Class LevenbergMarquardtOptimizer

  • All Implemented Interfaces:
    DifferentiableMultivariateVectorialOptimizer

    public class LevenbergMarquardtOptimizer
    extends AbstractLeastSquaresOptimizer
    This class solves a least squares problem using the Levenberg-Marquardt algorithm.

    This implementation should work even for over-determined systems (i.e. systems having more point than equations). Over-determined systems are solved by ignoring the point which have the smallest impact according to their jacobian column norm. Only the rank of the matrix and some loop bounds are changed to implement this.

    The resolution engine is a simple translation of the MINPACK lmder routine with minor changes. The changes include the over-determined resolution, the use of inherited convergence checker and the Q.R. decomposition which has been rewritten following the algorithm described in the P. Lascaux and R. Theodor book Analyse numérique matricielle appliquée à l'art de l'ingénieur, Masson 1986.

    The authors of the original fortran version are:

    • Argonne National Laboratory. MINPACK project. March 1980
    • Burton S. Garbow
    • Kenneth E. Hillstrom
    • Jorge J. More
    The redistribution policy for MINPACK is available here, for convenience, it is reproduced below.

    Minpack Copyright Notice (1999) University of Chicago. All rights reserved
    Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
    1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
    2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
    3. The end-user documentation included with the redistribution, if any, must include the following acknowledgment: This product includes software developed by the University of Chicago, as Operator of Argonne National Laboratory. Alternately, this acknowledgment may appear in the software itself, if and wherever such third-party acknowledgments normally appear.
    4. WARRANTY DISCLAIMER. THE SOFTWARE IS SUPPLIED "AS IS" WITHOUT WARRANTY OF ANY KIND. THE COPYRIGHT HOLDER, THE UNITED STATES, THE UNITED STATES DEPARTMENT OF ENERGY, AND THEIR EMPLOYEES: (1) DISCLAIM ANY WARRANTIES, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY IMPLIED WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE OR NON-INFRINGEMENT, (2) DO NOT ASSUME ANY LEGAL LIABILITY OR RESPONSIBILITY FOR THE ACCURACY, COMPLETENESS, OR USEFULNESS OF THE SOFTWARE, (3) DO NOT REPRESENT THAT USE OF THE SOFTWARE WOULD NOT INFRINGE PRIVATELY OWNED RIGHTS, (4) DO NOT WARRANT THAT THE SOFTWARE WILL FUNCTION UNINTERRUPTED, THAT IT IS ERROR-FREE OR THAT ANY ERRORS WILL BE CORRECTED.
    5. LIMITATION OF LIABILITY. IN NO EVENT WILL THE COPYRIGHT HOLDER, THE UNITED STATES, THE UNITED STATES DEPARTMENT OF ENERGY, OR THEIR EMPLOYEES: BE LIABLE FOR ANY INDIRECT, INCIDENTAL, CONSEQUENTIAL, SPECIAL OR PUNITIVE DAMAGES OF ANY KIND OR NATURE, INCLUDING BUT NOT LIMITED TO LOSS OF PROFITS OR LOSS OF DATA, FOR ANY REASON WHATSOEVER, WHETHER SUCH LIABILITY IS ASSERTED ON THE BASIS OF CONTRACT, TORT (INCLUDING NEGLIGENCE OR STRICT LIABILITY), OR OTHERWISE, EVEN IF ANY OF SAID PARTIES HAS BEEN WARNED OF THE POSSIBILITY OF SUCH LOSS OR DAMAGES.
      Since:
      2.0
      • Method Detail

        • setInitialStepBoundFactor

          public void setInitialStepBoundFactor​(double initialStepBoundFactor)
          Set the positive input variable used in determining the initial step bound. This bound is set to the product of initialStepBoundFactor and the euclidean norm of diag*x if nonzero, or else to initialStepBoundFactor itself. In most cases factor should lie in the interval (0.1, 100.0). 100.0 is a generally recommended value.
          Parameters:
          initialStepBoundFactor - initial step bound factor
        • setCostRelativeTolerance

          public void setCostRelativeTolerance​(double costRelativeTolerance)
          Set the desired relative error in the sum of squares.

          This setting is used only if the vectorial convergence checker is set to null.

          Parameters:
          costRelativeTolerance - desired relative error in the sum of squares
        • setParRelativeTolerance

          public void setParRelativeTolerance​(double parRelativeTolerance)
          Set the desired relative error in the approximate solution parameters.

          This setting is used only if the vectorial convergence checker is set to null.

          Parameters:
          parRelativeTolerance - desired relative error in the approximate solution parameters
        • setOrthoTolerance

          public void setOrthoTolerance​(double orthoTolerance)
          Set the desired max cosine on the orthogonality.

          This setting is always used, regardless of the vectorial convergence checker being null or non-null.

          Parameters:
          orthoTolerance - desired max cosine on the orthogonality between the function vector and the columns of the jacobian
        • setQRRankingThreshold

          public void setQRRankingThreshold​(double threshold)
          Set the desired threshold for QR ranking.

          If the squared norm of a column vector is smaller or equal to this threshold during QR decomposition, it is considered to be a zero vector and hence the rank of the matrix is reduced.

          Parameters:
          threshold - threshold for QR ranking
          Since:
          2.2