Help Topics
Concepts
Package
Class
Neural Classifer Agent Properties
and Use
The Neural Classifier Agent panel provides these options:
- Training File Name
- Enter the name of a definition
file. Use the Browse button to select a
definition file. This file will be used to specify the
network architecture's input value since it defines the
number of fields and their data type. One of the fields
must have a field name of class to support
supervised learning. The data itself must be a file by
the same name with a .dat extension. This will be used
when the mode is Train and also to Generate
Beans.
- Testing File Name
- Enter the name of a definition file with the same layout
as the training definition file. Use the Browse
button to select a definition file. This will be used
when you set the mode to Test and also when the
training process uses Test mode to conditionally end
training.
- Agent Mode
- Select one of the following agent modes:
Train implies that the network bean's weights are
unlocked, and the import bean referencing the Training
File provides the active data buffer connection.
Test implies that the network bean's weights are
locked, and the percent correct is calculated; the import
bean referencing the Testing File provides the active
data buffer connection.
Run implies that the network bean's weights are
locked and the percent correct is not calculated.
- Test/Train Ratio
- Once the desired accuracy specified by Minimum
Percent Correct is attained from training, this
value controls additional training required to improve
the network accuracy calculated from the test data. A
ratio of 10 means that the network trains with the
training data for 10 passes before switching to the test
data to calculate the number of records in the test data
that are classified correctly by the network bean.
- Network Architecture
- The architecture consists of five parameters:
- Inputs, which is calculated when beans are
generated from the Training File.
- Hidden1, which is the number of hidden
units in the first layer.
- Hidden2, which is the number of hidden
units in the second layer.
- Hidden3, which is the number of hidden
units in the third layer.
- Outputs, which is calculated when beans
are generated from the Training File.
- Minimum Percent Correct
- The threshold value for the minimum number of records
which the network must accurately classify. Training the
network stops when this threshold is attained. If test
data is supplied, training continues until the results
from the test data reach this level.
- Error Tolerance
- In Train or Test modes, errors are calculated for each
record and compared to the Tolerance value. Errors
greater than the tolerance value indicate a bad
calculation. If the error is within the tolerance, it is
treated as 0. Tolerance must be a real value between 0.0
and 1.0. Generally the tolerance for Test mode is higher
than the tolerance in Training mode.
- Maximum Number of Passes
- The threshold value for the maximum number of passes.
Training the network will stop if this threshold is
attained regardless of the calculated Minimum
Percent Correct.
The Neural Classifer Agent panel is used to generate an agent
containing import beans for
training and testing, a BackPropagation bean,
filter beans to
translate network inputs and outputs, and data connections. The Agent
Mode is set so that the network bean can be trained or used
to provide an independant data source to test that training is
sufficient.
Steps in using the panel for training include:
- Enter the name of the training file that defines the
record layout and name of the training data source. If
the file is read successfully, the Generate Beans
button should be enabled.
- Select Train for the Agent Mode.
- Set the values for hidden unit layers.
- Press the Generate Beans button. The Start
Training button should now be enabled.
- Change the tolerance values if desired - a lower training
tolerance will provide more accuracy but possibly
relatively less generalizing capability and more time
will be needed to train the network.
- Set the Minimum Percent Correct and Maximum
Passes values to control when training will stop. The
Minimum Percent Correct determines the network's
classification accuracy while the Maximum Passes provides
an alternate endpoint should the desired accuracy be
unachievable in a reasonable period of time.
- Press the Start Training button to begin
processing records through the data buffer connections.
Once training has begun, this button will toggle to Stop
Training. Training will continue until the toggled
button is pressed, the Minimum Percent Correct
value is met, or the Maximum Passes value is
attained.
- You may wish to press the Stop Training button,
change an accuracy threshold, and start again. If you
change the network architecture, press Generate
Beans for the changes to take effect. Press the Reset
Beans button to re-initialize the network weights
before starting training again if you wish.
- After the network bean has trained to the specified
accuracy with the training data, the training process
internally switches to Test mode. It will enable the data
buffer from the test import bean to calculate the
classification accuracy for the test data with the
network weights locked. If the required percentage of
records in the test data set are within the Tolerance,
training stops. Otherwise training
continues for the number of passes specified by Train/Test
Ratio before the accuracy is again calculated
from the test data.