Class implementing radial basis function networks for classification, trained in a fully supervised manner using WEKA's Optimization class by minimizing squared error with the BFGS method
Note that all attributes are normalized into the [0,1] scale.The initial centers for the Gaussian radial basis functions are found using WEKA's SimpleKMeans.
The initial sigma values are set to the maximum distance between any center and its nearest neighbour in the set of centers.There are several parameters.
The ridge parameter is used to penalize the size of the weights in the output layer.The number of basis functions can also be specified.
Note that large numbers produce long training times.Another option determines whether one global sigma value is used for all units (fastest), whether one value is used per unit (common practice, it seems, and set as the default), or a different value is learned for every unit/attribute combination.
It is also possible to learn attribute weights for the distance function.(The square of the value shown in the output is used.) Finally, it is possible to use conjugate gradient descent rather than BFGS updates, which can be faster for cases with many parameters, and to use normalized basis functions instead of unnormalized ones.
To improve speed, an approximate version of the logistic function is used as the activation function in the output layer.Also, if delta values in the backpropagation step are within the user-specified tolerance, the gradient is not updated for that particular instance, which saves some additional time.
Paralled calculation of squared error and gradient is possible when multiple CPU cores are present.Data is split into batches and processed in separate threads in this case.
Note that this only improves runtime for larger datasets.Nominal attributes are processed using the unsupervised NominalToBinary filter and missing values are replaced globally using ReplaceMissingValues.
(based on WEKA 3.7)
For further options, click the 'More' - button in the dialog.
All weka dialogs have a panel where you can specify classifier-specific parameters.