calhoun.analysis.crf.solver
Class StandardOptimizer

java.lang.Object
  calhoun.analysis.crf.solver.StandardOptimizer

All Implemented Interfaces:

CRFTraining

public class StandardOptimizer
extends java.lang.Object
implements CRFTraining

uses a L-BFGS algorithm to optimize the objective function. This is the algorithm normally used. Several configuration parameters all you to control optimization parameters:

This optimizer requires one property to be set:

gradFunc - This is the gradient function that the optimizer will use. It must be a reference to a bean that implements the CRFObjectiveFunctionGradient interface.

In addition, it has several optional properties that allow control over the optimization process:

• debugLevel - explicitly sets the debug level in the underlying L-BFGS solver. Normally it is set based on the debug level for this object.
• epsForConvergence - epsilon value used to determine when to halt the optimzation and declase convergence. Defaults to 0.0001.
• fixFirstWeight - if true, the first feature weight will be fixed at 1.0 and will not be allowed to change during the optimization
• maxIters - The maximum number of iterations (objective function evaluations) to attempt
• mForHessian - sets the mForHession parameter in the underlying L-BFGS solver. Defaults to 20.
• quadraticRegularization - if set to a nonzero value, regularizes feature weights by imposing a penalty on the objective function based on the absolute sizes of the weights. Can help in cases where weights get very large.
• requireConverge - if true, throws an error if convergence is not reached in the maximum number of iterations. Otherwise, the current feature weights are returned when maxIters is reached.
• starts - an initial set of guesses at feature weights. Defaults to 1.0
• unchangedObjective - the number of times the objectiveFunction must return an identical value before the optimization is assumed to have converged. This protects against a situation where the epsForConvergence is set low enough that even the smallest possible changes in weights can't find an optimum. Defaults to 5

Constructor Summary

StandardOptimizer()

Method Summary

int	getDebugLevel()
double	getEpsForConvergence()
int	getMaxIters()
int	getMForHessian()
CRFObjectiveFunctionGradient	getObjectiveFunction() returns the configured objective function gradient which will be used by the optimizer during the training process.
double	getQuadraticRegularization()
double[]	getStarts()
int	getUnchangedObjective()
boolean	isFixFirstWeight()
boolean	isRequireConvergence()
double[]	optimize(ModelManager fm, java.util.List<? extends TrainingSequence<?>> data) find the set of weights which maximizes the value of the objective function.
void	setDebugLevel(int debugLevel)
void	setEpsForConvergence(double epsForConvergence)
void	setFixFirstWeight(boolean fixFirstWeight)
void	setMaxIters(int maxIters)
void	setMForHessian(int forHessian)
void	setObjectiveFunction(CRFObjectiveFunctionGradient objectiveFunction) sets the objective function gradient.
void	setQuadraticRegularization(double quadraticRegularization)
void	setRequireConvergence(boolean requireConvergence)
void	setStarts(double[] starts) Sets the starting weights for the optimization.
void	setUnchangedObjective(int unchangedObjective)

Methods inherited from class java.lang.Object

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

Constructor Detail

StandardOptimizer

public StandardOptimizer()

Method Detail

optimize

public double[] optimize(ModelManager fm,
                         java.util.List<? extends TrainingSequence<?>> data)

Description copied from interface: CRFTraining

find the set of weights which maximizes the value of the objective function. The CRFObjectiveFunctionGradient object will already be configured witht eh appropriate training data, so it appears here as a pure function evaluation.

Specified by:

optimize in interface CRFTraining

Parameters:

fm - the model to train on.

data - the training data to use for training.

Returns:

an array of doubles containing the optimal weights.

getObjectiveFunction

public CRFObjectiveFunctionGradient getObjectiveFunction()

returns the configured objective function gradient which will be used by the optimizer during the training process.

Returns:

the configured objective function gradient

setObjectiveFunction

public void setObjectiveFunction(CRFObjectiveFunctionGradient objectiveFunction)

sets the objective function gradient. Called automatically during configuration.

getDebugLevel

public int getDebugLevel()

setDebugLevel

public void setDebugLevel(int debugLevel)

getEpsForConvergence

public double getEpsForConvergence()

setEpsForConvergence

public void setEpsForConvergence(double epsForConvergence)

isFixFirstWeight

public boolean isFixFirstWeight()

setFixFirstWeight

public void setFixFirstWeight(boolean fixFirstWeight)

getMaxIters

public int getMaxIters()

setMaxIters

public void setMaxIters(int maxIters)

getMForHessian

public int getMForHessian()

setMForHessian

public void setMForHessian(int forHessian)

getQuadraticRegularization

public double getQuadraticRegularization()

setQuadraticRegularization

public void setQuadraticRegularization(double quadraticRegularization)

isRequireConvergence

public boolean isRequireConvergence()

setRequireConvergence

public void setRequireConvergence(boolean requireConvergence)

getStarts

public double[] getStarts()

setStarts

public void setStarts(double[] starts)

Description copied from interface: CRFTraining

Sets the starting weights for the optimization.

Specified by:

setStarts in interface CRFTraining

Parameters:

starts - an array of weights to use as the starting point for the optimizer. The optimzier is not required to use this as a starting point.

getUnchangedObjective

public int getUnchangedObjective()

Returns:

Returns the unchangedObjective.

setUnchangedObjective

public void setUnchangedObjective(int unchangedObjective)

Parameters:

unchangedObjective - The unchangedObjective to set.