Difference between revisions of "User:Zeracles/Starmap/Minimal Spanning Tree"

From Ultronomicon
Jump to navigation Jump to search
(first version)
 
m (→‎Code: adjust links)
 
(9 intermediate revisions by the same user not shown)
Line 1: Line 1:
−
It definitely needs some improvement, but for now -
+
I originally coded up [[wikipedia:Prim's_algorithm|Jarnik/Prim's algorithm]]. In the process of optimising it, I unwittingly reinvented [[wikipedia:Bor%C5%AFvka%27s_algorithm|Boruvka's algorithm]], which handles large numbers of input points better, at least the way I've done it. Demonstration [http://www.star-control.com/forum/index.php/topic,1660.msg35209.html#msg35209 here].
−
% minimal spannng tree in matlab
+
==Minimally Connected Opaque Friend Graph==
−
% by Anthony Smith (Zeracles)
+
The matlab/octave code below for Boruvka's algorithm also contains an option to create a Minimally Connected Opaque Friend Graph (MCOFG). This is a graph I invented for the purpose of adding links to a completed MST to get closed loops, which were present in the original Star Control 1 starmaps. Actually, the MCOFG has another very interesting use in my real life work, but that's another story.
−
% astroant@hotmail.com
+
 
−
load('inputpositions')
+
The basic properties of the graph are
−
+
*all the points are connected
−
remainingpositions=inputpositions;
+
*each point is directly linked to ``nearby" points
−
+
*no individual link occupies a path in space similar to that traced by other link(s)
−
% detect input properties and set up index
+
It is ``minimal" because only close neighbours are connected, and with ``opaque friends" because links can't go ``through" nearby points (the ``friends"). Graph geeks will realise that the Delaunay Tessellation (which we also considered) seems to satisfy these criteria, but it has edge effects; it will produce a convex hull at the edges, linking points that may be nowhere near each other, meaning that it only partially satisfies the ``nearby" criterion. The MCOFG is also far easier to implement than the delaunay tessellation! The MST is a good starting point in the construction of this graph, because it guarantees the connection of all points. It is also used to define ``nearby" and thus ``friends", by using the length of the longest MST link hosted at any given node to define a radius within which links to additional points will be considered.
−
sizeremainingpositions=size(remainingpositions);
+
 
−
lengthremainingpositions=sizeremainingpositions(1);
+
So, to generate the graph, we first generate an MST. For each point, we examine its MST links, and remember the length of the longest one. Points within this distance around the point are considered local, and potential friends. To determine whether each of these potential friends should be linked to the point, we consider all points that are closer. If r_2 is the distance to the point we are considering linking, and r_1 is a point closer to the point we are finding friends for, and \theta is the angle between vectors '''r_2''' and '''r_1''', we compute the product
−
dimensions=sizeremainingpositions(2);
+
<br>((r_2)/(2r_1))*((\pi/2)/(\theta))
−
+
<br>If it's less than one, the link '''r_2''' is allowed. When there are many points closer to consider, the products for each of them are multiplied together, and if the result if less than one, the link is allowed. This approach is motivated by the third condition for this graph, that no individual link occupies a path in space similar to that traced by other link(s). If one has points 1, 2, and 3, in that order on a straight line, 1 should not be linked to 3, because this duplicates the paths 1 to 2 and 2 to 3. With our approach, 1 will not be linked to 3 because 2 is closer to one, and the angle \theta is zero, so the product will be very much greater then one.
−
index=transpose(linspace(1,lengthremainingpositions,lengthremainingpositions));
+
 
−
+
The motivation for the inclusion of the distance component (r_2/r_1) is that if two points subtend a small angle \theta, but are at similar distances, the link to the more distant one shouldn't be broken. As for normalisation, the factors of two in the angular and distance components are there so that we can actually get values of less than one.
−
remainingpositionxs=remainingpositions(:,1);
+
 
−
remainingpositionys=remainingpositions(:,2);
+
There are also two numbers that can be tweaked to vary the connectedness of the graph.
−
if dimensions>2
+
The product is divided by ``friendliness" f, i.e.
−
remainingpositionzs=remainingpositions(:,3);
+
<br>((r_2)/(2r_1))*((\pi/2)/(\theta))*(1/f)
−
end
+
<br>to vary the number of links. Higher values of f lead to more connections, more friends. The locality factor l influences the radius within which points are considered for friendship, which is by default the length of the longest local MST link. l = 2, say, would double this.
−
+
 
−
% initialise search parameters
+
==Efficiency Method==
−
minx=min(remainingpositionxs);
+
In discussion on the forum Death 999 contributed a method that has become our preferred method for closed loop generation - I'll insert a description when I get a moment. It is implemented in both versions of the code available below.
−
maxx=max(remainingpositionxs);
+
 
−
xextent=minx*maxx;
+
==Code==
−
+
When prototyping, I wrote a [http://www.physics.usyd.edu.au/~asmith/files/sc1/MSTclosure.m matlab/octave version available here], and have recently done a [http://www.physics.usyd.edu.au/~asmith/files/sc1/MSTclosure.c C conversion available here].
−
miny=min(remainingpositionys);
 
−
maxy=max(remainingpositionys);
 
−
yextent=miny*maxy;
 
−
 
−
if dimensions>2
 
−
minz=min(remainingpositionzs);
 
−
maxz=max(remainingpositionzs);
 
−
zextent=minz*maxz;
 
−
end
 
−
 
−
if dimensions==2
 
−
measure=xextent*yextent;
 
−
localmeasure=measure/lengthremainingpositions;
 
−
searchincrement=((10*localmeasure)/pi)^(1/2);
 
−
end
 
−
if dimensions==3
 
−
measure=xextent*yextent*zextent;
 
−
localmeasure=measure/lengthremainingpositions;
 
−
searchincrement=((10*localmeasure)/(4/3)/pi)^(1/3);
 
−
end
 
−
 
−
% initialise with a seed point
 
−
seedindex=1;
 
−
maskvector=zeros(lengthremainingpositions,1);
 
−
maskvector(seedindex)=1;
 
−
 
−
remainingpositionxs=remainingpositionxs(find(maskvector==0));
 
−
remainingpositionys=remainingpositionys(find(maskvector==0));
 
−
if dimensions>2
 
−
remainingpositionzs=remainingpositionzs(find(maskvector==0));
 
−
end
 
−
remainingpositionindices=index(find(maskvector==0));
 
−
 
−
lengthremainingpositions=lengthremainingpositions-1;
 
−
 
−
treeindices=index(seedindex,:);
 
−
lengthtreepositions=1;
 
−
 
−
branches=[];
 
−
 
−
while lengthremainingpositions>0
 
−
mindistances=[];
 
−
for i=linspace(1,lengthtreepositions,lengthtreepositions)
 
−
treepositionx=inputpositions(treeindices(i),1);
 
−
treepositiony=inputpositions(treeindices(i),2);
 
−
if dimensions>2
 
−
treepositionz=inputpositions(treeindices(i),3);
 
−
end
 
−
treeindex=treeindices(i);
 
−
 
−
occupiedbox=0;
 
−
searchfactor=1;
 
−
while occupiedbox==0
 
−
boxradius=searchfactor*searchincrement;
 
−
 
−
boxminx=treepositionx-boxradius;
 
−
boxmaxx=treepositionx+boxradius;
 
−
boxminy=treepositiony-boxradius;
 
−
boxmaxy=treepositiony+boxradius;
 
−
if dimensions>2
 
−
boxminz=treepositionz-boxradius;
 
−
boxmaxz=treepositionz+boxradius;
 
−
end
 
−
 
−
if dimensions==2
 
−
closerremainingmask=(boxminx<=remainingpositionxs).*(remainingpositionxs<=boxmaxx).*(boxminy<=remainingpositionys).*(remainingpositionys<=boxmaxy);
 
−
end
 
−
if dimensions==3
 
−
closerremainingmask=(boxminx<=remainingpositionxs).*(remainingpositionxs<=boxmaxx).*(boxminy<=remainingpositionys).*(remainingpositionys<=boxmaxy).*(boxminz<=remainingpositionzs).*(remainingpositionzs<=boxmaxz);
 
−
end
 
−
 
−
mindistancecandidates=[];
 
−
mindistancecandidateindices=[];
 
−
if sum(closerremainingmask)>0
 
−
for k=find(closerremainingmask==1)
 
−
branchcandidatex=remainingpositionxs(k);
 
−
branchcandidatey=remainingpositionys(k);
 
−
if dimensions>2
 
−
branchcandidatez=remainingpositionzs(k);
 
−
end
 
−
branchcandidateindex=remainingpositionindices(k);
 
−
 
−
if dimensions==2
 
−
mindistancecandidate=((branchcandidatex-treepositionx).^2+(branchcandidatey-treepositiony).^2).^(1/2);
 
−
end
 
−
if dimensions==3
 
−
mindistancecandidate=((branchcandidatex-treepositionx).^2+(branchcandidatey-treepositiony).^2+(branchcandidatez-treepositionz).^2).^(1/2);
 
−
end
 
−
if mindistancecandidate<=boxradius
 
−
mindistancecandidates=[mindistancecandidates;mindistancecandidate];
 
−
mindistancecandidateindices=[mindistancecandidateindices;branchcandidateindex];
 
−
end
 
−
end
 
−
end
 
−
if numel(mindistancecandidates)>0
 
−
mindistance=[min(mindistancecandidates),treeindex,mindistancecandidateindices(find(mindistancecandidates==min(mindistancecandidates)))];
 
−
mindistances=[mindistances;mindistance];
 
−
 
−
occupiedbox=1;
 
−
end
 
−
searchfactor=searchfactor+1;
 
−
end
 
−
end
 
−
branchdistance=min(mindistances(:,1));
 
−
branchindex=find(branchdistance==mindistances(:,1));
 
−
numberbranchcandidates=numel(branchindex);
 
−
if numberbranchcandidates>1
 
−
branchindex=branchindex(ceil(numberbranchcandidates*rand(1)));
 
−
end
 
−
branch=mindistances(branchindex,:);
 
−
branches=[branches;branch];
 
−
 
−
treeindices=[treeindices;branch(3)];
 
−
lengthtreepositions=lengthtreepositions+1;
 
−
 
−
remainingmask=ones(lengthremainingpositions,1);
 
−
removed=find(remainingpositionindices==branch(3));
 
−
remainingmask(removed)=0;
 
−
 
−
remainingpositionindices=remainingpositionindices(find(remainingmask));
 
−
remainingpositionxs=remainingpositionxs(find(remainingmask));
 
−
remainingpositionys=remainingpositionys(find(remainingmask));
 
−
if dimensions>2
 
−
remainingpositionzs=remainingpositionzs(find(remainingmask));
 
−
end
 
−
 
−
lengthremainingpositions=lengthremainingpositions-1;
 
−
end
 
−
 
−
save('branches','branches')
 
−
 
−
sizebranches=size(branches);
 
−
numberofbranches=sizebranches(1);
 
−
 
−
figure
 
−
hold on
 
−
box on
 
−
plot(inputpositions(:,1),inputpositions(:,2),'k.')
 
−
 
−
for i=linspace(1,numberofbranches,numberofbranches)
 
−
startindex=branches(i,2);
 
−
finishindex=branches(i,3);
 
−
 
−
startx=inputpositions(startindex,1);
 
−
starty=inputpositions(startindex,2);
 
−
 
−
finishx=inputpositions(finishindex,1);
 
−
finishy=inputpositions(finishindex,2);
 
−
 
−
line([startx;finishx],[starty;finishy],'Color','k')
 
−
end
 

Latest revision as of 06:22, 1 September 2013

I originally coded up Jarnik/Prim's algorithm. In the process of optimising it, I unwittingly reinvented Boruvka's algorithm, which handles large numbers of input points better, at least the way I've done it. Demonstration here.

Minimally Connected Opaque Friend Graph[edit]

The matlab/octave code below for Boruvka's algorithm also contains an option to create a Minimally Connected Opaque Friend Graph (MCOFG). This is a graph I invented for the purpose of adding links to a completed MST to get closed loops, which were present in the original Star Control 1 starmaps. Actually, the MCOFG has another very interesting use in my real life work, but that's another story.

The basic properties of the graph are

  • all the points are connected
  • each point is directly linked to ``nearby" points
  • no individual link occupies a path in space similar to that traced by other link(s)

It is ``minimal" because only close neighbours are connected, and with ``opaque friends" because links can't go ``through" nearby points (the ``friends"). Graph geeks will realise that the Delaunay Tessellation (which we also considered) seems to satisfy these criteria, but it has edge effects; it will produce a convex hull at the edges, linking points that may be nowhere near each other, meaning that it only partially satisfies the ``nearby" criterion. The MCOFG is also far easier to implement than the delaunay tessellation! The MST is a good starting point in the construction of this graph, because it guarantees the connection of all points. It is also used to define ``nearby" and thus ``friends", by using the length of the longest MST link hosted at any given node to define a radius within which links to additional points will be considered.

So, to generate the graph, we first generate an MST. For each point, we examine its MST links, and remember the length of the longest one. Points within this distance around the point are considered local, and potential friends. To determine whether each of these potential friends should be linked to the point, we consider all points that are closer. If r_2 is the distance to the point we are considering linking, and r_1 is a point closer to the point we are finding friends for, and \theta is the angle between vectors r_2 and r_1, we compute the product
((r_2)/(2r_1))*((\pi/2)/(\theta))
If it's less than one, the link r_2 is allowed. When there are many points closer to consider, the products for each of them are multiplied together, and if the result if less than one, the link is allowed. This approach is motivated by the third condition for this graph, that no individual link occupies a path in space similar to that traced by other link(s). If one has points 1, 2, and 3, in that order on a straight line, 1 should not be linked to 3, because this duplicates the paths 1 to 2 and 2 to 3. With our approach, 1 will not be linked to 3 because 2 is closer to one, and the angle \theta is zero, so the product will be very much greater then one.

The motivation for the inclusion of the distance component (r_2/r_1) is that if two points subtend a small angle \theta, but are at similar distances, the link to the more distant one shouldn't be broken. As for normalisation, the factors of two in the angular and distance components are there so that we can actually get values of less than one.

There are also two numbers that can be tweaked to vary the connectedness of the graph. The product is divided by ``friendliness" f, i.e.
((r_2)/(2r_1))*((\pi/2)/(\theta))*(1/f)
to vary the number of links. Higher values of f lead to more connections, more friends. The locality factor l influences the radius within which points are considered for friendship, which is by default the length of the longest local MST link. l = 2, say, would double this.

Efficiency Method[edit]

In discussion on the forum Death 999 contributed a method that has become our preferred method for closed loop generation - I'll insert a description when I get a moment. It is implemented in both versions of the code available below.

Code[edit]

When prototyping, I wrote a matlab/octave version available here, and have recently done a C conversion available here.

Retrieved from "https://wiki.uqm.stack.nl/index.php?title=User:Zeracles/Starmap/Minimal_Spanning_Tree&oldid=27413"