RF Class Reference

The RF class represents a receptive field. More...

#include <rf.hpp>

Inheritance diagram for RF:

Inheritance graph
[legend]
Collaboration diagram for RF:

Collaboration graph
[legend]
List of all members.

Public Member Functions

 RF ()
 Default constructor for the RF.
 RF (Types::Input c)
 A constructor taking only a center point for the RF.
 RF (Types::Input c, const boost::program_options::variables_map &vm)
 A constructor taking a center point and a variables_map.
virtual Types::RValue getActivation (Types::Input x) const =0
 Return the activation for a test point.
Types::OutputT predict (Types::Input x) const
 Return the predicted value for an unput x.
Types::OutputT confidence (Types::Input x) const
 return the predictive variance for a point
bool trustworthy () const
 Check whether the RF is trustworthy.
virtual Types::RValue learn (Types::Input x, Types::Output y, Types::RValueP w)=0
 Update the receptive field and have local model update itself.

Protected Attributes

Types::vector z
 store the projection of x
Types::vector temp2
Types::vector q
Types::vector H_temp
Algorithm Parameters
This group of variables control the behavior of the algorithm.

All paramters are defined in the paper (see PAPER).

Types::RParam d_def
 d_def is the initial value for diagonal of distance matrix in RF.
Types::RParam lambda
 A forgetting factor for sufficient statistics.
Types::RParam gamma
 A weight on the distance matrix penalty.
Types::RParam alpha_init
 alpha_init is the initial learning rate for all distance elements
bool meta
 determine if meta updates are done
Types::RValue meta_alpha
 meta_learning rate
RF Descriptors
These variables describe the receptive field.

They include the center, distance matrix, and local regression model

Types::vector center
 The center point of the RF.
Model localModel
 the receptive fields local model
Sufficient Statistics
These variables hold the sufficient statistics for updating the RF and local model.

Types::RValue W
 Activation history.
Types::vector a_H
Types::vector a_G
Types::RValue a_E
Types::RValue a_pk
 confidence intervals ss
Types::RValue sigma
Types::RValue e_2

Friends

std::ostream & operator<< (std::ostream &out, const RF &rf)
 Print the receptive field to a stream in simple format.

Detailed Description

The RF class represents a receptive field.

This class handles updates of the distance matrix for the receptive field and hold the local model associated with the field.

Definition at line 38 of file rf.hpp.


Constructor & Destructor Documentation

RF::RF ( Types::Input  c  ) 

A constructor taking only a center point for the RF.

Parameters:
c An input vector corresponding to the center of the RF.

Definition at line 36 of file rf.cpp.

References a_G, a_H, alpha_init, d_def, e_2, gamma, lambda, meta, and W.


Member Function Documentation

virtual Types::RValue RF::getActivation ( Types::Input  x  )  const [pure virtual]

Return the activation for a test point.

This currently uses ublas vectro expressions, but should be changed to use blas calls once everything is working correctly.

Parameters:
x the input vector

Implemented in RFDiag, and RFFull.

Referenced by confidence().

Types::OutputT RF::predict ( Types::Input  x  )  const [inline]

Return the predicted value for an unput x.

Parameters:
x the input vector

Definition at line 162 of file rf.hpp.

References localModel, and Model::predict().

Here is the call graph for this function:

Types::OutputT RF::confidence ( Types::Input  x  )  const

return the predictive variance for a point

Parameters:
x the input point

Definition at line 88 of file rf.cpp.

References a_pk, Model::check_derivatives(), Model::get_a_zz(), getActivation(), localModel, Model::project(), q, and z.

Here is the call graph for this function:

bool RF::trustworthy (  )  const [inline]

Check whether the RF is trustworthy.

See also:
Model::trustworthy()

Definition at line 177 of file rf.hpp.

References localModel, and Model::trustworthy().

Here is the call graph for this function:

virtual Types::RValue RF::learn ( Types::Input  x,
Types::Output  y,
Types::RValueP  w 
) [pure virtual]

Update the receptive field and have local model update itself.

Update the distance matrix as in Equation 3.6 and Table 4 $ M^{n+1} = M^n - \alpha \frac{\partial J}{\partial M} $

Parameters:
x the input vector
y the actual output

Implemented in RFDiag, and RFFull.


Member Data Documentation

Types::RParam RF::lambda [protected]

A forgetting factor for sufficient statistics.

This parameter must have a value in [0,1]. It may be updated over time.

Definition at line 59 of file rf.hpp.

Referenced by RFFull::learn(), RFDiag::learn(), and RF().


The documentation for this class was generated from the following files:
Generated on Fri Jul 27 00:24:02 2007 for LWPR by  doxygen 1.5.1