Package pulse.search.direction

Class GradientBasedOptimiser

java.lang.Object

pulse.util.UpwardsNavigable

pulse.util.Group

pulse.util.Accessible

pulse.util.PropertyHolder

pulse.search.direction.PathOptimiser

pulse.search.direction.GradientBasedOptimiser

All Implemented Interfaces:

Serializable, Descriptive, Reflexive

Direct Known Subclasses:

CompositePathOptimiser, LMOptimiser

public abstract class GradientBasedOptimiser
extends PathOptimiser

See Also:

Serialized Form

Constructor Summary

Constructors

Modifier	Constructor	Description
protected	GradientBasedOptimiser()	Abstract constructor that sets up the default ITERATION_LIMIT, ERROR_TOLERANCE and GRADIENT_RESOLUTION for this PathSolver.

Method Summary

Modifier and Type	Method	Description
protected double	dx(NumericProperty defProp, double value)	Calculates the gradient step.
NumericProperty	getGradientResolution()
double	getGradientStep()
Vector	gradient(GeneralTask task)	Calculates the Vector gradient of the target function (the sum of squared residuals, SSR, for this task.
Set<NumericPropertyKeyword>	listedKeywords()	The types of the listed parameters for this class include: GRADIENT_RESOLUTION, ERROR_TOLERANCE, ITERATION_LIMIT.
void	reset()	Resets the default ITERATION_LIMIT, ERROR_TOLERANCE and GRADIENT_RESOLUTION values for this PathSolver.
void	set(NumericPropertyKeyword type, NumericProperty property)	The accepted types are: GRADIENT_RESOLUTION, ERROR_TOLERANCE, ITERATION_LIMIT.
void	setGradientResolution(NumericProperty resolution)

Methods inherited from class pulse.search.direction.PathOptimiser

compatibleWith, genericProperties, getErrorTolerance, getInstance, getMaxIterations, getSolver, ignoreSiblings, initState, iteration, prepare, setErrorTolerance, setInstance, setMaxIterations, setSolver, toString, update

Methods inherited from class pulse.util.PropertyHolder

addListener, areDetailsHidden, data, describe, firePropertyChanged, getDescriptor, getListeners, getPrefix, initListeners, isListedNumericType, isListedParameter, listedTypes, numericData, parameterListChanged, removeListeners, setPrefix, updateProperties, updateProperty

Methods inherited from class pulse.util.Accessible

accessibleChildren, genericProperty, numericProperties, numericProperty, property, update

Methods inherited from class pulse.util.Group

access, children, contents, getSimpleName, subgroups

Methods inherited from class pulse.util.UpwardsNavigable

addHierarchyListener, getHierarchyListeners, getParent, identify, removeHierarchyListener, removeHierarchyListeners, setParent, specificAncestor, tellParent

Methods inherited from class java.lang.Object

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

Constructor Detail

GradientBasedOptimiser

protected GradientBasedOptimiser()

Abstract constructor that sets up the default ITERATION_LIMIT, ERROR_TOLERANCE and GRADIENT_RESOLUTION for this PathSolver. In addition, sets up a list of search flags defined by the Flag.defaultList method.

Method Detail

reset

public void reset()

Resets the default ITERATION_LIMIT, ERROR_TOLERANCE and GRADIENT_RESOLUTION values for this PathSolver. In addition, sets up a list of search flags defined by the Flag.defaultList method.

Overrides:

reset in class PathOptimiser

gradient

public Vector gradient(GeneralTask task)
                throws SolverException

Calculates the Vector gradient of the target function (the sum of squared residuals, SSR, for this task.

If $$Δf(Δx_i) is the change in the target function associated with the change of the parameter $$x_i, the i-th component of the gradient is equal to $$g_i = (Δf(Δx_i)/Δx_i). The accuracy of this calculation depends on the $$Δx_i value, which is roughly the GRADIENT_RESOLUTION. Note however that instead of using a forward-difference scheme to calculate the gradient, this method utilises the central-difference calculation of the gradient, which significantly increases the overall accuracy of calculation. This means that to evaluate each component of this vector, the Problem associated with this task is solved twice (for $$x_i ± Δx_i).

Parameters:

task - a SearchTask that is being driven to the minimum of SSR

Returns:

the gradient of the target function

Throws:

SolverException

dx

protected double dx(NumericProperty defProp,
                    double value)

Calculates the gradient step. Ensures dx is not zero even if the parameter values is. Applicable to discrete properties.

Parameters:

defProp - the default property

value - the value of the parameter under the optimisation vector

Returns:

the gradient step

setGradientResolution

public void setGradientResolution(NumericProperty resolution)

getGradientResolution

public NumericProperty getGradientResolution()

listedKeywords

public Set<NumericPropertyKeyword> listedKeywords()

The types of the listed parameters for this class include: GRADIENT_RESOLUTION, ERROR_TOLERANCE, ITERATION_LIMIT. Also, all the flags in this class are treated as separate listed parameters.

Overrides:

listedKeywords in class PathOptimiser

Returns:

a set of NumericPropertyKeyword instances, which have been explicitly marked as a listed parameter for this PropertyHolder.

See Also:

NumericPropertyKeyword

set

public void set(NumericPropertyKeyword type,
                NumericProperty property)

The accepted types are: GRADIENT_RESOLUTION, ERROR_TOLERANCE, ITERATION_LIMIT.

Overrides:

set in class PathOptimiser

Parameters:

type - the type, which must be equal by definition to property.getType().

property - the property, which contains new information.

getGradientStep

public double getGradientStep()