hex.glm

Class LSMSolver

java.lang.Object

water.Iced

hex.glm.LSMSolver

All Implemented Interfaces:

java.lang.Cloneable, Freezable

Direct Known Subclasses:

LSMSolver.ADMMSolver, LSMSolver.ProxSolver

public abstract class LSMSolver
extends Iced

Distributed least squares solvers

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	LSMSolver.ADMMSolver
static class	LSMSolver.LSMSolverException
static class	LSMSolver.LSMSolverType
static class	LSMSolver.ProxSolver

Field Summary

Fields

Modifier and Type	Field and Description
protected boolean	_converged
java.lang.String	_id
Key	_jobKey

Constructor Summary

Constructors

Constructor and Description
LSMSolver(double lambda, double alpha)

Method Summary

Methods

Modifier and Type	Method and Description
boolean	converged()
double[]	grad(Gram gram, double[] beta, double[] xy)
protected double	lsm_objectiveVal(double[] xy, double yy, double[] beta, double[] xb) Compute the LSM objective.
abstract java.lang.String	name()
protected double	objectiveVal(double[] xy, double yy, double[] beta, double[] xb) Compute least squares objective function value: lsm_obj(beta) = 0.5*(y - X*b)'*(y - X*b) + l1 + l2 = 0.5*y'y - (X'y)'*b + 0.5*b'*X'X*b) + l1 + l2 l1 = alpha*lambda*l1norm(beta) l2 = (1-alpha)*lambda*l2norm(beta)/2
protected static double	shrinkage(double x, double kappa)
abstract boolean	solve(Gram gram, double[] xy, double yy, double[] newBeta)
static void	subgrad(double alpha, double lambda, double[] beta, double[] grad)

Methods inherited from class water.Iced

clone, frozenType, init, newInstance, read, toDocField, write, writeJSON, writeJSONFields

Methods inherited from class java.lang.Object

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

Field Detail

_jobKey

public Key _jobKey

_id

public java.lang.String _id

_converged

protected boolean _converged

Constructor Detail

LSMSolver

public LSMSolver(double lambda,
         double alpha)

Method Detail

grad

public final double[] grad(Gram gram,
            double[] beta,
            double[] xy)

subgrad

public static void subgrad(double alpha,
           double lambda,
           double[] beta,
           double[] grad)

solve

public abstract boolean solve(Gram gram,
            double[] xy,
            double yy,
            double[] newBeta)

Parameters:

xy - - guassian: -X'y binomial: -(1/4)X'(XB + (y-p)/(p*1-p))

yy - - < y,y > /2

newBeta - - resulting vector of coefficients

Returns:

true if converged

converged

public final boolean converged()

name

public abstract java.lang.String name()

shrinkage

protected static double shrinkage(double x,
               double kappa)

objectiveVal

protected double objectiveVal(double[] xy,
                  double yy,
                  double[] beta,
                  double[] xb)

Compute least squares objective function value: lsm_obj(beta) = 0.5*(y - X*b)'*(y - X*b) + l1 + l2 = 0.5*y'y - (X'y)'*b + 0.5*b'*X'X*b) + l1 + l2 l1 = alpha*lambda*l1norm(beta) l2 = (1-alpha)*lambda*l2norm(beta)/2

Parameters:

xy: - X'y

yy: - 0.5*y'y

beta: - b (vector of coefficients)

xb: - X'X*beta

Returns:

0.5*(y - X*b)'*(y - X*b) + l1 + l2

lsm_objectiveVal

protected double lsm_objectiveVal(double[] xy,
                      double yy,
                      double[] beta,
                      double[] xb)

Compute the LSM objective. lsm_obj(beta) = 0.5 * (y - X*b)' * (y - X*b) = 0.5 * y'y - (X'y)'*b + 0.5*b'*X'X*b) = 0.5yy + b*(0.5*X'X*b - X'y)

Parameters:

xy - X'y

yy - y'y

beta -

xb - X'X*beta

Returns: