time (up to reasonably exhaustively checking all tactical alternatives) or, where applicable, verify whether I have applied my own theory correctly and understand why (in other cases: to which extent) the theory itself is correct.

Hint: read Positional Judgement 1 - Territory, Positional Judgement 2 - Dynamics, Tactical Reading (also do not overlook that the problems are created by myself), Fighting Fundamentals. You might read further books to better appreciate that my variations consider every move, regardless how unlikely its shape might appear.

This wikipedia page will tell you a lot about Feng Qing Yang: https://en.wikipedia.org/wiki/List_of_T ... characters

Have you seen the Hon Dojo's review of Master's games? It's interesting, since they go over the game quickly, break out the interesting points, and then work backwards to where they think either Master played an excellent move or his opponent played a blunder, and then a final conclusion.

The problem here being that these were the quick games played against strong players, but the format and review by multiple pros was an interesting style.

Sorry, brain fart. I'm aware of those 3 books but never saw which games he actually discusses there-in.

Commented Games of Lee Sedol Volume 1 in English seems to be out of print.

_________________
http://tchan001.wordpress.com
A blog on Asian go books, go sightings, and interesting tidbits
Go is such a beautiful game.