Help Topics
Concepts
Package
Class
Neural Prediction Agent Properties
and Use
The Neural Prediction 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 predicted 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 Maximum
RMS Error 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 error from the records in the test data.
- 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.
- Maximum RMS Error
- Enter the threshold value for the maximum root mean-squared
error between the actual value and the value calculated
by the network. Training the network will stop if when
the error falls below this threshold.
- Explicit error mode
- Select explicit error mode if the input buffer during
training contains the actual error value. Otherwise it is
assumed to be the desired or target value.
- Tolerance
- In Test mode the error is 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. In Training mode, Tolerance is set at the
network bean level.
- Maximum Number of Passes
- The threshold value for the maximum number of passes.
Training the network will stop if this threshold is
attained.
The Neural Prediction 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.
- Set the Maximum RMS Error to control when training will
stop, and the Tolerance as desired.
- 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 accuracy threshold is met, or the Maximum
Passes value is attained.
- You may wish to press the Stop Training button,
change the Maximum RMS Error or Tolerance, 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.
- To use the test capability, enter a file in the Testing
File Name field that uses same layout as the
Training File definition. Once provided with the test
data and 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 RMS error from the test data with the network weights
locked. If the required error calculated in the test data
set is less than the maximum specified, 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.