Hi.
On Fri, 01 Aug 2014 18:29:33 +0200, Benjamin Eltzner wrote:
Dear mailing list,
I would like to do a Levenberg-Marquardt least mean square fit (i. e.
min( |f(x)|^2 ) for f: R^d -> R^n, d in [2,8], n in [~50, ~10000])
using
org.apache.commons.math3.fitting.leastsquares.LevenbergMarquardtOptimizer
and
I am somewhat confused by the relevant API.
tl/dr: Is there a way to do this in a simple way, providing only a
function f, its derivative function df and an initial guess for x?
Long form: it seems that I have to create a LeastSquaresProblem which
requires a lot of input
* MultivariateVectorFunction model: done
* MultivariateMatrixFunction jacobian: done
* double[] observed: [0,...,0] in my case, since i want least mean
squares, right?
Not necessarily. The algorithm will look for the set of parameters
x = ( x_1, x_2, ..., x_d )
for which f(x) = c_1, c_2, ..., c_n is closest to the "observed"
values o_1, o_2, ..., o_n.
* double[] start: done
* RealMatrix weight: should this be a "new DiagonalMatrix([1,...,1])"
of
the dimension of "observed"?
A diagonal matrix of unit weights is the default.
* ConvergenceChecker<LeastSquaresProblem.Evaluation> checker: no clue
It's a user-defined check for convergence.
But you can set it to "null": the standard Levenberg-Marquardt
procedure
for checking convergence will then be used.
* int maxEvaluations: no clue
* int maxIterations: no clue
These are intended to avoid running "forever": "maxEvaluations" sets a
limit to the number of evaluations of f(x), "maxIterations" limits the
number of iteration of the optimizer's main loop.
If you are confident that the problem is well behaved, you can set them
to "Integer.MAX_VALUE".
Is there a way to get "reasonable defaults" for the three or four
last
arguments? Or put differently, what would be "typical" inputs for the
last four arguments?
I would be very glad, if you could help me with this problem. I have
been trying to figure out, what would be sensible inputs be reading
the
source code, but I guess it would take ages until I really understood
what is going on there. The results of a trial-and-error approach
have
also been rather disappointing so far.
The user guide should be updated; sorry for that.
Best regards,
Gilles
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
