Learn more about Model-Based Calibration

Prune Functionality

You can use the Prune function to reduce the number of centers in a radial basis function network. This helps you decide how many centers are needed.

To use the Prune facility:

1. Select an RBF global model in the model tree.

2. Either click the toolbar button or select the menu item Model > Utilities > Prune.

The Number of Centers Selector dialog box appears.

The graphs show how the fit quality of the network builds up as more RBFs are added. It makes use of the fact that most of the center selection algorithms are greedy in nature, and so the order in which the centers were selected roughly reflects the order of importance of the basis functions.

The default fit criteria are the logarithms of PRESS, GCV, RMSE, and Weighted PRESS. Additional options are determined by your selections in Summary Statistics. See Model Menu (Global Level). Weighted PRESS penalizes having more centers, and choosing number of centers to minimize weighted PRESS is often a good option.

All four criteria in this typical example indicate the same minimum at eight centers.

If the graphs all decrease, as in the preceding example, this suggests that the maximum number of centers is too small, and the number of centers should be increased.

Clicking the Minimize button selects the number of centers that minimizes the criterion selected in the drop-down menu on the left. It is good if this value also minimizes all the other criteria. The Clear button returns to the previous selection.

Note that reducing the number of centers using Prune only refits the linear parameters (RBF weights). The nonlinear parameters (center locations, width, and lambda) are not adjusted. You can perform a cheap width refit on exiting the dialog box by selecting the Refit widths on close check box. If a network has been pruned significantly, you should use the Update Model Fit toolbar button. This performs a full refit of all the parameters.

Width Selection Algorithms

Statistics

© 1984-2009- The MathWorks, Inc.    -   Site Help   -   Patents   -   Trademarks   -   Privacy Policy   -   Preventing Piracy   -   RSS