Leela Zero analysis of 'Making good shape' and other books.

zermelo · **#61**

Here's one more book example, this time from Ishida-Davies 'Attack and Defense', a book that I appreciate very much.

There is a problem quite early in the book, asking which one is better for black, A or B:

Leela Zero thinks that the book right answer is 4.5% worse than book wrong answer, and a 7.5% mistake (also compared to the best local move). You can navigate the above diagram for some notes and variations.

I am not taking LZ opinion as gospel, by the way. Indeed, maybe alphago or the next year's LZ would have another verdict. However I have no reason to think that Ishida-Davies know better. LZ strength is evidence, and I think the only evidence, that suggests it may know better.

What I did learn is that no amateur has anything to worry if he has no idea which of the moves is better. The book would be a bit better if it just said that some options in this kind of running fight are A and B, instead of trying to judge them.

(One caveat is the komi, which was I think 5.5 when the book was written. But if the difference between A and B is so subtle that their ordering depends on the komi, then it is certainly high above the pay grade of the book's audience.)

Kirby · **#62**

Quote:

Ponnuki without capturing is impossible. It's what the word means.

Thanks. I used to study Japanese, but funnily enough, I didn't know this. Kind of makes sense.

Bill Spight · **#63**

Tryss wrote:

Quote:

The position doesn't even have to be all that wacky. IIUC, the more any bot is trained by millions of games of self play, the more it has been trained on high level positions played in its style. How many of those positions are going to be similar to those that I face at move 75? The stronger the bot gets, the less similar its training positions will be to my games, n'est-çe pas?

But these deep neural networks are good at this kind of generalization. Even if you create an unnatural position (but equal and not a 120 moves deep tsumego), it will still play sensible moves.

Yes, but so will I.

OC, the nets are good at generalization, but as the bulk of the positions that they generalize becomes narrower (less varied), their generalization won't get better, will it?

Bill Spight · **#64**

zermelo wrote:

Here's one more book example, this time from Ishida-Davies 'Attack and Defense', a book that I appreciate very much.

There is a problem quite early in the book, asking which one is better for black, A or B:

Using my "miai method" of comparing plays, I compared these two diagrams.

Black A - White B

Click Here To Show Diagram Code: [go]$$Bc Black A - White B $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . O . . . X X 1 . O X . . . | $$ | . . . O . . . . . , X O . . O , X . . | $$ | . . O . . . . . . . X O . 3 . . . . . | $$ | . . . . . . . O . . . O . . . . X . . | $$ | . . . . . . . . . X . . . . . . . . . | $$ | . . . . . . . . . . . O . . . . . . . | $$ | . . X . X . . . . X . . . . . . . . . | $$ | . . . , . . . . . , . 2 . . . , . . . | $$ | . . O . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . O . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . X , . . . . . , . . . . . O . . . | $$ | . . . . X . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Edit:

added later.

Black B - White A

Click Here To Show Diagram Code: [go]$$Bc Black B - White A $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . O . . . X X 2 . O X . . . | $$ | . . . O . . . . . , X O . . O , X . . | $$ | . . O . . . . . . . X O . . . . . . . | $$ | . . . . . . . O . . . O . . . . X . . | $$ | . . . . . . . . . X . . . . . . . . . | $$ | . . . . . . . . . . . O . . . . . . . | $$ | . . X . X . . . . X . . . . . . . . . | $$ | . . . , . . . . . , . 1 . . . , . . . | $$ | . . O . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . O . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . X , . . . . . , . . . . . O . . . | $$ | . . . . X . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Despite a certain thinness with the keima, I prefer the second diagram.

Edit: Hmmm. The existence of a play like :b3:

in the first diagram suggests that this is not a good example for my miai method.

Out of courtesy I am hiding my further remarks.

John Fairbairn · **#65**

The following is part I was transcribing a game this morning. My first impression was that I must have made a mistake somewhere along the line as the moves looked so bad. On one of the previous forums, T Mark and I used to run a series where you had to guess whether the game was by pros or amateurs. This one I would have ranked at kyu level. The shapes (at first) struck me as horrendous, although I realised later it was the suji (the flow of stones) that felt most wrong. Looking at the resulting actual shapes (katachi), yes there were empty triangles etc, but on the whole you could say Black's bad shape cancelled out White's bad shape.

But even then, something felt off, and my next thought was that the two top young Japanese players here had been at the forefront of training with DeepZen - perhaps they were trying to play like an AI bot?

It's a weird game worth looking at anyway, but I took a look with LeelaZero. I was taken aback by the results. There were no 3-3 invasions or shoulder hits but almost every move was in a margin-of-error range that I presume was very acceptable. What I mean by that is the winrate changed by a very modest amount, usually down though not always, within less than a percentage point in most cases. The only moves that showed a big change in the early stages were 49 and 56 (-10% and -11%). These two big mistakes cancelled each other out but the trend was for Black to accumulate more tiny mistakes and he eventually lost.

The impression I have been getting up to now, perhaps wrongly but mainly from myself and others pointing up apparent pro mistakes in problem books, is that pros have been making lots more big mistakes than in this weird game, and we should be more prepared to question their judgement.

So I took a look at some other recent (i.e. AI-influenced) games, and it seems again that pros are getting winrate changes that are being held within a very low range. I certainly did not get that sort of pattern when I looked at some famous Edo period games before.

At best, what I think I have found can only be indicative, but I'm curious to know whether other people have an impression that modern pros have already been able to adopt some AI knowledge/feeling (and, as I've said, outside the usual 3-3/shoulder hit novelty paradigm)?

Bill Spight · **#66**

Yes, I think that you can see the imitation of the bots in this game. Imitation means learning by observation. There is surface, monkey-see-monkey-do imitation, but there is also a deeper, more thoughtful imitation, called programmatic imitation. I think we can see both in this game.

It would be interesting to see a professional commentary.

I have made a few comments in the sgf file, in an attempt to understand some of what is going on.

Gomoto · **#67**

You can make one eye at the edge? Dont! Run!

Leela Zero analysis of 'Making good shape' and other books.

Who is online