de.jungblut.online.regularization

Class L1Regularizer

java.lang.Object

de.jungblut.online.regularization.GradientDescentUpdater

de.jungblut.online.regularization.L1Regularizer

All Implemented Interfaces:

WeightUpdater

public class L1Regularizer
extends GradientDescentUpdater

Ported to "real" Java from Spark's mllib org.apache.spark.mllib.optimization.Updater. L1 regularizer: R(w) = ||w||_1. Uses a step-size decreasing with the square root of the number of iterations. Instead of subgradient of the regularizer, the proximal operator for the L1 regularization is applied after the gradient step. This is known to result in better sparsity of the intermediate solution.

Constructor Summary

Constructors

Constructor and Description
L1Regularizer()

Method Summary

Modifier and Type	Method and Description
CostWeightTuple	computeNewWeights(de.jungblut.math.DoubleVector theta, de.jungblut.math.DoubleVector gradient, double learningRate, long iteration, double lambda, double cost) Simplistic gradient descent without regularization.

Methods inherited from class de.jungblut.online.regularization.GradientDescentUpdater

computeGradient

Methods inherited from class java.lang.Object

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

Constructor Detail

L1Regularizer

public L1Regularizer()

Method Detail

computeNewWeights

public CostWeightTuple computeNewWeights(de.jungblut.math.DoubleVector theta,
                                         de.jungblut.math.DoubleVector gradient,
                                         double learningRate,
                                         long iteration,
                                         double lambda,
                                         double cost)

Description copied from class: GradientDescentUpdater

Simplistic gradient descent without regularization.

Specified by:

computeNewWeights in interface WeightUpdater

Overrides:

computeNewWeights in class GradientDescentUpdater

Parameters:

theta - the old weights.

gradient - the gradient.

learningRate - the learning rate.

iteration - the number of the current iteration.

lambda - the regularization parameter.

cost - the computed cost for this gradient update.

Returns:

the already updated weights for a particular updated gradient.