It is currently Sun Jul 05, 2020 11:54 am

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 8 posts ] 
Author Message
Offline
 Post subject: Reviewing book positions with AI
Post #1 Posted: Tue Jun 16, 2020 10:37 am 
Lives in gote

Posts: 648
Liked others: 281
Was liked: 323
There are several books whose value in the AI age I'm questioning:

  • The Direction of Play (Kajiwara)
  • Whole Board Thinking in Joseki (Yang Yilun)
  • A Way of Play for the 21st Century (Go Seigen)
  • Improve Your Intuition (Takagawa, 3 booklets)
  • Get Strong at the Opening
  • A Dictionary of Modern Fuseki, The Korean Style

I intend to take some positions from some books and check them with my AI configuration of choice: KataGo 20-block (g170-b20c256x2-s4384473088-d968438914.bin.gz), using around 10k playouts, often more. I believe that stating these parameters is necessary. Too often it is stated that "AI would play like this" as though those were absolute sequences.

I also invite others to share their analysis of book positions in this thread.

Often the AI wouldn't play the moves described as "correct" in the books and likes the "mistakes" better. These differences might be explained by some of the following:

  • playing for the side when the corner is open
  • having a different opinion about how strong or weak a group is
  • having a different opinion about how secure a moyo is or what aji there is inside it
  • having a different opinion about defect-free or useful a wall is

The books show some sequences; the AI shows some sequences. Who to believe?

Maybe the book moves are easier to understand for kyu and low dan players - the books' intended audience - but if I have a question, I can't ask the author. I can ask the AI - about any move, any sequence, any position, to any depth, for as long as I want.

Books show a very small subset of what is possible, and the AI doesn't even agree with the conclusions.

Related threads:


---

The first position is from "The Direction of Play". The Japanese original version was written in 1970, at a time when the komi was 5.5, so that is what I am using in Sabaki. With this komi, White starts out with a win rate of 44%, just for reference.

This position is shown on page 21, and the book says:

Quote:
If White encloses the [lower right] corner, Black extends to [the star point on the right side] and the game is over. Obviously that's an exaggeration, but White's forces are overconcentrated on one side of the board, whereas Black has taken control of the fourth line in a big way. This gives him such outside influence that it is hard to see how White's profit can match. If, after this, Black gets the sanrensei position with [the star point at the top], the game really will be over.


KataGo says that White can get 45.2% by invading the upper right 3-3, so it's a slight gain over the starting position. This is in stark contrast to the strongly-worded text.

Attachment:
direction-of-play-p021.png
direction-of-play-p021.png [ 877.56 KiB | Viewed 462 times ]


In addition I had White invade both 3-3 points and play the outdated hane-connect, further increasing Black's influence. But far from the game being over, White has actually gained 10% over the previous position.

Attachment:
direction-of-play-p021a.png
direction-of-play-p021a.png [ 910.97 KiB | Viewed 462 times ]


I suppose the author wanted to stress that it is good if stones are working together, but this seems to be a poor example.

There are some fundamental concepts mentioned in the book that I completely agree with:

Quote:
[...] it is important not to play [a move] simply because it *is* joseki. Rather, your move should take into account the overall position. Necessity comes first. If a joseki happens to result from that, then all well and good.


And:

Quote:
If you play as your opponent dictates, you cannot expect any good to come of it. Your thoughts must be directed towards avoiding such submissiveness.


So I believe that the fundamental concepts mentioned in the book are useful, but this specific position, and the variations given in the book, are not worth studying. They produce an average result, just like hundreds of related positions simply because neither player has made a serious mistake so far.


This post by Marcel Grünauer was liked by 2 people: Bill Spight, sorin
Top
 Profile  
 
Offline
 Post subject: Re: Reviewing book positions with AI
Post #2 Posted: Tue Jun 16, 2020 11:43 am 
Oza

Posts: 2574
Liked others: 16
Was liked: 3698
Without disagreeing with any specific element of what you are saying, Marcel, I think you are being too pessimistic.

I think we can even leave aside the probable fact that many book examples have been researched by amateurs under pro guidance which can vary from lackadaisical to attentive.

For a time I shared your pessimism. I felt it most keenly when I looked at Katsugo Shinpyo. This tesuji book had been touted by Fujisawa Hideyuki as the best thing since sliced bread, and so I chose this for my first venture into checking books with AI. The problem wasn't just finding apparent "mistakes" - you expect that anyway given the difficulty of setting up tesuji positions on a big board. It was the sheer number. I gave up part way through, not having found a single "correct" problem. It was doubly disheartening because the solutions looked correct to me!

A couple of things have changed my view. One is that looking at full games showed that pros make a HUGE number of moves that bots like. Pros genuinely are of a very, very high standard. That must include Fujisawa. So, my reasoning goes, surely he couldn't really have been wrong about KS. That didn't help much at the time, though, because I still couldn't explain the discrepancy.

The next thing that turned my view was that I kept coming across recent throw-away comments by top pros discussing their games, and the gist of these was that they knew that a move they played may not have been objectively the best but they preferred to make a move they understood. Playing move after move they way you think a bot might play is an incoherent, scattergun approach, and unless you are a Meijin you probably are wrong in most cases amyway. You quickly lose control. If you play a move you understand, you know what the follow-ups are likely to be. You retain a measure of control - and with knobs on when you are doing this against another human, and not a bot, which for a pro applies almost 100% of the time in their professional lives.

Seeing those comments also made me realise (at least so I think) what was going on with KS. I had been looking at examples of specific tesujis. What I should have been looking at was general suji. The flow of the combination. And control of the flow. KS is very, very good at that. I now think I see some of what Fujisawa saw. It was something like learning to pay one tune at a time by memorising how to plonk on a keyboard, always in the key of C. With suji, you can play in any key. And learning to play any tune any time from sheet music is like keeping control of a game. It still doesn't make you a concert pianist, of course. A busker with a monkey churning out tunes on a pianola might even play more accurately than you. But if you wanted improve your technique, would you spend a fortune on peanuts and turn to the organ grinder, or would you dig out your old book of Hanon exercises again?

And I don't think any of this glimmer of understanding of control, flow, suji or whatever is something I (and most others) would get from the present generation of bots. Bots wave flags but they don't do semaphore.


This post by John Fairbairn was liked by 2 people: Ferran, sorin
Top
 Profile  
 
Offline
 Post subject: Re: Reviewing book positions with AI
Post #3 Posted: Tue Jun 16, 2020 12:45 pm 
Lives with ko

Posts: 131
Liked others: 0
Was liked: 25
Marcel Grünauer wrote:
The books show some sequences; the AI shows some sequences. Who to believe?

Without referring to specific positions, two thoughts:

1. Bots are stronger than pros. This does not mean their "opinion" in a position is necessarily correct over pros', only that it is more often correct than pros'. In some positions it may be the bot that is wrong.

2. What does "correct" mean? Correct in minimax sense (ie. better assuming perfect play from that position)? Or correct in strategic sense (ie. - more often than not - better in similar positions, even if - from random noise - that strategy doesn't happen to work to a meaningful result in that particular position? I think bot opinion and human opinion target different correctness...

Top
 Profile  
 
Offline
 Post subject: Re: Reviewing book positions with AI
Post #4 Posted: Tue Jun 16, 2020 1:14 pm 
Honinbo

Posts: 9943
Liked others: 3219
Was liked: 3254
Marcel Grünauer wrote:
There are several books whose value in the AI age I'm questioning:

  • The Direction of Play (Kajiwara)
  • Whole Board Thinking in Joseki (Yang Yilun)
  • A Way of Play for the 21st Century (Go Seigen)
  • Improve Your Intuition (Takagawa, 3 booklets)
  • Get Strong at the Opening
  • A Dictionary of Modern Fuseki, The Korean Style

I intend to take some positions from some books and check them with my AI configuration of choice: KataGo 20-block (g170-b20c256x2-s4384473088-d968438914.bin.gz), using around 10k playouts, often more. I believe that stating these parameters is necessary. Too often it is stated that "AI would play like this" as though those were absolute sequences.
(emphasis mine)

Hear, hear! :)

Quote:
The first position is from "The Direction of Play". The Japanese original version was written in 1970, at a time when the komi was 5.5, so that is what I am using in Sabaki. With this komi, White starts out with a win rate of 44%, just for reference.

This position is shown on page 21, and the book says:

Quote:
If White encloses the [lower right] corner, Black extends to [the star point on the right side] and the game is over. Obviously that's an exaggeration, but White's forces are overconcentrated on one side of the board, whereas Black has taken control of the fourth line in a big way. This gives him such outside influence that it is hard to see how White's profit can match. If, after this, Black gets the sanrensei position with [the star point at the top], the game really will be over.


KataGo says that White can get 45.2% by invading the upper right 3-3, so it's a slight gain over the starting position. This is in stark contrast to the strongly-worded text.

When the book came out, Japanese amateurs, and even this foreigner, knew that Kajiwara was creative, eccentric, and prone to overstatement.

Quote:
I suppose the author wanted to stress that it is good if stones are working together, but this seems to be a poor example.


I know from experience creating problems that the law of unintended consequences applies. You cannot assume that what you mean to illustrate actually holds in your constructed position, especially a whole board position.

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

Banana Republic. It's not just a store anymore.

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: Reviewing book positions with AI
Post #5 Posted: Tue Jun 16, 2020 1:46 pm 
Honinbo

Posts: 9943
Liked others: 3219
Was liked: 3254
John Fairbairn wrote:
The next thing that turned my view was that I kept coming across recent throw-away comments by top pros discussing their games, and the gist of these was that they knew that a move they played may not have been objectively the best but they preferred to make a move they understood. Playing move after move they way you think a bot might play is an incoherent, scattergun approach, and unless you are a Meijin you probably are wrong in most cases amyway. You quickly lose control. If you play a move you understand, you know what the follow-ups are likely to be. You retain a measure of control - and with knobs on when you are doing this against another human, and not a bot, which for a pro applies almost 100% of the time in their professional lives.


Well, there are a few caveats. First, particularly in the early opening, a lot of bot plays are fairly predictable. Furthermore, it is well within the capability of pros to understand how to follow up those choices. The bot moves are not hard, just not what humans came up with.

Second, the principle of starting from where you are applies. Pros, having devoted their lives to the game, have a bank of userful knowledge, experience, and intuition. It makes no sense for them to throw all that away and try to play like a bot. As SDK, OTOH, has little to lose by trying to play like a bot. As a 2 kyu I tried to play like Go Seigen. OC, I did not come close, but I believe it helped.

Third. you do not lose control by imitation. You still search for the truth as best you can. Imitation guides your path. And if you choose a good guide, so much the better. :)

Quote:
And I don't think any of this glimmer of understanding of control, flow, suji or whatever is something I (and most others) would get from the present generation of bots. Bots wave flags but they don't do semaphore.


As lightvector and Marcel indicate, you use a bot as a tool, not a master. How many of us, coming along, could consult a pro every time we had a question? Yes, the bots do not speak, but used aright, they help us find answers to our questions.

To paraphrase the Buddha, Work out your own salvation, with the help of a bot. :lol:

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

Banana Republic. It's not just a store anymore.

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: Reviewing book positions with AI
Post #6 Posted: Tue Jun 16, 2020 1:52 pm 
Honinbo

Posts: 9943
Liked others: 3219
Was liked: 3254
jann wrote:
2. What does "correct" mean? Correct in minimax sense (ie. better assuming perfect play from that position)? Or correct in strategic sense (ie. - more often than not - better in similar positions, even if - from random noise - that strategy doesn't happen to work to a meaningful result in that particular position? I think bot opinion and human opinion target different correctness...


Yes. Humans do tend to strive for objective optimality. Bots try to win games.

That's by design. And one reason I would like to try to train a bot on difference games. I expect that such a bot would think more like a human. :)

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

Banana Republic. It's not just a store anymore.

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: Reviewing book positions with AI
Post #7 Posted: Wed Jun 17, 2020 12:00 am 
Lives with ko

Posts: 131
Liked others: 0
Was liked: 25
Bill Spight wrote:
Yes. Humans do tend to strive for objective optimality. Bots try to win games.

That's by design. And one reason I would like to try to train a bot on difference games. I expect that such a bot would think more like a human. :)

I'm not sure what "objective" mean here, but would guess it is closer to the minimax meaning of correctness, thus humans may be the less objective of the two.

And I doubt differences would change bot nature significantly. I think the fundamental difference is that - despite popular opinion - neural networks are nothing like human intuition. A NN learns (an approximation for) a complex function. Once learning is over that function can be pretty sharp, and is evaluated exactly. If you move a single stone a single step in a complicated position, NN output may change drastically (unlike human intuition, and closer to minimax truth).

A perfect player could feel similar, the limits of human strategy. Again: human wisdom is only guaranteed to be correct more often than incorrect (>50%), but bot wisdom may even be correct most of the time (>90%).

Top
 Profile  
 
Offline
 Post subject: Re: Reviewing book positions with AI
Post #8 Posted: Wed Jun 17, 2020 3:04 am 
Honinbo

Posts: 9943
Liked others: 3219
Was liked: 3254
Here is what I mean by using difference games to compare plays.
https://www.lifein19x19.com/viewtopic.php?f=15&t=17082

Neural networks trained on difference games will play differently on single boards than those trained as they are today. OC, they will be weaker per training time and resources, because difference games double the size of play and because they compare moves one at a time.

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

Banana Republic. It's not just a store anymore.

Everything with love. Stay safe.

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 8 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