It is currently Mon Nov 28, 2022 9:35 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 14 posts ] 
Author Message
Offline
 Post subject: Using AI to study general board positions
Post #1 Posted: Sat Nov 19, 2022 5:29 pm 
Lives in gote

Posts: 360
Liked others: 29
Was liked: 169
GD Posts: 1072
One of the down sides of AIs is that they only gives an opinion on a specific board position. My understanding is that this is because the input to the AI engine are two matrices (a black matrix and a white matrix) where the white and black stones are marked as a 1 at their respective positions. I am curious what would happen if a fractional value were to be input into these matrices?

The general problem I'm interested in is trying to get AI analysis of a corner position without great regard for the rest of the board. 15-20 years ago we would have used GoGOD and Kombilo to search on a corner pattern. Suppose we place a group of black stones in an adjacent corner with weights of 0.6 or something, to indicate some kind of black position there without being specific about what it is. Would the AI networks provide sensible output in that case?


This post by pwaldron was liked by: Elom0
Top
 Profile  
 
Offline
 Post subject: Re: Using AI to study general board positions
Post #2 Posted: Sat Nov 19, 2022 8:14 pm 
Lives in gote

Posts: 652
Location: Austin, Texas, USA
Liked others: 53
Was liked: 216
Interesting idea pwaldron! I linked this on the Computer Go Community discord. You may know that most the bots also have additional history inputs where they give the board state for the previous N moves. But katago trained sometimes with similar idea to yours -- the board state was not given so the net has to deal with sometimes the history is not there (I don't know if the history is given as all 1s, all 0s, all 0.5 or what).

Anyways for the near-term you have to use workarounds. Katago author posted some research on corner stuff where he would usually put some minimal stones in the other corners, like maybe just 4-4 stone or maybe 4-4 stone with some approach or similar.

Also some katago GUIs support a feature where it locks katago into only playing in a specific part of the board which could help.

Top
 Profile  
 
Offline
 Post subject: Re: Using AI to study general board positions
Post #3 Posted: Mon Nov 21, 2022 6:16 am 
Honinbo

Posts: 9488
Liked others: 1599
Was liked: 1673
KGS: Kirby
Tygem: 커비라고해
yoyoma wrote:

Also some katago GUIs support a feature where it locks katago into only playing in a specific part of the board which could help.


I'd imagine that this would be somewhat straightforward. If you can limit AI analysis to a 9x9 board or a 13x13 board, it'd seem possible to generalize: limit the set of valid moves to an arbitrary boundary.

I'd imagine that the trained policy network may produce suboptimal moves for arbitrary "board boundaries" since it was trained on a 19x19 board, but maybe it'd still produce a decent analysis...

_________________
be immersed

Top
 Profile  
 
Offline
 Post subject: Re: Using AI to study general board positions
Post #4 Posted: Mon Nov 21, 2022 8:57 am 
Dies with sente

Posts: 86
Location: Munich, Germany
Liked others: 336
Was liked: 17
Rank: EGF 5kyu
KGS: Akura
DGS: Akura361
yoyoma wrote:
I linked this on the Computer Go Community discord.


Would mind sharing a link to that discord server?

Top
 Profile  
 
Offline
 Post subject: Re: Using AI to study general board positions
Post #5 Posted: Wed Nov 23, 2022 6:13 am 
Lives in gote

Posts: 458
Location: Adelaide, South Australia
Liked others: 180
Was liked: 231
Rank: Australian 2 dan
GD Posts: 200
pwaldron wrote:
The general problem I'm interested in is trying to get AI analysis of a corner position without great regard for the rest of the board.


Kirby wrote:
yoyoma wrote:
I'd imagine that this would be somewhat straightforward. If you can limit AI analysis to a 9x9 board or a 13x13 board, it'd seem possible to generalize: limit the set of valid moves to an arbitrary boundary.


Yes, but that doesn't solve the stated problem, as the AI still "knows" what the rest of the board looks like. You've stopped it offering tenukis, but it will still evaluate moves in the chosen region by how well they work in the global position.

Consider a joseki choice problem: in the top right, you might prefer move "A" if black has thickness in the lower right, or move "B" if black is weak in the lower right. The AI will always be biased in one direction or another, while a human can give a judgement along the lines of "move A is the shape move for this position, but B would be a special-purpose move" or vice versa. How do you get the AI to give this "all else being equal" kind of opinion?

Top
 Profile  
 
Offline
 Post subject: Re: Using AI to study general board positions
Post #6 Posted: Wed Nov 23, 2022 9:21 am 
Gosei

Posts: 1661
Location: Earth
Liked others: 616
Was liked: 284
What is the underlying truth?

The whole board position always matters.

Only looking in a database for the moves in a restricted area without considering the whole board position was only half the task in the past. You had to check the games to get the complete picture.

Top
 Profile  
 
Offline
 Post subject: Re: Using AI to study general board positions
Post #7 Posted: Wed Nov 23, 2022 12:04 pm 
Lives in gote

Posts: 360
Liked others: 29
Was liked: 169
GD Posts: 1072
Gomoto wrote:
Only looking in a database for the moves in a restricted area without considering the whole board position was only half the task in the past. You had to check the games to get the complete picture.


Let's take a very hypothetical position to illustrate what motivated the question. Suppose we have the corner position below. It's a complicated position, with multiple variations.
Click Here To Show Diagram Code
[go]$$
$$ . . . . . . . . . |
$$ . . . . . . . . . |
$$ . . . . . . . X . |
$$ . . . . . X X O . |
$$ . . . . X O O . . |
$$ . . . . . X O . . |
$$ . . . . . X O . . |
$$ . . . . . . . . . |
$$ . . . . . . . . . |
$$ ---------------------[/go]


As you say, context is everything, but surely we ought to be able to draw conclusions on generalities as well--we've been playing go for a long time doing just that. So what variation might be expected when Black has some influence radiating from the bottom left? Here are three positions in that general category.

Click Here To Show Diagram Code
[go]$$
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . X . |
$$ | . O . . . . . . . . . . . . . X X O . |
$$ | . . . X . . . . . . . . . . X O O . . |
$$ | . . O X . . . . . . . . . . . X O . . |
$$ | . . O X . . . . . . . . . . . X O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------[/go]


Click Here To Show Diagram Code
[go]$$
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . O X . . . . . . . . . . . . . . X . |
$$ | . O X . . . . . . . . . . . . X X O . |
$$ | . O X . . . . . . . . . . . X O O . . |
$$ | . . O X . . . . . . . . . . . X O . . |
$$ | . . O X . . . . . . . . . . . X O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------[/go]


Click Here To Show Diagram Code
[go]$$
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . X . |
$$ | . . . . . . . . . . . . . . . X X O . |
$$ | . . . . . . . . . . . . . . X O O . . |
$$ | . . X . . X . . . . . . . . . X O . . |
$$ | . . . . . . . . . . . . . . . X O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------[/go]


The positions are different, but they have some similarities. When I posted the original question, the idea was whether I could distribute some group of black stones with fractional weight into the area to indicate that black, in essence, has friends in an adjacent area without fixing the position too much.

Top
 Profile  
 
Offline
 Post subject: Re: Using AI to study general board positions
Post #8 Posted: Wed Nov 23, 2022 1:19 pm 
Gosei

Posts: 1661
Location: Earth
Liked others: 616
Was liked: 284
My point is, probably disappointingly, that every single stone on the board is important :-)

Top
 Profile  
 
Offline
 Post subject: Re: Using AI to study general board positions
Post #9 Posted: Thu Nov 24, 2022 2:58 am 
Oza
User avatar

Posts: 2382
Location: Tokyo, Japan
Liked others: 2306
Was liked: 1325
Rank: Jp 6 dan
KGS: ez4u
Click Here To Show Diagram Code
[go]$$
$$ . . . . . . . . . |
$$ . . . . . . . . . |
$$ . . . . . . . X . |
$$ . . . . b X X O . |
$$ . . . c X O O . . |
$$ . . . . a X O . . |
$$ . . . . . X O . . |
$$ . . . . . . . . . |
$$ . . . . . . . . . |
$$ ---------------------[/go]

When I put this corner into katago, the general choice is between White a and White b. Along the way, it seems that whether the ladder after White b Black c works or not is at least as important as the lower left corner position. This seems to imply that "vague" information on the other parts of the board will not yield meaningful analysis.

_________________
Dave Sigaty
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21

Top
 Profile  
 
Offline
 Post subject: Re: Using AI to study general board positions
Post #10 Posted: Thu Nov 24, 2022 6:39 am 
Lives in gote

Posts: 360
Liked others: 29
Was liked: 169
GD Posts: 1072
ez4u wrote:
When I put this corner into katago, the general choice is between White a and White b. Along the way, it seems that whether the ladder after White b Black c works or not is at least as important as the lower left corner position. This seems to imply that "vague" information on the other parts of the board will not yield meaningful analysis.


We could just as easily ask about inserting vague information into the opposite corner to influence ladder questions. Fundamentally my question was whether providing stones with fractional weights as inputs to the neural net would provide a practical way of providing vague information that would influence the results.

Top
 Profile  
 
Offline
 Post subject: Re: Using AI to study general board positions
Post #11 Posted: Thu Nov 24, 2022 6:44 am 
Lives in gote

Posts: 360
Liked others: 29
Was liked: 169
GD Posts: 1072
Gomoto wrote:
My point is, probably disappointingly, that every single stone on the board is important :-)


We know this can't be the universal situation. We have dozens of joseki books that look at corner sequences, and we know that, to a first approximation, the direct 3-3 invasion is a reasonable move in the opening regardless of what the adjacent corners are.

Whether that level of vagueness can be conveyed to an AI is a different question, and that was the one I was interested in.

Top
 Profile  
 
Offline
 Post subject: Re: Using AI to study general board positions
Post #12 Posted: Thu Nov 24, 2022 7:18 am 
Oza

Posts: 3395
Liked others: 20
Was liked: 4394
Quote:
The general problem I'm interested in is trying to get AI analysis of a corner position without great regard for the rest of the board.


I think this question is bedevilled by the confusing word 'analysis'.

From previous discussions of positions put through the AI meat grinders, most people seem to want to know the evaluations for various game moves based on the point score for the AI's best move. But as you are happy to disregard the rest of the board to some degree, I infer that is not really what you are looking for.

I further infer that you want to see various lines locally, and to get some sense of which ones may generally be the best. I would say that corresponds to the usual sense of 'analysis' in chess, which of course is a highly tactical game. If this is your area of interest, I'd further infer that you are hoping to want to use AI simply to get more tactical insights.

If all that is correct (probably not, in light of the thread title 'general' board positions), you may wish to be aware that Japanese magazines take this latter approach quite often, and Go Weekly even has a regular column devoted to it. They never talk about point scores or percentages (hooray!) but do talk about various unexpected tactical lines in corner positions that have been found recently (I assume AI input is to be taken as read). There is then a brief evaluation in terms either of what a particular line achieves (e.g. sabaki or thickness) or whether a line is playable or the like. There is rarely any mention of a wider board position, and no indication of how an AI search, if any, has been carried out, but I'm guessing that they just run a position through a bot and see what unexpected moves turn up, and the humans take it from there.

I think the column has been running for close to two years now, but maybe just looking at two or three may be enough to show you how to do it all for yourself, plus the sort of rewards this approach might give.

Incidentally, I may have just missed it as I don't pay much attention to that column, but I've been vaguely surprised that there has not yet been any use of tewari in discussing these positions in Go Weekly. People of my generation will recall the wonder of discovering tewari in Takagawa's Vital Points, a wonder that has stayed with me.

Top
 Profile  
 
Offline
 Post subject: Re: Using AI to study general board positions
Post #13 Posted: Thu Nov 24, 2022 8:01 am 
Lives in gote

Posts: 360
Liked others: 29
Was liked: 169
GD Posts: 1072
Your summary largely hits what I'm after, John, including the vagaries of the word 'analysis'.

My general interest was trying to get back up to speed after a hiatus away from go. There's new josekis since the last time I played seriously and I'm trying to figure what plays are viable in different contexts. It got me thinking about how to use an AI to get that kind of information.

My phrasing about 'general board positions' was the idea that a line of play (joseki?) certainly depends on the full board position, but likely not the exact position. Positions with similar general properties--outside influence for one colour in an adjacent corner or (un)favourable ladders--will likely have similar continuations. Maybe that's not entirely true, but I suspect it's reasonable. If every single stone mattered critically in a position then we wouldn't have joseki that survive from game to game.

But how to get the AI to deal with that? If I firmly plant down a stone on the board and start the AI then it's going to do its AI thing and give me a specific line of play for a specific position. The idea was that maybe a single stone weight (=1.0) could be distributed around 5 neighbouring points with a fractional weight. Might a neural network interpret that as some kind of vague statement about the type of position it was dealing with?

The initial question was largely to the AI programming experts who would understand what input to a neural network actually represented.

Top
 Profile  
 
Offline
 Post subject: Re: Using AI to study general board positions
Post #14 Posted: Fri Nov 25, 2022 4:19 pm 
Lives in gote

Posts: 458
Location: Adelaide, South Australia
Liked others: 180
Was liked: 231
Rank: Australian 2 dan
GD Posts: 200
pwaldron wrote:
My phrasing about 'general board positions' was the idea that a line of play (joseki?) certainly depends on the full board position, but likely not the exact position. Positions with similar general properties--outside influence for one colour in an adjacent corner or (un)favourable ladders--will likely have similar continuations. Maybe that's not entirely true, but I suspect it's reasonable. If every single stone mattered critically in a position then we wouldn't have joseki that survive from game to game.

But how to get the AI to deal with that?

I'm with you on this, and I believe it should be possible, but it would take a fair bit of work. Then again, I'm a dabbler in neural nets, not by any means an expert.

To put it crudely, a network contains a bunch of layers, the earlier layers capture local information about shape (e.g. recognising when a stone is in atari, or learning that empty triangles are likely to be bad), and the later layers are about building up that local knowledge into full-board judgement. A human is able to look at the whole board and choose what we think is the best move; a human can also choose to be a bit vague about the whole-board position and talk about the best local move, or the best move for a class of whole-board positions (e.g. "when black has thickness in the lower left" without specifying the exact shape). I believe it should be possible for an AI to "filter" the network in a similar way and pull out different sorts of judgements. The network contains more knowledge about go than is in any one human brain, so the answers you're looking for must be in there somewhere! But I'm not aware that anyone has tried.

I have a feeling that your suggestion of assigning fractional values to stones will cause weird side effects. But it's worth trying it anyway to see exactly what does go wrong. If I somehow end up with a spare 20 hours when I feel like hacking away at some code, I might give it a go. But I don't foresee that happening soon, sorry. Hopefully one of the real experts is up for such an experiment.

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 14 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: Google [Bot], jeromie 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