LinearAssociator

Module that learns a linear mapping from INPUT to OUTPUT based on training samples in T-INPUT and T-OUTPUT. The module can operate in two different modes:

In gradient descent the, a gradient descent method is used to update the weights based on each new sample. By default, only a single sample is used on every tick. By increasing the values of memory_max, it is possible to store a number of samples and train on all of them every tick. My setting mamory_training to a value larger than 1, the data will be used several times.
In least square mode all stored samples are used to directly set the optimal associative weights that minimizes the missmatch between the inputs and outputs in the training data. The amount of trainsing data used is controlled by the memory_max parameter. If this parameter is larger than the current tick then all data so far seen will be used. If it is smaller, a random selection of training points will be throws away. This means that is memory_max is set to a low value, the trained asssociations will track changes in the data. For least square mode, it is important that memory_max is not set too low.

Example XML definition

A simple example

  <module
      class = "LinearAssociator"
      name = "LinearAssociator"
      alpha = "0.1"
  />

Parameters

Name	Description	Type	Default
class	The class name for the module; must be "LinearAssociator".	string
name	The name of this instance of the module.	string
mode	The learning mode	choices: gradient_descent LMS	gradient_descent
alpha	The learning rate	float	0.1
beta	The momentum rate	float	0.0
memory_max	Maximum number of stored training samples	int	1
memory_training	Number of times to train on each memorized sample	int	1

Module Connections

Inputs

Name	Description
LEARNING	The learning rate
INPUT	The input
T-INPUT	The training input
T-OUTPUT	The training target input

Outputs

Name	Description
OUTPUT	The output
MATRIX	The association matrix
ERROR	The error for the last training sample.
CONFIDENCE	1-ERROR

Author

Christian Balkenius
christian.balkenius@lucs.lu.se
Lund University Cognitive Science

Files

LinearAssociator.h
LinearAssociator.cc
LinearAssociator.ikc