hex.optimization

Class L_BFGS.GradientSolver

java.lang.Object

hex.optimization.L_BFGS.GradientSolver

Direct Known Subclasses:

GLM.GLMGradientSolver, GLM.ProximalGradientSolver

Enclosing class:

L_BFGS

public abstract static class L_BFGS.GradientSolver
extends java.lang.Object

Provides gradient computation and line search evaluation specific to given problem. Typically just a wrapper around MRTask calls.

Constructor Summary

Constructors

Constructor and Description
L_BFGS.GradientSolver()

Method Summary

Methods

Modifier and Type	Method and Description
L_BFGS.LineSearchSol	doLineSearch(L_BFGS.GradientInfo ginfo, double[] beta, double[] direction, int nSteps, double tdec) Perform line search at given solution and search direction.
abstract L_BFGS.GradientInfo	getGradient(double[] beta) Evaluate gradient at solution beta.
abstract double[]	getObjVals(double[] beta, double[] pk, int nSteps, double initialStep, double stepDec) Evaluate objective values at k line search points beta_k.

Methods inherited from class java.lang.Object

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

Constructor Detail

L_BFGS.GradientSolver

public L_BFGS.GradientSolver()

Method Detail

getGradient

public abstract L_BFGS.GradientInfo getGradient(double[] beta)

Evaluate gradient at solution beta.

Parameters:

beta -

Returns:

getObjVals

public abstract double[] getObjVals(double[] beta,
                  double[] pk,
                  int nSteps,
                  double initialStep,
                  double stepDec)

Evaluate objective values at k line search points beta_k. When used as part of default line search behavior, the line search points are expected to be beta_k = beta + direction * _startStep * _stepDec^k

Parameters:

beta - - initial vector of coefficients

pk - - search direction

Returns:

objective values evaluated at k line-search points beta + pk*step[k]

doLineSearch

public L_BFGS.LineSearchSol doLineSearch(L_BFGS.GradientInfo ginfo,
                                double[] beta,
                                double[] direction,
                                int nSteps,
                                double tdec)

Perform line search at given solution and search direction.

Parameters:

ginfo - - gradient and objective value at current solution

beta - - current solution

direction - - search direction

Returns: