org.ddogleg.optimization.impl

Class LevenbergMarquardtDampened

java.lang.Object

org.ddogleg.optimization.impl.LevenbergBase

org.ddogleg.optimization.impl.LevenbergDenseBase

org.ddogleg.optimization.impl.LevenbergMarquardtDampened

public class LevenbergMarquardtDampened
extends LevenbergDenseBase

Modification of LevenbergDampened which incorporates the insight of Marquardt. The insight was to use the function's curvature information to increase dampening along directions with a larger gradient. In practice this method seems to do better on nearly singular systems.

The step 'x' is computed using the following formula: [J(k)'*J(k) + μ*diag(J(k)'*J(k))]x = -g = -J'*f
where J is the Jacobian, μ is the damping coefficient, g is the gradient, f is the functions output.

The linear solver it uses is specified in the constructor. Cholesky based solver will be the fastest but can fail if the J(x)'J(x) matrix is nearly singular. In those situations a pseudo inverse type solver should be used, which is immune to that type of problem but much more expensive.

Constructor Summary

Constructors

Constructor and Description
LevenbergMarquardtDampened(org.ejml.interfaces.linsol.LinearSolver<org.ejml.data.DenseMatrix64F> solver, double initialDampParam) Specifies termination condition and linear solver.

Method Summary

Methods inherited from class org.ddogleg.optimization.impl.LevenbergDenseBase

setFunction

Methods inherited from class org.ddogleg.optimization.impl.LevenbergBase

addToMessage, getFnorm, getIterationCount, getMessage, getParameters, initialize, isConverged, isUpdatedParameters, iterate, setConvergence, setInitialDampParam

Methods inherited from class java.lang.Object

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

Constructor Detail

LevenbergMarquardtDampened

public LevenbergMarquardtDampened(org.ejml.interfaces.linsol.LinearSolver<org.ejml.data.DenseMatrix64F> solver,
                                  double initialDampParam)

Specifies termination condition and linear solver. Selection of the linear solver an effect speed and robustness.

Parameters:

solver - Linear solver. Cholesky or pseudo-inverse are recommended.

initialDampParam - Initial value of the dampening parameter. Tune.. try 1e-3;