SANN - Automated Network Search (ANS) - Quick Tab

Select the Quick tab of the SANN - Automated Network Search (ANS) dialog box to access the options described here. For information on the options that are common to all tabs, see SANN - Automated Network Search (ANS).

Network types. Use the options in this group box to specify the type of network (MLP or RBF). For each selected type, you can also specify a range for the complexity of the neural network models to be tried by the Automated Network Search (ANS). Specify the complexity of networks to be tested in terms of a range of figures for the number of hidden units. Specifying the number of hidden units exactly (i.e., by setting the minimum equal to the maximum) may be beneficial if you know, or have good cause to suspect, the optimal number. In this case, it allows the Automated Network Search (ANS) to concentrate its search algorithms on other dimensions in the search such as activation functions. The larger the number of hidden units in a neural network model the stronger the model is, i.e. the more capable the network is to model complex relationships between the inputs and the target variables.

MLP. Select the MLP check box to include multilayer perceptron networks in the network search. The multilayer perceptron is the most common form of network. It requires iterative training, which may be quite slow for a large number of hidden units and data sets, but the networks are quite compact, execute quickly once trained, and in most problems yield better results than the other types of networks.

Min. hidden units. Specify the minimum number of hidden units to be tried by the Automated Network Search (ANS) when using MLP networks.

Max. hidden units. Specify the maximum number of hidden units to be tried by the Automated Network Search (ANS) when using MLP networks.

RBF. Select the RBF check box to include radial basis function networks in the network search. Radial basis function networks tend to be slower and larger than multilayer perceptron, and often have relatively inferior performance, but they train extremely quickly when the output activation functions are the identity. They are also usually less effective than multilayer perceptrons if you have a large number of input variables (they are more sensitive to the inclusion of unnecessary inputs).

Min. hidden units. Specify the minimum number of hidden units to be tried by the Automated Network Search (ANS) when using RBF networks.

Max. hidden units. Specify the maximum number of hidden units to be tried by the Automated Network Search (ANS) when using RBF networks.

Note: What effect does the number of hidden units have? In general, increasing the number of hidden units increases the modeling power of the neural network (it can model a more convoluted, complex underlying function), but also makes it larger, more difficult to train, slower to operate, and more prone to over-fitting (modeling noise instead of the underlying function). Decreasing the number of hidden units has the opposite effect.

If your data is from a fairly simple function or is very noisy, or if you have too few cases, a network with relatively few hidden units is preferable. If, in experimenting with different numbers of hidden units you find that larger networks have better training performance, but worse selection performance, then you are probably over-fitting and should revert to smaller networks.

To combat overfitting SANN uses a test sample (which you can specify in the Data selection dialog box) which can help the SANN training algorithm. This test sample is never used to train the neural network (i.e., to learn the data) but rather used to monitor performance throughout training at the end of each iteration cycle. See Overfitting for more details.  

Train/Retain networks. Use the options in this group box to specify how many networks should be trained and how many networks should be retained by the ANS.

Networks to train. Specify how many networks the Automated Network Search (ANS) should perform. The larger the number of networks trained the more detailed is the search carried out by the ANS. It is recommended that you set the value for this option as large as possible depending on your hardware speed and resources.

Networks to retain. Specify how many of the neural networks tested by the Automated Network Search (ANS) should be retained (for testing, and then insertion into the current network set). Networks with the lowest error for regression and highest classification rate for classification will be retained.

Error function. Specify the error function to be used in training a network.

Sum of squares. Select the Sum of Squares check box to generate networks using the sum of squares error function. This error function is one of the most commonly used in training neural networks and available in SANN for both MLP and RBF types of networks in both regression and classification tasks.

Cross entropy. Select the Cross entropy check box to generate networks using cross entropy error functions. This error function assumes that the data is drawn from the multinomial family of distributions (see Bishop 1995 for more details) and supports a direct probabilistic interpretation of the network outputs. Note that this error function is only available for classification problems. It will be disabled for regression type analyses. While using the cross entropy error function, the output activation functions are automatically set to softmax. This restriction ensures that the network outputs are true class membership probabilities, which is known to enhance the performance of classification neural networks.