Class LevenbergMarquardtOptimizer

java.lang.Object
org.hipparchus.optim.nonlinear.vector.leastsquares.LevenbergMarquardtOptimizer
All Implemented Interfaces:
LeastSquaresOptimizer

public class LevenbergMarquardtOptimizer extends Object implements LeastSquaresOptimizer
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.
  • Constructor Details

    • LevenbergMarquardtOptimizer

      public LevenbergMarquardtOptimizer()
      Default constructor.

      The default values for the algorithm settings are:

      • Initial step bound factor: 100
      • Cost relative tolerance: 1e-10
      • Parameters relative tolerance: 1e-10
      • Orthogonality tolerance: 1e-10
      • QR ranking threshold: Precision.SAFE_MIN
    • LevenbergMarquardtOptimizer

      public LevenbergMarquardtOptimizer(double initialStepBoundFactor, double costRelativeTolerance, double parRelativeTolerance, double orthoTolerance, double qrRankingThreshold)
      Construct an instance with all parameters specified.
      Parameters:
      initialStepBoundFactor - initial step bound factor
      costRelativeTolerance - cost relative tolerance
      parRelativeTolerance - parameters relative tolerance
      orthoTolerance - orthogonality tolerance
      qrRankingThreshold - threshold in the QR decomposition. Columns with a 2 norm less than this threshold are considered to be all 0s.
  • Method Details

    • withInitialStepBoundFactor

      public LevenbergMarquardtOptimizer withInitialStepBoundFactor(double newInitialStepBoundFactor)
      Build new instance with initial step bound factor.
      Parameters:
      newInitialStepBoundFactor - 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 non-zero, or else to newInitialStepBoundFactor itself. In most cases factor should lie in the interval (0.1, 100.0). 100 is a generally recommended value. of the matrix is reduced.
      Returns:
      a new instance.
    • withCostRelativeTolerance

      public LevenbergMarquardtOptimizer withCostRelativeTolerance(double newCostRelativeTolerance)
      Build new instance with cost relative tolerance.
      Parameters:
      newCostRelativeTolerance - Desired relative error in the sum of squares.
      Returns:
      a new instance.
    • withParameterRelativeTolerance

      public LevenbergMarquardtOptimizer withParameterRelativeTolerance(double newParRelativeTolerance)
      Build new instance with parameter relative tolerance.
      Parameters:
      newParRelativeTolerance - Desired relative error in the approximate solution parameters.
      Returns:
      a new instance.
    • withOrthoTolerance

      public LevenbergMarquardtOptimizer withOrthoTolerance(double newOrthoTolerance)
      Build new instance with ortho tolerance.
      Parameters:
      newOrthoTolerance - Desired max cosine on the orthogonality between the function vector and the columns of the Jacobian.
      Returns:
      a new instance.
    • withRankingThreshold

      public LevenbergMarquardtOptimizer withRankingThreshold(double newQRRankingThreshold)
      Build new instance with ranking threshold.
      Parameters:
      newQRRankingThreshold - 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.
      Returns:
      a new instance.
    • getInitialStepBoundFactor

      public double getInitialStepBoundFactor()
      Gets the value of a tuning parameter.
      Returns:
      the parameter's value.
      See Also:
    • getCostRelativeTolerance

      public double getCostRelativeTolerance()
      Gets the value of a tuning parameter.
      Returns:
      the parameter's value.
      See Also:
    • getParameterRelativeTolerance

      public double getParameterRelativeTolerance()
      Gets the value of a tuning parameter.
      Returns:
      the parameter's value.
      See Also:
    • getOrthoTolerance

      public double getOrthoTolerance()
      Gets the value of a tuning parameter.
      Returns:
      the parameter's value.
      See Also:
    • getRankingThreshold

      public double getRankingThreshold()
      Gets the value of a tuning parameter.
      Returns:
      the parameter's value.
      See Also:
    • optimize

      Solve the non-linear least squares problem.
      Specified by:
      optimize in interface LeastSquaresOptimizer
      Parameters:
      problem - the problem definition, including model function and convergence criteria.
      Returns:
      The optimum.