Help Topics     Concepts     Package     Class


Radial Basis Function Net Bean Properties and Use

Properties

The Back Propagation Bean panel provides these options:

Architecture
The network architecture consists of these parameters:
  1. Inputs, which must match the number of outputs provided any bean with a data buffer connections.
  2. Hidden, which is the number of basis units in the hidden layer.
  3. Outputs, which is calculated when beans are generated from the Training File.
Learn Rate
Enter a value to control how much the network weights are changed during a weight update. Larger values cause more change. Learn rate is a real value between 0.0 and 10.0, with a typical starting value of 0.2.
Momentum
Enter a value to control the amount that previous network weight updates should influence the current network weight update. This acts as a smoothing parameter that reduces oscillation and helps attain convergence. Momentum is a real value between 0.0 and 1.0, with a typical value of 0.5.
Tolerance
During training 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. A typical value is 0.1.
All Widths
Controls the width or selectivity of the selected radial basis function. If set to a non-zero value, all of the hidden units will use the parameter value; smaller values denote smaller, more focused selection. A zero value is not currently supported; for some implementations, a 0 value signals that each hidden unit takes a unique width parameter from the Widths array.
Mode
Select one of the following agent modes:
Train implies that the network bean's weights are unlocked, and network weights will be adjusted as data is processed.
Test implies that the network bean's weights are locked, and that error calculations will be performed as data is processed.
Run implies that the network bean's weights are locked and no error calculations are made.
Basis Function
Selects which radial basis function is used for the activation function in the hidden units. Three functions are currently supported:
0. Gaussian = exp(- v**2 / 2 width**2)
1. Thin Plate Spline = v**2 * log(v)
2. Multi-quadratic = sqrt( v**2 + width**2)

where v is the Euclidian Norm, the distance between the Input vector and the hidden unit Center, calculated as the square root of the sum of the squared element differences:
sqrt(∑(Input[i] - Center[i])**2)

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. This is used in control applications.
Normalized
Controls whether the hidden unit activations are normalized so that they sum to 1. If Normalized = 1, then all hidden unit values are summed and then divided by the sum. If Normalized = 0, then the hidden unit activation values are set to the value returned by the selected radial basis function. Normalization tends to force a single result rather than hedging between several responses.
Epoch (batch) updates
Select batch updates if network weights are updated only after a complete training epoch. Otherwise weights are updated after each record.
Auto-find center weights
Controls whether the first layer weights are learned or set. A value of 1 indicates the center weights are determined using self-organizing learning. A value of 0 indicates the center weights must be explicitly set.
Adaptive learn rate
Select adaptive learning if the Learn Rate is to be lowered as training progresses.

Use

The Back Propagation Bean panel is used to create a network with specified architecture and training parameters. The 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:

  1. Set the architecture input value to the number of outputs from the bean providing data.
  2. Set the values for hidden unit layer.
  3. Set the number of outputs.
  4. Select the basis function.
  5. Set Auto-find center to true.
  6. Press the Set Architecture button.
  7. Train the network by pressing the Step, Cycle, or Run buttons on the Agent Editor toolbar.
  8. You may wish to press the Stop toolbar button, change a parameter such as Learn Rate or Tolerance, and start again. If you change the network architecture, press Set Architecture for the changes to take effect. Press the Reset Weights button to re-initialize the network weights before starting training again if you wish.