aureservoir::TrainPI< T > Class Template Reference

#include <train.h>

Inheritance diagram for aureservoir::TrainPI< T >:

[legend]

Collaboration diagram for aureservoir::TrainPI< T >:

[legend]

---

Detailed Description

template<typename T>
class aureservoir::TrainPI< T >

Trains the ESN offline in two steps, as described in Jaeger's "Tutorial on training recurrent neural networks"

See also:

http://www.faculty.iu-bremen.de/hjaeger/pubs/ESNTutorial.pdf

1. teacher-forcing/sampling: collects the internal states and desired outputs
2. computes output weights usings LAPACK's xGELSS routing, which performs a singular value decomposition and gets the pseudo inverse

See also:

http://www.netlib.org/lapack/single/sgelss.f

The difference to TrainLS is, that TrainPI is computationally more expansive, but TrainLeastSquare can have stability problems.

See also:

class TrainLeastSquare

For a more mathematical description:

See also:

http://en.wikipedia.org/wiki/Linear_least_squares

http://www.netlib.org/lapack/lug/node27.html

Public Member Functions
	TrainPI (ESN< T > *esn)
virtual	~TrainPI ()
class TrainPI Implementation
virtual void	train (const typename ESN< T >::DEMatrix &in, const typename ESN< T >::DEMatrix &out, int washout) throw (AUExcept)
Protected Member Functions
void	clearData ()
class TrainBase Implementation
void	checkParams (const typename ESN< T >::DEMatrix &in, const typename ESN< T >::DEMatrix &out, int washout) throw (AUExcept)
void	collectStates (const typename ESN< T >::DEMatrix &in, const typename ESN< T >::DEMatrix &out, int washout)
void	squareStates ()
Protected Attributes
ESN< T > *	esn_
ESN< T >::DEMatrix	M
ESN< T >::DEMatrix	O

---

Constructor & Destructor Documentation

template<typename T>

aureservoir::TrainPI< T >::TrainPI

(

ESN< T > *

esn

)

[inline]

template<typename T>

virtual aureservoir::TrainPI< T >::~TrainPI

(

)

[inline, virtual]

---

Member Function Documentation

template<typename T>

void aureservoir::TrainPI< T >::train	(	const typename ESN< T >::DEMatrix &	in,
		const typename ESN< T >::DEMatrix &	out,
		int	washout
	)			throw (AUExcept) [inline, virtual]

training algorithm

Implements aureservoir::TrainBase< T >.

template<typename T>

void aureservoir::TrainBase< T >::checkParams	(	const typename ESN< T >::DEMatrix &	in,
		const typename ESN< T >::DEMatrix &	out,
		int	washout
	)			throw (AUExcept) [inline, protected, inherited]

check parameters

template<typename T>

void aureservoir::TrainBase< T >::collectStates	(	const typename ESN< T >::DEMatrix &	in,
		const typename ESN< T >::DEMatrix &	out,
		int	washout
	)			[inline, protected, inherited]

collect network states with simulation algorithm

template<typename T>

void aureservoir::TrainBase< T >::squareStates

(

)

[inline, protected, inherited]

squares states for SIM_SQUARE

template<typename T>

void aureservoir::TrainBase< T >::clearData

(

)

[inline, protected, inherited]

frees allocated data for M and O

---

Field Documentation

template<typename T>

ESN<T>* aureservoir::TrainBase< T >::esn_ [protected, inherited]

reference to the data of the network

template<typename T>

ESN<T>::DEMatrix aureservoir::TrainBase< T >::M [protected, inherited]

matrix for network states and inputs over all timesteps

template<typename T>

ESN<T>::DEMatrix aureservoir::TrainBase< T >::O [protected, inherited]

matrix for outputs over all timesteps

---

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

• train.h
• train.hpp

---