T.R | Title | User | Personal Name | Date | Lines |
---|
732.1 | | KIRK::KOLKER | Conan the Librarian | Wed Jul 15 1987 13:28 | 13 |
|
reply .0
I came up with the barycenter of the points, however I can prove it only
in the case that the points are arrange symmetrically. The problem is
to minimize the mean distance as opposed to the mean squared distance to the
N points. In the latter case, a direct computation with partial derivatives
show a solution at the barycenter of the points.
Do you have a configuration of points where the mean distance minimizer and
the mean squared distance minimizer are different?
|
732.2 | Minimal distance for equipment repairs | 51528::OLAV | Use VAX Ada! | Wed Jul 15 1987 13:55 | 5 |
| The N points are actually location of equipment at customer sites.
I need an algoritm to compute the optimal position (minimum mean
distance to equipment), for a FS office.
Olav
|
732.3 | a quick and dirty estimate | EAGLE1::BEST | R D Best, Systems architecture, I/O | Wed Jul 15 1987 15:19 | 36 |
|
> The N points are actually location of equipment at customer sites.
> I need an algoritm to compute the optimal position (minimum mean
> distance to equipment), for a FS office.
Before I take a shot this, you should note that unless you
can fly straight to a customer site, this is not really a problem in
geometry, but one in graph theory. Remember that you have to get to
the customer site by roads. You might also want to build the FS office near a
road :-). You might also want to think about weighting the sites by the amount
of equipment there, since you will probably be making more frequent trips
to those sites with more equipment (given the dubious assumption that
the reliability of all our equipment is the same) or those sites with
the least reliable equipment.
Having noted these, here's a cheap, dirty, and probably inaccurate
way to get a first shot at an appropriate site.
I assume that these sites are close enough that Euclidean plane geometry will
suffice. The earth's surface is not really flat, but over small areas, flatness
is a reasonable approximation, especially for lazy engineers like myself.
Express each of the site positions as an (x,y) vector from an arbitrarily
chosen reference position (a convenient choice of reference position is
probably one of the customer sites) and arbitrarily oriented orthogonal
axes. Let V(n) = (x(n), y(n)). Form:
Sum( 1, n, V )/n
where Sum is the real sum from the first integer parameter to the second of the
real vector in the third. This result is a vector to the mean position.
All the sites are weighted equally.
I'm not sure if this is the same as the point having the lowest mean
distance to any of the sites, but it seems plausible. Does anyone have
a proof/disproof ?
|
732.4 | | CLT::GILBERT | Builder | Wed Jul 15 1987 19:12 | 65 |
732.5 | Peter's idea done the lazy way | SQM::HALLYB | Like a breath of fresh water... | Thu Jul 16 1987 17:33 | 27 |
732.6 | | CLT::GILBERT | Builder | Fri Jul 17 1987 00:02 | 14 |
732.7 | Is this NP-complete? | AKQJ10::YARBROUGH | Why is computing so labor intensive? | Fri Jul 17 1987 12:20 | 4 |
| The general solution is probably best found by relaxation methods, in any
case. The reality of city traffic patterns is such that 'taxicab' geometry may
be quite valuable... I don't think we are going to be able to find a solution
that is both realistic and mathematically interesting!
|
732.8 | Grist for grad students | SQM::HALLYB | Like a breath of fresh water... | Fri Jul 17 1987 13:56 | 12 |
| re: .6 - Peter, I yield to the reality of hard numbers, but now wonder
where is the error in the mathematics. After all, it should be
nothing more than your work minus the square roots. Of course,
my calculus was over 20 years ago...
Is there a reasonable worst-case error bound on the arithmetic mean?
Assuming some maximum distance M between points, or total distance T,
or some similar criterion. If we want to provide a real-world answer,
is the arithmetic mean good enough? If not, how fast does iteration
converge?
John
|
732.9 | | KIRK::KOLKER | Conan the Librarian | Fri Jul 17 1987 14:41 | 9 |
| re .5
A function is monotonic iff it is order preserving. It needs a domain
that is ordered and a co-domain that is ordered. A function defined
on a region in an N-space is not monotone since the N-space is not
ordered. I think that is the cause of the disonence between your
result and the counter-example.
|
732.10 | | CLT::GILBERT | Builder | Fri Jul 17 1987 17:05 | 3 |
| Consider the one-dimensional case. The two proposed solutions yield
the mean and the median (or, if the number of points is even, any point
between the two center points). The median gives the correct solution.
|
732.11 | It was wrong this way | SQM::HALLYB | Like a breath of fresh water... | Fri Jul 17 1987 17:28 | 11 |
| Minimzing
/ N 2 \
sqrt ( Sum Z )
\ i=1 i /
does not necessarily minimize
N / 2 \
Sum sqrt( Z )
i=1 \ i /
|
732.12 | | KIRK::KOLKER | Conan the Librarian | Fri Jul 17 1987 18:23 | 11 |
| re priors
Is there any physical process which is an analogue for this problem?
I saw an article in Scientific American where a number of NP complete
problems could be represented physically and "solved" in quick time.
I will try to remember the year and the month of the issue and post
it somewhere.
|
732.13 | Solution from Nature | SEMI::NG | | Fri Jul 17 1987 19:19 | 21 |
| Re .12:
I think I remember coming across that article too, but I can't
remember what issue it was.
Solution in similiar line of that article may be as follows:
The N points are N fixed pins and the optimal point corresponds
to a floating pin. Then somehow attach the floating pin with all
the N fixed pins. Maybe by springs or rubber-bands. After the system
goes into equilibrium, the position of the floating pin will be
the optimal point.
However, this will yield a different answer from .5.
In my case, the system will minimize the total potential energy
which is kind of related to the square of the distance between 2
points for a linear spring and not just the distance. Hence, the
resulting solution actually minimizes the sum of the squares of
the distances instead of just the sum of the distances.
David Ng
|
732.14 | Something more to say | SEMI::NG | | Fri Jul 17 1987 19:29 | 12 |
| I agree with what .3 said. This is more of a problem in graph theory
than in geometry.
About whether this a NP-complete or not, I am not sure if you want
to treat it as a graph theory problem. If the solution of .5 is
what you want, that is definitely not a NP-complete problem since
.5 is not a combinatorial algorithm and I think the word "NP-complete"
only describes combinatorial problem. (??)
David Ng
|
732.15 | | CLT::GILBERT | Builder | Sat Jul 18 1987 05:18 | 5 |
| re .13
If you use a table with holes in it and hang equal weights, ..., then
this will yield the solution that minimizes the sum of the distances,
since the potential energy of the system is minimized, and that implies
that the lengths of the strings beneath the table is maximized.
|
732.16 | | TLE::BRETT | | Sun Jul 19 1987 01:12 | 3 |
| Actually, you can only be sure it is a local minimum,...
/Bevin
|
732.17 | | ENGINE::ROTH | | Mon Jul 20 1987 11:19 | 7 |
| What is more, in even the simplest case of two points any point on
the line segment between them meets the criteria. With many points
all sorts of things can happen.
Off hand I don't know an easy way of obtaining the global minimum.
- Jim
|