Interactive Go Maps: Art, Not Science
Frank de Groot
influence simulator in Interactive Go Maps:
http://www.moyogo.com/blog/2006/12/interactive-go-maps-art-not-science.html
In my opinion (elaborated on my blog), the author (who is the author?) did
not intend to make an accurate influence simulator, but an visually pleasing
approximation of "influence" and various other concepts he invented.
It looks stunningly beautiful but I am quite sure that compared to a Bouzy
map, those on
http://canut-ki-in.jeudego.org/simulation_influence/
are (much) less accurate.
I mention this because it is not the first time that people asked me to
include such a gimmick in my software, and I always reply: "I am not an
artist but a scientist".
And just a few hours ago a kid named David Windsor claimed that my carefully
researched influence algorithm (based on the research of computer Go
scientists like Bruno Bouzy) is "crap" and that the aforementioned artwork
is much better.
I am never too old to learn and admit I am a lousy Go player - please,
someone unbiased and preferably a Go programmer - tell me what's wrong with
my analysis in my blog posting:
Interactive Go Maps: Art, Not Science
http://canut-ki-in.jeudego.org/simulation_influence/ has a beautiful
"influence simulator" and quite a few people asked me: "Why don't you put
something like that in Moyo Go"? Well.. Because I'm a computer-Go scientist,
not an artist :-)
I want to make it clear that I have great respect for the programmer of this
stunningly beautiful display, and I'm sure the algorithms used are not
simple at all. Yet, it does not appear to be a genuine influence simulator,
more a kind of "Go art" or perhaps the programmer's own interpretation of
"influence".
Because "influence" in comp. Go terms is not something subjective, or
something "beautiful", no, we design influence functions and measure their
performance with tens of thousands of end positions of pro games and keep
tuning those functions until they give the most accurate estimation of
territory and influence (which is more subjective). Territory is subjective
as well, but the abberations average eachother out so we can establish
accurate performance parameters for our influence functions.
Take this example of influence maps. (see image on blog)
In computer Go, the sub-problem of "influence" has been quite acceptably
solved by Bruno Bouzy, who has published several papers on his method.
Both Bouzy's and David Fotland's (of Many Faces of Go fame) influence
functions "bounce" influence off the edges of the board and let the
"wavefronts" combine. This is visible in the lower right corner of Moyo Go's
example (right), but not in the "interactive Go map" on the left because a
much simpler, more inaccurate algorithm is used that does not even take
Manhattan distances into account. The purpose of Interactive Go maps appears
to test and present Go-concept ideas in an esthetical manner, without laying
claim to being "accurate".
And there are more things that are sub-optimal in the example on the left.
Look at those white stones in the lower left. They project quite some
influence leftward, but in the Interactive Go map there is nothing!
The analogy Go stones = magnets goes a long way, but in the end, Go stones
in a Go position are not magnets and more realistic algorithms need to be
employed to predict likely territory and influence, especially in more
complex positions.
Let the Interactive Go maps be pretty, and let Moyo Go's influence maps be
as accurate as possible. No effort has obviously been spared to make those
maps as pleasing to the eye as possible, while I read my way through a pile
of state-of-the art research in influence algorithms. The Bouzy algorithm
remains unbeaten, but I found a way to greatly speed it up and get almost
the same (or better?) results. Rest assured that although it doesn't look as
good as the Interactive Go maps, no efforts have been spared to make them as
useful as possible, and not merely good-looking.
Bantari
> And there are more things that are sub-optimal in the example on the left.
> Look at those white stones in the lower left.
Don't you mean "lower right" here?
____________
-Bantari
http://www.bantari.net/
Bill Spight
> In computer Go, the sub-problem of "influence" has been quite
> acceptably
> solved by Bruno Bouzy, who has published several papers on his method.
>
Bouzy's method is interesting, but I do not think that it can be
considered a solution, or even close. If go had a good influence
function, it would have a good evaluation function, and programs would
be much stronger.
Cheers,
Bill
Frank de Groot
> Bouzy's method is interesting, but I do not think that it can be
> considered a solution, or even close. If go had a good influence
> function, it would have a good evaluation function, and programs would
> be much stronger.
Bouzy's method has to be combined with a lot of other stuff (esp. L&D) but
it's the best "we" (Go programmers) have, AFAIK.
Do you think that the only thing needed in an evaluation function is the
"influence factor"?
You mean that "influence" is in fact "everything combined"?
Bill Spight
I suspect that the best evaluation function currently achievable has
thousands of features (variables), none of which could be called
influence. Influence is a vague, informal term that indicates the effect
of stones on enemy stones and on empty points that are not yet
territory. Various influence functions are attempts to formalize that
vague intuition. I am unaware of one that more than partially captures
the notion.
That is not necessarily a bad thing. Informal notions often break down
into a number of more concrete concepts. Perhaps it is best to think of
the different influence functions as defining different things. Their
eventual utility is hard to predict at present.
Ciao,
Bill
Dav...@gmail.com
On Dec 1, 4:19 pm, "Frank de Groot" <f...@moyogo.com> wrote:
> And just a few hours ago a kid named David Windsor claimed that my carefully
> researched influence algorithm (based on the research of computer Go
> scientists like Bruno Bouzy) is "crap" and that the aforementioned artwork
> is much better.
Get your names right, that wasnt me.
Dav...@gmail.com
"Vit Brunner" <vit.brun...@gmail.com> wrote
> I bet that most people read your blog just for amusement value of
> things like your "influence diagram" (yeah, I know, you are not a go
> player and you have no idea what influence is, but maybe this is why it
> might help to know something about go before writing a go program). By
> the way I highly suggest you to check this excellent influence
> simulator: http://canut-ki-in.jeudego.org/simulation_influence/
Vit Brunner
Yeah, it was me, although I didn't say it was crap (but yeah, I did
think so).
I still want to hear why the dead stones in your diagram (
http://moyogo.com/blog/uploaded_images/influence.png ) exert influence.
If you enter the same position to that French influence diagram the
dead stones exert *no* influence. This shows very clearly which one is
better.
The rest of the your influence diagram is also inferior to the French
one, but it's impossible to explain why to someone who is not a go
player...
Frank de Groot
> The rest of the your influence diagram is also inferior to the French
> one, but it's impossible to explain why to someone who is not a go
> player...
Nonsense. Influence doesn't equal life.
Influence is influence. The stones influence the points. It's up to you to
make them live.
L&D status has nothing to do with influence. as soon as a tactical situation
gets very complex, computers do not know the L&D status, but they still know
the Influence.
So you're attacking windmills here.
I gave several arguments why the French one didn't make sense (backed up by
various strong players in email to me), you keep your "arguments" secret
because you fear critiscism of stronger players.
It is clear that there is no L&D factor in my influence diagram. It's
mentioned in my blog posting and in the Help file. It's a work-in-progress
and I blog about that. I also roll out to customers features that are not
100% perfect yet. It's called "agile development" (aspects of XP are present
as well).
BTW it's all a matter of definition. Those stones DO exert influence. It's
just that "influence" DOES NOT EQUAL LIFE.
Get it?
Bakc to the study books for you!
Frank de Groot
> Nonsense. Influence doesn't equal life.
> Influence is influence. The stones influence the points. It's up to you to
> make them live.
I have to add that I'm speaking purely from a comp. Go POV.
There is an influence module and a L&D module and a pattern module and a
ladder module etc.
You can't "fuzzy it up" like Go players who are not programmers often do.
Influence is influence. It always exists.
Saddam Hussein had influence over his generals, even though he was a "dead
man".
I approach (also due to me not being an expert Go player) the comp. Go field
in a scientific manner.
I do not try to mimic the thought processes of the masters, I rather measure
my progress in exact ways.
When I add the future L&D module to the influence module, it will look
differently (but will take much longer to scroll real-time through a game
with influence-display turned on).
Tools are not always for people who can't use them.
At the moment, you should not confuse "influence" in my software with
"life".
So when you are strong enough to keep stones alive (or judge their status),
the influence function is pretty cool.
And it works wonders at the opening stage, where there are little if any L/D
fights.
This is how other major Go software vendors market their influence feature
as well.
"Very useful in Fuseki".
Frank de Groot
Addendum II:
I'm open for /constructive/ critiscism about that influence function.
From customers /only/.
My goal is to make it as best as I can, not a pissing contest.
Vit Brunner
> Nonsense. Influence doesn't equal life.
That's true, however, in the normal understanding of influence (
http://senseis.xmp.net/?Influence is pretty close, I think), the stones
have to be alive to exert influence.
> Influence is influence. The stones influence the points. It's up to you to make them live.
So you want to say that the point B17 in this diagram:
http://moyogo.com/blog/uploaded_images/influence.png
is in white's sphere of influence? Sorry but that's pure nonsense...
> L&D status has nothing to do with influence.
Wow... now that's a bold (and wrong) statement...
http://senseis.xmp.net/?Influence
> as soon as a tactical situation gets very complex, computers do not know the L&D status, but they still know the Influence.
For a group to exert influence it has to be alive (also, if you feel
that a group exerts a lot of influence you can consider it alive).
> So you're attacking windmills here.
I'm not attacking anything, I am just trying to help you understand
what "influence" is.
> I gave several arguments why the French one didn't make sense (backed up by
> various strong players in email to me), you keep your "arguments" secret
> because you fear critiscism of stronger players.
I provided one very clear argument. The other are not relevant at this
stage. As soon as some strong player enters the discussion and asks I
will start discussing it with him (there is no point in discussing the
finer points with you when you are not a go player, I am just trying to
explain basics).
> It is clear that there is no L&D factor in my influence diagram. It's
> mentioned in my blog posting and in the Help file. It's a work-in-progress
> and I blog about that. I also roll out to customers features that are not
> 100% perfect yet. It's called "agile development" (aspects of XP are present
> as well).
I know that there is no L&D in your diagram, the French diagrams also
don't feature L&D and yet are able to recognize that a white stone
surrounded by black ones doesn't exert influence.
> BTW it's all a matter of definition. Those stones DO exert influence. It's
> just that "influence" DOES NOT EQUAL LIFE.
This is a valid point, if you redefined the term influence to mean "the
color of a stone that is near", then you are right. But how should I
have known?
Vit Brunner
So why are you writing this to rgg? Write it to your customers and not
here...
Bill Spight
> L&D status has nothing to do with influence. as soon as a tactical
> situation
> gets very complex, computers do not know the L&D status, but they
> still know
> the Influence.
>
Influence, as understood by go players, is definitely related to life
and death. Dead stones exert little influence, if any. More generally,
strong stones exert more influence than weak stones.
But all of that has to do with the vague, intuitive notion of influence.
If you define influence as the result of an influence function, that's
another matter. Then each different function defines its own brand of
influence.
Ciao,
Bill
Dav...@gmail.com
> "Bill Spight" <bspi...@pacbell.net> wrote
> Bouzy's method has to be combined with a lot of other stuff (esp. L&D) but
> it's the best "we" (Go programmers) have, AFAIK.
>
> Do you think that the only thing needed in an evaluation function is the
> "influence factor"?
>
> You mean that "influence" is in fact "everything combined"?
How does influence software deal with aspects of go that are more
intuition based (in the context of actual play)?
For instance, what ways would an intuitional algo deal with concepts of
shape (good shape having better influence) and proximity with regards
to shape (say a wall of one color exerting less influence than it
ought' because its sitting around some opponents stones (who arent
necissarilly cose, but have a very strong/solid/thick shape)?
Does such an program take into account that groups susceptible to bieng
cut excerpt less influence than those that are thick? and if so, in
what ways does it approach the subject of potential cuts?
My understanding of Bouzy's method is very limited in that im no
programmer, but the idea that a program (theoretical or otherwise) can
(calculate?) influence is very cool; provided it approaches the subject
with more thought than assuming all stones just radiate influence with
regard to their proximity to certain parts of the board.
Regards.
Dav...@gmail.com
On Dec 4, 5:43 am, David Gillen <Bel...@RedBrick.DCU.IE> wrote:
> > Get your names right, that wasnt me.Lets see how long it takes Frank to apologise.
>
> D.
> --
henricb...@gmail.com
Vit Brunner skrev:
> On Dec 2, 6:33 am, "Davo...@gmail.com" <Davo...@gmail.com> wrote:
> > On Dec 1, 4:19 pm, "Frank de Groot" <f...@moyogo.com> wrote:
> > > And just a few hours ago a kid named David Windsor claimed that my carefully
> > > researched influence algorithm (based on the research of computer Go
> > > scientists like Bruno Bouzy) is "crap" and that the aforementioned artwork
> > > is much better.
> > Get your names right, that wasnt me.
>
> Yeah, it was me, although I didn't say it was crap (but yeah, I did
> think so).
>
> I still want to hear why the dead stones in your diagram (
> http://moyogo.com/blog/uploaded_images/influence.png ) exert influence.
> If you enter the same position to that French influence diagram the
> dead stones exert *no* influence. This shows very clearly which one is
> better.
Hello Vit.
Which stones do you mean, the two white ones in the upper right? I
would think that dead stones provide for forcing moves in the vicinity.
Also, if a ko comes up somewhere on the board they may turn out to be
very valuable and become a menace in the local area, compared to the
situation without them, that should count for something too, shouldn't
it? And what about Franks statement that an isolated should excert
more influence horisontally and vertically than diagonally, is that
likely to be correct or not?
I'm not familiar at all with the subject of this thread, I'm missing
references to what's been published by Bruno Bouzy and others. Is there
any review article anywhere online? What is Manhattan distance, for
instance?
> The rest of the your influence diagram is also inferior to the French
> one, but it's impossible to explain why to someone who is not a go
> player...
This is a pretty arrogant line of argument and not very interesting to
anyone IMHO. What's wrong with commenting in concrete terms on what's
being proposed here instead. I would have thought that this thread is
unusually constructive and interesting from a go point of view, for
being one of Frank's.
I guess that "influence" is bound to remain a vague concept, at the
very least it must depend on many local tactical details. The same for
these kinds of maps. Still, they may have great value, didactically for
example. I played around a little with the french program, but I don't
get the feeling that it's particularly sensitive to the presence of
possible cuts and such things. It would be nice to know a bit more
about what the maps, Franks, the French one and others are based on.
But I wouldn't like to spend too much time on any detailed technical
descriptions. A sensible discussion like the one Frank appears to
suggest might be interesting to follow, though.
best regards,
Henric
Bill Spight
On Tue, 2006-12-05 at 08:09 -0800, henricb...@gmail.com wrote:
> And what about Franks statement that an isolated should excert
> more influence horisontally and vertically than diagonally, is that
> likely to be correct or not?
>
I doubt it. See later with regard to Manhattan distance.
> I'm not familiar at all with the subject of this thread, I'm missing
> references to what's been published by Bruno Bouzy and others. Is
> there
> any review article anywhere online? What is Manhattan distance, for
> instance?
>
Manhattan distance is just the distance moving point to point.
(Euclidean distance is the straight line distance, going between points
if necessary.)
4 3 2 3 4
3 2 1 2 3
2 1 # 1 2
3 2 1 2 3
4 3 2 3 4
The numbers show the Manhattan distance of each point from #.
One aspect of influence (in the informal sense) is connectivity. If
White cannot cut two Black stones apart, it is in part because of the
influence the stones exert on each other.
I am not sure what Frank means when he says that an isolated stone
exerts more influence horizontally or vertically than diagonally, but
let's look at connectedness in terms of Manhattan distance.
. . . . .
. . . . .
. . # # .
. . . . .
. . . . .
(This is a 5x5 region of the full board.)
These two stones are 1 point apart, and solidly connected.
. . . . .
. . . # .
. . # . .
. . . . .
. . . . .
These two stones are 2 points apart. They are connected, but the
connection may be threatened. It is possible that they may be cut in the
course of the game.
. . . . .
. . . . .
. # . # .
. . . . .
. . . . .
These stones are also 2 points apart, but they are not connected.
(Usually it will not pay for White to disconnect them, but that's
another matter.)
. . . . .
. . . . .
. . . # .
. # . . .
. . . . .
These stones are 3 points apart, but they are connected on an empty
board, because of the possibility of a ladder. (Note that these stones
are further apart than the previous two in terms of Euclidean distance
as well as Manhattan distance.)
All of this suggests that connectivity, and hence influence, is stronger
in a diagonal direction than horizontally or vertically. In a real game,
horizontal influence is usually the strongest, because of the
interaction with the edge of the board.
> I guess that "influence" is bound to remain a vague concept, at the
> very least it must depend on many local tactical details. The same for
> these kinds of maps.
Yes. As I said in another post, I think that eventually the vague notion
of influence will break down into several more concrete concepts.
Cheers,
Bill
Vit Brunner
> Which stones do you mean, the two white ones in the upper right?
Yes.
> I would think that dead stones provide for forcing moves in the vicinity.
> Also, if a ko comes up somewhere on the board they may turn out to be
> very valuable and become a menace in the local area, compared to the
> situation without them, that should count for something too, shouldn't it?
You are right, the stones still have quite a lot of aji. But aji is
just what it is - it's aji, not influence (at least in usual meaning).
Nevertheless, you can't really say that P17 is under white's sphere of
influence (I know that you didn't say it, but Frank's diagram shows
just that). It's a solid black territory, not white's sphere influence.
> And what about Franks statement that an isolated should excert
> more influence horisontally and vertically than diagonally, is that
> likely to be correct or not?
I think this is correct, but it's just a feeling which is very
difficult to describe...
I am sorry for sounding arrogant but it's difficult not to be arrogant
to someone who is not even a go player (eg ~ 20kyu) and yet is
"educating" me about what influence is (and gets it totally wrong, as
expected).
Vit
henricb...@gmail.com
Hello Bill,
thanks for the reply, I think it's interesting to see how far it's
possible to get by just general arguments.
Bill Spight skrev:
> Manhattan distance is just the distance moving point to point.
> (Euclidean distance is the straight line distance, going between points
> if necessary.)
>
> 4 3 2 3 4
> 3 2 1 2 3
> 2 1 # 1 2
> 3 2 1 2 3
> 4 3 2 3 4
Ok, thanks, I should have guessed :-)
> One aspect of influence (in the informal sense) is connectivity. If
> White cannot cut two Black stones apart, it is in part because of the
> influence the stones exert on each other.
>
> I am not sure what Frank means when he says that an isolated stone
> exerts more influence horizontally or vertically than diagonally, but
> let's look at connectedness in terms of Manhattan distance.
>
> . . . . .
> . . . . .
> . . # # .
> . . . . .
> . . . . .
>
> (This is a 5x5 region of the full board.)
>
> These two stones are 1 point apart, and solidly connected.
>
> . . . . .
> . . . # .
> . . # . .
> . . . . .
> . . . . .
>
> These two stones are 2 points apart. They are connected, but the
> connection may be threatened. It is possible that they may be cut in the
> course of the game.
Yes, or at any rate the opponent has forcing moves from either side
which may prove useful, e.g. in making eyes for a group in the area.
This makes the diagonal connection a bit weaker than the horisontal
one, isn't it fair if this is reflected in the scale length of
influence attenuation?
> . . . . .
> . . . . .
> . # . # .
> . . . . .
> . . . . .
>
> These stones are also 2 points apart, but they are not connected.
> (Usually it will not pay for White to disconnect them, but that's
> another matter.)
> . . . . .
> . . . . .
> . . . # .
> . # . . .
> . . . . .
>
> These stones are 3 points apart, but they are connected on an empty
> board, because of the possibility of a ladder. (Note that these stones
> are further apart than the previous two in terms of Euclidean distance
> as well as Manhattan distance.)
>
> All of this suggests that connectivity, and hence influence, is stronger
> in a diagonal direction than horizontally or vertically.
I'm not sure that I follow you here. If we only care for the number of
moves required to get connection over a distance x between two points,
then as you say, a diagonal move is as connected as a horisontal move.
By this argument neither Manhattan distance, nor Euclidean distance (
in the sense that the diagonal of a square is longer than the side) is
relevant, only the number of spaces, regardless whether they are
diagonal or horisontal. With that measure a keima only covers 2 units.
Under this assumption every isolated stone is surrounded by a "square
of influence". However, if you admit, as you did, that the diagonal
connection is weaker than the horisontal one, then this result should
be modified in the direction that Frank prefers, shouldn't it? Also,
the one space diagonal 'elephant' jump is very poorly connected, in
fact often wrong, if the vertical one space jump is considered "never
wrong".
The saying "extend N+1 spaces from an N-stone wall" is a bit related to
connectivity for N<3 at least. Is there any rational argument for it
for larger N? What do you think should be the rule of thumb for how far
to extend from a diagonal "wall"?
Of course there are other considerations, besides connectivity. Like
"if I extend a distance x, how probable is it that an invasion by the
opponent will find enough space to live?" But also ladders! Where a
fight breaks out it's usually important to control the ladders, and
ladderbreaking power propagates diagonally. Perhaps any influence map
that does not take ladders into account is seriously flawed?
> In a real game,
> horizontal influence is usually the strongest, because of the
> interaction with the edge of the board.
I'm not sure how the edge should come in, in a general discussion.
Frank is talking about waves of influence being "reflected" at the
wall. That seems hard to justify, to me. I would expect there to be a
huge difference in influence towards the edge, if it's possible to live
there or not - like the influence of the hoshi towards the corner
should be very limited by the possibility for the opponent to get easy
life there.
> Yes. As I said in another post, I think that eventually the vague notion
> of influence will break down into several more concrete concepts.
Probably :-)
Ciao,
Henric
Bill Spight
>
> > One aspect of influence (in the informal sense) is connectivity. If
> > White cannot cut two Black stones apart, it is in part because of the
> > influence the stones exert on each other.
> >
> > I am not sure what Frank means when he says that an isolated stone
> > exerts more influence horizontally or vertically than diagonally, but
> > let's look at connectedness in terms of Manhattan distance.
Since then I found where Frank opined that the influence function around
tengen on a sparse board should be diamond shaped rather than square,
because the points on the board are connected horizontally or
vertically. That's not exactly the same thing. However, since I talked
about connectivity, I think that my remarks apply to what he said.
> >
> > . . . . .
> > . . . . .
> > . . # # .
> > . . . . .
> > . . . . .
> >
> > (This is a 5x5 region of the full board.)
> >
> > These two stones are 1 point apart, and solidly connected.
> >
> > . . . . .
> > . . . # .
> > . . # . .
> > . . . . .
> > . . . . .
> >
> > These two stones are 2 points apart. They are connected, but the
> > connection may be threatened. It is possible that they may be cut in the
> > course of the game.
>
> Yes, or at any rate the opponent has forcing moves from either side
> which may prove useful, e.g. in making eyes for a group in the area.
> This makes the diagonal connection a bit weaker than the horisontal
> one, isn't it fair if this is reflected in the scale length of
> influence attenuation?
>
It is an interesting question whether to take the slight probability of
a cut into account in a connectivity function. I suspect that it is
better to deal separately with the interaction with enemy stones that
would make allowing the cut advisable. E. g., damezumari.
The keima is better connected than either the elephant jump or the one
space jump. That's a curious fact that does not fit well with either
Manhattan or Euclidean distance. Certainly Manhattan distance does not
have a simple relation to connectivity.
The main reason, I think, that connectivity is better along a diagonal
is that the number of potential connections matters. Why is the diagonal
connection connected, even though the Manhattan distance is 2? One
reason is that there are two minimal connecting paths, and the opponent
cannot block both. The one space jump is cuttable because there is only
one minimal path, which can be blocked in one play.
> The saying "extend N+1 spaces from an N-stone wall" is a bit related to
> connectivity for N<3 at least. Is there any rational argument for it
> for larger N? What do you think should be the rule of thumb for how far
> to extend from a diagonal "wall"?
>
In fact, as I learned it, the reason for that rule of thumb is that the
extension is connected to the wall. (That depends upon the edge of the
board, which allows some invading stones to be captured.)
> Of course there are other considerations, besides connectivity. Like
> "if I extend a distance x, how probable is it that an invasion by the
> opponent will find enough space to live?" But also ladders! Where a
> fight breaks out it's usually important to control the ladders, and
> ladderbreaking power propagates diagonally. Perhaps any influence map
> that does not take ladders into account is seriously flawed?
>
I don't think so. It's OK to leave out specific tactical considerations.
For instance, while on an empty board the keima is connected, that
depends upon a ladder. Once the board is sufficiently filled in, maybe
we can take the chance that the ladder works as 1/2. Then it makes sense
to say that the keima is 1/2 connected.
Cheers,
Bill
Frank de Groot
> I'm not sure how the edge should come in, in a general discussion.
> Frank is talking about waves of influence being "reflected" at the
> wall. That seems hard to justify, to me.
Amongst comp. Go scientists, this is quite undisputed, with the currently
best influence functions we have.
And before anyone tells me that I am "not a Go player", and want to politely
remind them of the fact that it will be computer scientists that will
eventually and inevitably make the program that will win from a 9p player.
Frank de Groot
> huge difference in influence towards the edge, if it's possible to live
> there or not - like the influence of the hoshi towards the corner
> should be very limited by the possibility for the opponent to get easy
> life there.
I found a new Law of Nature:
"Everyt time a computer Go scientist shows of part of / progress of his
work, his work will be misunderstood and thus deemed flawed by those who
rank higher in Go strength".
henricb...@gmail.com
Frank de Groot skrev:
Actually, I at least am not likely to deem anything flawed in this
subject. I'm not familiar with any relevant literature, I wouldn't even
know where to find it, but I'm a bit interested in what those influence
maps are based on, so I'll appreciate to see any further rational
discussion along the lines that you initiated.
cheers,
H.
Vit Brunner
> The saying "extend N+1 spaces from an N-stone wall" is a bit related to
> connectivity for N<3 at least. Is there any rational argument for it
> for larger N? What do you think should be the rule of thumb for how far
> to extend from a diagonal "wall"?
Hi,
I think that the walls are usually horizontal or vertical, or just
irregular. I've never really seen a diagonal wall yet. :-)
Mhmm... let's try:
. . . . . . . . . . . |
. . . . . . . . . . . |
. . . . . . . a c . . |
. . . . . . . d b . . |
. . . . . . . . . . . |
* . . . . . . * . . . |
. . . . . . . . . . . |
. . . . . . . . . . . |
. . . O . . . . . . . |
. . . . O . . . . . . |
. . . . . O . . . . . |
. . . . . . O . . . . |
* . . . . . . O . . . |
. . . . . . . . O . . |
. . . . . . . . . . . |
. . . . . . . . . . . |
----------------------+
a b c d are the extensions that I'd consider appropriate (sorted by my
personal preference), concerning that there are no other stones on the
board and we have to make an extension (normally I'd prefer not to
extend but to occupy a big point), and if there were stones somewhere,
they would influence my decision about extending rather heavily.
I didn't even try making an extension from a diagonal wall into the
center - that's the kind of extension you only make when your stones
are endangered, and we'd really need to know the surroundings for that.
Vit
Rich
> > L&D status has nothing to do with influence. as soon as a tactical
> > situation
> > gets very complex, computers do not know the L&D status, but they
> > still know
> > the Influence.
Whether they know it correctly or not depends entirely on the
definition and assessment of influence, of course.
> Influence, as understood by go players, is definitely related to life
> and death. Dead stones exert little influence, if any. More generally,
> strong stones exert more influence than weak stones.
It seems possible that if Frank's algorithm combines (Go-players')
influence with aji, dead (or perhaps "capturable" is a more relevant
term) stones could have (Frank's) influence - at least by reducing the
effectiveness of the opponent's nearby influence.
Cheers
Rich
Rich
> > And what about Franks statement that an isolated should excert
> > more influence horisontally and vertically than diagonally, is that
> > likely to be correct or not?
>
> I think this is correct, but it's just a feeling which is very
> difficult to describe...
There's also the point that ladder-breaking only works diagonally, not
vertically and horizontally. Should the influence of a tengen play be
an X-shape? ;)
Rich
Bill Spight
The thing is, Frank, that you appear to be making comments about go. If
that is not the case, please clarify what you have said, to help clear
up misunderstanding.
You wrote:
> I would like to start a new discussion about the practical use of the
> influence simulator in Interactive Go Maps:
>
> http://www.moyogo.com/blog/2006/12/interactive-go-maps-art-not-science.html
>
> In my opinion (elaborated on my blog), the author (who is the author?)
> did
> not intend to make an accurate influence simulator, but an visually
> pleasing
> approximation of "influence" and various other concepts he invented.
>
The idea of influence in go has been around for a long time. No one
person has invented it.
>
> It looks stunningly beautiful but I am quite sure that compared to a
> Bouzy
> map, those on
> http://canut-ki-in.jeudego.org/simulation_influence/
>
> are (much) less accurate.
>
Here you appear to be saying that the influence map is a less accurate
estimator of influence than Bouzy's function. On its face, that is a
statement about go, since influence is a go idea.
[snip]
> I am never too old to learn and admit I am a lousy Go player -
> please,
> someone unbiased and preferably a Go programmer - tell me what's wrong
> with
> my analysis in my blog posting:
>
> Interactive Go Maps: Art, Not Science
>
> http://canut-ki-in.jeudego.org/simulation_influence/ has a beautiful
> "influence simulator" and quite a few people asked me: "Why don't you
> put
> something like that in Moyo Go"? Well.. Because I'm a computer-Go
> scientist,
> not an artist :-)
>
> I want to make it clear that I have great respect for the programmer
> of this
> stunningly beautiful display, and I'm sure the algorithms used are
> not
> simple at all. Yet, it does not appear to be a genuine influence
> simulator,
> more a kind of "Go art" or perhaps the programmer's own interpretation
> of
> "influence".
>
> Because "influence" in comp. Go terms is not something subjective, or
> something "beautiful", no, we design influence functions and measure
> their
> performance with tens of thousands of end positions of pro games and
> keep
> tuning those functions until they give the most accurate estimation
> of
> territory and influence (which is more subjective). Territory is
> subjective
> as well, but the abberations average eachother out so we can
> establish
> accurate performance parameters for our influence functions.
>
Here you are talking about a function that estimates both territory and
influence for computer programs. Things are getting fuzzier as to
whether you are talking about go or about computer programs. For
instance, the interaction with the rest of the program may matter. In
that case, the best influence function in one program may be different
from the best function in a different program. However, if the influence
function is independent and simply gives an estimate of the value of the
board, then shouldn't it do the same for human players? In that case,
aren't we talking about go?
>
> Take this example of influence maps. (see image on blog)
>
> In computer Go, the sub-problem of "influence" has been quite
> acceptably
> solved by Bruno Bouzy, who has published several papers on his method.
>
I respect Bouzy's work, but I disagree. I suspect that Bouzy would
disagree, as well. His function may be the best one we currently know of
for the current version of GnuGo, but that's a more limited claim.
As I said, you appear to be talking about a program-independent estimate
of the value of the board. And go players know that the strength of
stones and especially their life and death status is a very important
factor in that estimation. A dead stone with no threat to live is worth
one point for the opponent by territory scoring. It's value is -1 (0 by
area scoring). Any independent estimator of value that does not take
that fact into account is far from perfect.
I do not know the details of GnuGo or other programs that use Bouzy's
influence function. (Actually, doesn't he have more than one? I'll still
use the singular here, though.) Perhaps they use it as an input for an
evaluation function. Then it is not independent. I can imagine a program
that used both Bouzy's function and the "influence map" function as
inputs to an evaluation function, among others. Why not? The art of go
programming is still young.
Ciao,
Bill
Vit Brunner
> The keima is better connected than either the elephant jump or the one
> space jump. That's a curious fact that does not fit well with either
> Manhattan or Euclidean distance. Certainly Manhattan distance does not
> have a simple relation to connectivity.
Hi Bill,
while I agree that the elephant jump is not very connected, I can't
agree with you saying that keima is more connected than a one point
jump. While none of them is connected in the way "can't be cut", when
cutting a one point jump you severely strengthen your opponent in the
surroundings, and when cutting keima you don't need to make him
anywhere near strong.
One could also say that it's usual to use keima for attacking (that is
when you don't really care about the cut), and to use one point jump
when you are escaping with something (if the danger of cutting is too
big you might be better of with kosumi, but almost never with a
keima)).
Just my 2c,
Vit
Vit Brunner
> And before anyone tells me that I am "not a Go player", and want to politely
> remind them of the fact that it will be computer scientists that will
> eventually and inevitably make the program that will win from a 9p player.
Just curious - why is it inevitable?
(please note that I'm NOT saying that it won't happen, I just think
that one can't be sure that it's inevitable, at least not yet)
-
Bill Spight <bsp...@pacbell.net> wrote:
> ... All of this suggests that connectivity, and hence influence, is
> In a real game, horizontal influence is usually the strongest,
> because of the interaction with the edge of the board.
Interaction with an edge of the board depends upon proximity
to an edge of the board. Too close to the edge also results
in a "snake house" effect. This ratio for "stronger horizontal
influence" is a dampened exponential, maybe Poisson dist.
Mere "diagonal connectivity" has less ability to form territory,
so it's not much to crow about. How to combine Manhattan
with Euclidean distance? One approach to a Go Pincushion
uses: dist = ((dx)^e + (dy)^e)^(1/e) . All depends on how
such interim calculations are put to use in subsequent math.
Once getting more than 6 or 7 from a wall, it's disconnecting.
This pincushion suggests that beyond a 5,4 interval becomes
disconnected.
> henricb...@gmail.com wrote:
>> I guess that "influence" is bound to remain a vague concept,
>> at the very least it must depend on many local tactical details.
>> The same for these kinds of maps.
> Yes. As I said in another post, I think that eventually the vague notion
> of influence will break down into several more concrete concepts.
There's an unwritten rule concerning deconstructions of Go:
a reassembly should eventually occur to arrive at territory ideas.
- regards
- jb
----------------------------------------------------------
Is President Bush Sane?
http://www.vdare.com/roberts/061201_sane.htm
----------------------------------------------------------
FBI Taps Cell Phone Mic as Eavesdropping Tool
http://www.nationalvanguard.org/story.php?id=10936
----------------------------------------------------------
marc gonzalez-carnicer
to throw an idea. I think an screen saver that drawed these influence
figures on a real go game, could make quite a cool screen saver. I
don't know if there are others besides Arnoud vd Loeff's, but in any
case it should not be difficult to modify.
Frank de Groot ha escrit:
-
jazze...@hotmail.com (-) wrote:
> Interaction with an edge of the board depends upon proximity
> to an edge of the board. Too close to the edge also results
> in a "snake house" effect. This ratio for "stronger horizontal
> influence" is a dampened exponential, maybe Poisson dist.
> Mere "diagonal connectivity" has less ability to form territory,
> so it's not much to crow about. How to combine Manhattan
> with Euclidean distance? One approach to a Go Pincushion
> uses: dist = ((dx)^e + (dy)^e)^(1/e) . All depends on how
> such interim calculations are put to use in subsequent math.
> Once getting more than 6 or 7 from a wall, it's disconnecting.
> This pincushion suggests that beyond a 5,4 interval becomes
> disconnected.
I have been asked to explain this with more elaboration.
A general form for this pincushion distance metric is given by:
Maxima:
n:?exp(1); ROMBERG( (cos(x)^n + sin(x)^n)^(1/n) , x,0,?pi/2);
= 1.5000916878482187 ...
-( AND )-
Pari:
n=exp(1): m=1/n: intnum( x=0,Pi/2,(cos(x)^n+sin(x)^n)^m )
= 1.500091687848216971414238860 ...
These numerical integrations provide a result very close to 3/2
for the positive quadrant, and thereby very close to 6 for the sum
of all four quadrants. A Go Stone should have power of 6 liberties
when considering its "virtual liberties" along diagonals, and/or
off the horizontal or vertical directions of extended-order liberties.
These aren't "move to capture" liberties, but "influence" liberties,
where "influence" may be directly converted (territory) or indirectly
convertible (connections).
For n=1, yields Manhattan Distance with virtual power of 8 (too strong).
For n=2, yields Pythagorean Distance with some virtual power of `2*pi.'
For n=e (natural logarithm) yields here (estimated) virtual power of ~6.
At n --> oo, this form of estimated virtual power converges to Sqrt(32).
When considering how stones must eventually all fit together into the
endgame, virtual power becomes actual power and is thereby an integer.
So Pythagorean Distance cannot be the correct measure. I do not find
close integers for virtual power (influence), other than n=1 and/or n=e,
nor multiples of `pi' & `1/pi'. Due to irrationality for infinity limit
I do not expect to discover any other close integers. All appearances
indicate only further fractal effects.
Perhaps there is connection between this characteristic and qualities of
Fermat's Last Theorem. If having an existing solution in rationals for
x^n + y^n = z^n implies that multiple pairs of x',y' must occur for
some z, then these would need to be indicated linearly in some
fashion, along a constant curvature (straight line or arc). Results
from checking numerical integrations do not indicate much promise
for non-trivial values of `n' (other than n=1 & n=2). There doesn't
seem to be sufficient room in the margin for additional remarks...
- regards
- jb
------------------------------------------------------------------------
Maxima is based on a 1982 version of Macsyma ...
http://en.wikipedia.org/wiki/Maxima
------------------------------------------------------------------------
Max
Possibly. But it won't necessarily be you. And it is irrelevant to the
discussion of the propriety of the influence functions of specific
programs.
I don't know what comp. Go "scientists" agree on and don't really care.
The reflection of influence of waves of merit, given that the sides
have an effect on fighting and thus influence. But it should not be a
perfect mirror. It should be a darkened one. Also, it should be
understood that this is just an approximation until a better one is
found or an it can be proven or demonstrated that is is an accurate
approach.