de.jungblut.math.minimize

Class GradientDescent.GradientDescentBuilder

java.lang.Object

de.jungblut.math.minimize.GradientDescent.GradientDescentBuilder

Enclosing class:

GradientDescent

public static class GradientDescent.GradientDescentBuilder
extends Object

Method Summary

Modifier and Type	Method and Description
GradientDescent.GradientDescentBuilder	annealingAfter(int iteration) Sets a simple annealing (alpha / (1+current_iteration / phi)) where phi is the given parameter here.
GradientDescent.GradientDescentBuilder	boldDriver() BoldDriver will change the learning rate over time by observing the cost of the costfunction.
GradientDescent.GradientDescentBuilder	boldDriver(double increasedCostPercentage, double decreasedCostPercentage) BoldDriver will change the learning rate over time by observing the cost of the costfunction.
GradientDescent.GradientDescentBuilder	breakOnDifference(double delta) Breaks minimization process when the given delta in costs have been archieved.
GradientDescent.GradientDescentBuilder	breakOnDivergence() If called, this breaks when the gradient descent minimizer starts to diverge (costs are growing).
GradientDescent	build()
static GradientDescent.GradientDescentBuilder	create(double alpha) Creates a new builder.
GradientDescent.GradientDescentBuilder	momentum(double momentum) Add momentum to this gradient descent minimizer.

Methods inherited from class java.lang.Object

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

Method Detail

build

public GradientDescent build()

momentum

public GradientDescent.GradientDescentBuilder momentum(double momentum)

Add momentum to this gradient descent minimizer.

Parameters:

momentum - the momentum to use. Between 0 and 1.

Returns:

the builder again.

boldDriver

public GradientDescent.GradientDescentBuilder boldDriver()

BoldDriver will change the learning rate over time by observing the cost of the costfunction. If the cost decreases, it will increase the learning rate by 5%. If the cost increases it will cut the learning rate in half.

Returns:

the builder again.

boldDriver

public GradientDescent.GradientDescentBuilder boldDriver(double increasedCostPercentage,
                                                         double decreasedCostPercentage)

BoldDriver will change the learning rate over time by observing the cost of the costfunction. If the cost decreases, it will increase the learning rate (typically by 5%). If the cost increases it will (typically) cut the learning rate in half.

Parameters:

increasedCostPercentage - the percentage of the learning rate that will be used when cost increases.

decreasedCostPercentage - the percentage of the learning rate that will be used when cost decreases.

Returns:

the builder again.

breakOnDivergence

public GradientDescent.GradientDescentBuilder breakOnDivergence()

If called, this breaks when the gradient descent minimizer starts to diverge (costs are growing).

Returns:

the builder again.

breakOnDifference

public GradientDescent.GradientDescentBuilder breakOnDifference(double delta)

Breaks minimization process when the given delta in costs have been archieved. Usually a quite low value of 1e-4 to 1e-8.

Parameters:

delta - the delta to break in difference between two costs.

Returns:

the builder again.

annealingAfter

public GradientDescent.GradientDescentBuilder annealingAfter(int iteration)

Sets a simple annealing (alpha / (1+current_iteration / phi)) where phi is the given parameter here. This will gradually lower the global learning rate after the given amount of iterations.

Parameters:

iteration - the iteration to start annealing.

Returns:

the builder again.

create

public static GradientDescent.GradientDescentBuilder create(double alpha)

Creates a new builder.

Parameters:

alpha - the learning rate to set.

Returns:

a new builder.