Supervised Neural Network Classifier

A supervised neural network classifier learned to associate driver behavior with roadway traffic controls. The input layer of the two layer network had 11 nodes and the output layer had three nodes. The neural network was fully connected: all output nodes received input from every one of the input nodes.

We constructed the neural network classifier using the Aspirin/Migraines package [4]. It used a learning rate of 0.05 and inertia of 0.95. It ran 50,000 iterations to train and updated the weights each time training data were presented.

We used K-fold cross validation, where K = 10, to evaluate the performance of the network. K-fold cross-validation works as follows:

1.

Split a data set of N instances into K cuts containing (N/K) random instances.

2.

Form a testing set with each cut.

3.

Form a training set for every testing set with the remaining (N - N/K) instances.

4.

Train and test the neural network using each of the pair of training and testing sets.

5.

Record and average the results for the testing sets to determine the performance of the network.