---

Terminator Class Reference

Abstract superclass for early stopping strategies for AnyNetwork. More...

#include <termination.h>

Inheritance diagram for Terminator:

List of all members.

Public Methods
	Terminator (const PatternSet &validationset, int striplen)
bool	validate (const ANNetwork &net, Trainer &trainer, int cyclesTrained)
virtual bool	restore (ANNetwork &net)
double	generalizationLoss (double min=-666, double opt=-666) const
double	validationError () const
double	minimumError () const
int	howManyTrained () const
Protected Attributes
double	mMinValidError
int	mMinCycle
double	mLastValidError
const PatternSet &	mValidationSet
int	mStripLength

---

Detailed Description

Abstract superclass for early stopping strategies for AnyNetwork.

The idea behing early stopping is as follows. When neural networks are trained with a training set, they often learn the training set so well that they "overfit" to it and start behaving badly with other patterns. With early stopping we try to stop the learning at the time when the networks starts to generalize badly. Generalization ability is tested with a separate validation set.

Terminates the subject when it is about to learn too much.

Definition at line 52 of file termination.h.

---

Constructor & Destructor Documentation

Terminator (  const PatternSet &    validationset, int    striplen )

Standard constructor. Parameters: validationset  Validation set that is used for measuring the generalization ability of the neural network taught with the separate training data. striplen  Interval of training cycles between validation and thus early stopping test. Definition at line 69 of file termination.cc. References mLastValidError, mMinCycle, and mMinValidError.

---

Member Function Documentation

double generalizationLoss (  double    min = -666, double    opt = -666 )  const

Returns the current generalization loss (GL), OR calculates it for the given parameters, if given. Parameters: min  OPTIONAL Lowest validation MSE measured so far. opt  OPTIONAL Optimal validation MSE measured so far. [TODO: check this] Definition at line 75 of file termination.cc. References mLastValidError, and mMinValidError. Referenced by Trainer::train(), and validate().

int howManyTrained (    )  const [inline]

Returns the number of training cycles the network has been trained before the minimum validation error point. Definition at line 103 of file termination.h. References mMinCycle. Referenced by Trainer::train().

double minimumError (    )  const [inline]

Returns the lowest measured validation error. Definition at line 98 of file termination.h. References mMinValidError. Referenced by Trainer::train().

virtual bool restore (  ANNetwork &    net )  [inline, virtual]

Restores the network state state in Terminator::validate(). The most important implementor of this method is SavingTerminator::restore(). Returns: Value 'true' if an earlier network state was restored, 'false' if not. See also: SavingTerminator Reimplemented in SavingTerminator. Definition at line 84 of file termination.h. Referenced by Trainer::train().

bool validate (  const ANNetwork &    net, Trainer &    trainer, int    cyclesTrained )

Tests the networks with the validation set and returns 'true' if the training should be terminated. The Terminator::restore method should be called to restore the optimal network state. See also: SavingTerminator Definition at line 82 of file termination.cc. References generalizationLoss(), mLastValidError, mValidationSet, Trainer::setGeneralizLoss(), and Learner::test(). Referenced by Trainer::train().

double validationError (    )  const [inline]

Returns the last measured validation error. Definition at line 95 of file termination.h. References mLastValidError. Referenced by Trainer::train().

---

Member Data Documentation

double mLastValidError [protected]

Last validation error. Definition at line 113 of file termination.h. Referenced by generalizationLoss(), Terminator(), validate(), and validationError().

int mMinCycle [protected]

The number of training cycles at minimum. Definition at line 110 of file termination.h. Referenced by howManyTrained(), SavingTerminator::save(), and Terminator().

double mMinValidError [protected]

Smalled validation error so far. Definition at line 107 of file termination.h. Referenced by generalizationLoss(), minimumError(), and Terminator().

int mStripLength [protected]

Validation interval. Definition at line 119 of file termination.h.

const PatternSet& mValidationSet [protected]

Validation pattern set. Definition at line 116 of file termination.h. Referenced by validate().

---

The documentation for this class was generated from the following files:

• termination.h
• termination.cc

---