Hoen Shinpo--Are historical handicap openings still good?

[John Tilley]

You are correct about the title of the first Fukui book - the hardback one. The main difference between the two books by Fukui is that the first one I mentioned has lengthy analysis and comments on the games - the first game has 15 pages and is 13 figures and 23 diagrams long.

The second book has about 5 figures and 5 pages per game and no explanatory diagrams.

Fukui wrote a series of seven volumes on Classical Japanese Go - Jowa, Dosaku, Great Senchi, Shuwa, Shusaku, Shuho and Shuei whereas the Nihon Kiin series has eight volumes - you gain Genan Inseki and Chitoku, but lose Senchi.

I haven't attempted to list and compare the chosen games between the volumes (yet).... Its always interesting as to the games chosen for these anthologies.

Take care - John

[John Fairbairn]

There are a few example pages to read on Amazon, including mention of the ear-reddening move, which I was also discussed here on L19 using AI. I do not read Japanese. I only know enough to read tsumego problems and whether there is a ko. This book seems to have a lot of expanded discussion without saying much. The amount of text is too much for me while translating. I'd prefer Shuho's brevity. And I am not as interested in the AI analysis. I appreciate Marcel's suggestion though. If anyone orders the book, please share a review of it.

I haven't got round to reading it properly yet, and I'm not sure when I will - all the numbers and graphs represent trials and tribulations for me and I haven't got a magic flute to help me with the task of overcoming them.

On the plus side, part author Ohashi Hirofumi is an engaging Papageno-like figure, and certainly knows the AI field. He is the go-to man from Japan whenever AI seminars and tournaments send out invitations (via Hedwig, no doubt ). He wrote a "Go AI Encyclopaedia - from AlphaGo to Zen" (the wrong chronology but A to Z had too much pull for him or his editor). Frankly this was a bit of a jumping-on-the-bandwagon pot-boiler, certainly by his standards. A very much more interesting book on "New Fusekis in the AI age" followed. Rather than trying to tell us what the new principles of fuseki should be, he looked instead at unusual fusekis such as o-takamoku (6-4), the tiger-jimari, 5-5, tengen and the Black Hole, and tells us what AI says about these openings. Not being a proper student of the game any more I never read all of it, but I read enough to be convinced it was at the least a very entertaining book. As I say, his whole tone is "Der Fusekifänger bin ich ja, stets lustig, heissa hopsa-sa!" Whether it is also useful I don't feel qualified to judge.

The latest book, invoking Shuwa and Shusaku, is one that I think I will get round to reading (the history pull is too strong) and I'm already certain it will be very useful. But this is not quite Ohashi's usual hopsa-sa style, no doubt because the book is almost entirely a conversation between him and co-author Terayama Rei, who is still only 20 (but already 6-dan). This means there is no solid presentation of theory or other ideas. You have to glean these from throw-away comments in the conversation. But there's gold in them thar comments (e.g. Ohashi's observation that AI prefers high extensions to low ones).

Actually, despite the title and the presence of graphs, in many respects the book is not even about AI. The book has a sub-title that says you can become strong just by playing over old games. That is really the main message (and the stress on improving by playing over pro games is one that I've seen repeated over and over in recent months). AI is just a tool to help you do that, by enabling you to look deeper and more accurately. Having said that, there is a human endgame brilliancy here that Katago 6809 (the version used) allegedly missed, despite searching for over a million moves.

I don't think it's a spoiler to say the humans come out rather well in this book. It wouldn't have been worth writing otherwise. But even that's not the real point, which is still that the way to improve is to play over these games with the latest tools at your disposal. You still have to think for yourself.

The structure of the book:

Introductory Section: The Ear-reddening game
Section 1: Shuwa
Section 2: Shusaku
Section 3: Shuwa vs Shusaku.

Dotted throughout the book (i.e. between the games, which do not have conventional commentaries) are also what would normally be mini-essays, but these too are in conversational format. They cover on the one hand things like how to use AI to study old games, AI and thickness; but on the other they talk about Edo go history.

Because of the conversational nature of the book, I think you will need to be able to read Japanese beyond the usual go-words level. And there are no variation diagrams to provide an alternative to what the text says. You may get something out of the graphs, but actually there aren't as many as I may have implied (one is too many for me), and you could run your own AI to generate pretty similar graphs yourself, I imagine.

But if you can read the text, even with my cursory engagement with it I'd recommend the book very highly, both as entertainment and as a way to improve.

[CDavis7M]

There is quite a lot of detail of Shuwa's life in my Meijin of Meijins.

Well, that was depressing.

My hype went up in flames, so to speak.

[John Fairbairn]

My hype went up in flames, so to speak.

Worse for him, though - his godown went down in flames.

I have a question for the AI buffs. Having posted the (semi-)review above, I bestirred myself enough to read some more of the book in bed. It was numerology stuff, and so more trials and tribulations, but a pre-bed toot on a magic harmonica instead of a flute gave me some inspiration.

Ohashi (I assume it is he - he just says "I") points out that the choice of whether to look at win rate or points difference (in Katago) depends on the state of the game. Points difference should normally be used as this is more familiar to a human, but AI is unreliable as regards points difference in intense fighting situations, such as capturing races and ko fights (and even may make inaccurate moves). Then you should rely on the win rate. I can follow that.

There is, however, I would suggest, a powerful psychological reason for preferring points difference, which he does not discuss directly. Herein lies my question. He shows graphs with both win rate and points difference, the lines superimposed. The win rate line often veers up and down wildly with huge fluctuations, making it look as if one side is far ahead it's not worth playing on, or that a player has made a truly horrendous mistake. Yet the points difference line typically trundles on, on an almost flat path, telling us that the game is really rather close, and any points change is in a range a typical amateur wouldn't even notice..

Ohashi says that, by the endgame, in modern komi go, if a player is leading by half a point, that may represent about 90% win rate. OK, but then he says, "If he loses 1 point, this will drop 80% and may become 10%." Forget the mix-up between % and percentage points - Japanese are as prone to grocer's mathematics as we are. But in my grocer's head, I say "losing a point in a half-point game means the lead switches" so I expect the win rate now to be 90% for the other guy. I assume Ohashi may have been a bit sloppy and omitted some detail about the 1-point mistake (sente/gote?), but I also assume that I have probably missed something. What?

There are a couple more statements by him that have not yet passed the Pearly Gates test for entry into my brain. One is: "In the case of no-komi go [where jigo is possible], if you make a 1-point loss from a win rate of 90%, you go down to 50%, or 0 points. To a mind used to counting as in modern go, when you look at the count in a no-komi game, don't forget that if it is a close game the amplitude of the swing is about half."

I can sense some sense in that, but basically it's smoke and mirrors to me. Can someone provide the baby steps to understanding?

A further remark: In the case of 1-point [no-komi] game, we can perhaps say that with a win rate between 25% and 75% there is a high probability that it will be a jigo." Again, I can feel an inkling of understanding, but a voice in my head says, "So what can we say about the 50 percentage points in that range - are they just statistical noise?"

Here are a couple of graphs. The thick line denotes win rate and the thin line is points difference. The middle dotted line is the 50% bar.

In the first case, we can see the possible psychological impact if we rely on win rate rather than points difference.

Ohashi1.jpg (70.21 KiB) Viewed 63789 times

In the case below, we can see that Black (Shuwa) apparently made not a single significant mistake. There are similar examples in other games, which is why I say the humans seem to come out of this book rather well.

Ohashi2.jpg (50.54 KiB) Viewed 63789 times

But these graphs also reveal what I think is a defect of the book. There are no scale markings, and so it seems that in one of the above (non-handicap) games, Black started off with a bigger win rate than in the other. There are also no move numbers, and while it seems likely that the whole game is represented, we cannot be certain in some portions of the book. FWIW, in the case of the "complete victory" by Shuwa (then "at his peak" say the authors), the margin of victory was apparently 5 points, but the graph line shows a gradual increase, which implies the initial advantage to Black was less than 5 points????

Note: I can't find a way to turn the pictures round. They appear in the right orientation elsewhere on my system.

[kvasir]

Ohashi says that, by the endgame, in modern komi go, if a player is leading by half a point, that may represent about 90% win rate. OK, but then he says, "If he loses 1 point, this will drop 80% and may become 10%." Forget the mix-up between % and percentage points - Japanese are as prone to grocer's mathematics as we are. But in my grocer's head, I say "losing a point in a half-point game means the lead switches" so I expect the win rate now to be 90% for the other guy. I assume Ohashi may have been a bit sloppy and omitted some detail about the 1-point mistake (sente/gote?), but I also assume that I have probably missed something. What?

I think what you are missing, actually you don't seem to be missing anything because it is a case that you are adding sente / gote considerations were there is none. The predicted points and point differences are estimates of the final result of the game. If you had a choice of gote moves x and y, where x > y, but choice y your point loss is approximately y - x. If you instead make a mistake of sente or gote, then the estimate is still of the final result of the game, but one one component of the point difference will be an estimate of the value of sente. Basically, you can speak of 1 point mistake without specifying sente or gote, it could have been that you played a move in sente that is just 1 point less than it could have been or you could have played a really big move in gote and this 1 point is the difference compared to playing some sente moves first.


The "point difference" and "percentages" are both estimates that are
1. affected by uncertainty
2. are mapping win-loss (or 0-1 result of the game) in two different ways that introduces ambiguity of presentation

The uncertainty is less toward the end of the game and in quiet positions, it also depends on the quality of the analysis. The percentages should get closer and closer to 0% or 100% as the game moves closer to the end, even if the game is quite close.

The ambiguity of mapping a win-loss result to a point difference can lead to values that are difficult to interpret. For example the swing between who wins a capturing race can be heavily biased in terms of points, as was mentioned by Ohashi, basically what does it really mean to say that the average of x and y is (x + y) / 2, and that this is the outcome of the game, if that is not a likely outcome of the game and only x or y are plausible outcomes. This is just a simple example to show that the point estimate can be almost meaningless.

We probably shouldn't discuss "percentages" or "point differences" very much when reviewing games. I sometimes feel one can just as well say "I checked it with a computer" or "computer approved" than say "a is 16.4% better than b" or "a is 3.4 pts better than b".