Help Topics
Concepts
Package
Class
Radial Basis Function Net Bean Properties
and Use
The Back Propagation Bean panel provides these options:
- Architecture
- The network architecture consists of these parameters:
- Inputs, which must match the number of
outputs provided any bean with a data buffer
connections.
- Hidden, which is the number of basis units
in the hidden layer.
- 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.
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:
- Set the architecture input value to the number of outputs
from the bean providing data.
- Set the values for hidden unit layer.
- Set the number of outputs.
- Select the basis function.
- Set Auto-find center to true.
- Press the Set Architecture button.
- Train the network by pressing the Step, Cycle, or Run
buttons on the Agent Editor toolbar.
- 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.