Neural Network for Classification Active Walker Patterns

A radial-basis-function neural network is constructed to classify patterns generated by the boundary probabilistic active walker model. The patterns generated by this model fall into five classes referred to as blob, jellyfish, diamond, lollipop, and needle. Some of these exhibit abnormal growth behavior such as reproducible or ~sfo~ation~ growth. For a test sample of 300 patterns, our neural network classifier cannot distinguish jellyfish from diamond, but can recognize blob, lollipop, needle, and jellyfish/diamond as four different classes with a success rate of about 99%. The failures occur within some parts of the sensitive zone, i.e., the zone in a parameter space in which more than one type of pattern are generated from different computer runs, due to the differing sequence of random numbers employed in each run. In developing this neural network classifier a number of issues are addressed, including the large dimension of the input space, the rotational variation within the classes, the intricacies of the patterns, and the de~nition of the metric.