mpicbg.imglib.algorithm.peak

Class LevenbergMarquardtSolver

java.lang.Object

mpicbg.imglib.algorithm.peak.LevenbergMarquardtSolver

public class LevenbergMarquardtSolver
extends Object

A collection of static utils implementing a plain Levenberg-Marquardt least-square curve fitting algorithm.

It was adapted and stripped from jplewis (www.idiom.com/~zilla) and released under the GPL. There are various small tweaks for robustness and speed, mainly a first step to derive a crude estimate, based on maximum-likelihood analytic formulae.

Author:

Jean-Yves Tinevez 2011, 2012

Constructor Summary

Constructors

Constructor and Description
LevenbergMarquardtSolver()

Method Summary

Modifier and Type	Method and Description
static double	chiSquared(double[][] x, double[] a, double[] y, FitFunction f) Calculate the current sum-squared-error
static int	solve(double[][] x, double[] a, double[] y, FitFunction f, double lambda, double termepsilon, int maxiter) Minimize E = sum {(y[k] - f(x[k],a)) }^2 Note that function implements the value and gradient of f(x,a), NOT the value and gradient of E with respect to a!

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

LevenbergMarquardtSolver

public LevenbergMarquardtSolver()

Method Detail

chiSquared

public static final double chiSquared(double[][] x,
                                      double[] a,
                                      double[] y,
                                      FitFunction f)

Calculate the current sum-squared-error

solve

public static final int solve(double[][] x,
                              double[] a,
                              double[] y,
                              FitFunction f,
                              double lambda,
                              double termepsilon,
                              int maxiter)
                       throws Exception

Minimize E = sum {(y[k] - f(x[k],a)) }^2 Note that function implements the value and gradient of f(x,a), NOT the value and gradient of E with respect to a!

Parameters:

x - array of domain points, each may be multidimensional

y - corresponding array of values

a - the parameters/state of the model

lambda - blend between steepest descent (lambda high) and jump to bottom of quadratic (lambda zero). Start with 0.001.

termepsilon - termination accuracy (0.01)

maxiter - stop and return after this many iterations if not done

Returns:

the number of iteration used by minimization

Throws:

Exception