|
Abstract:
|
A Piecewise Linear Network (PLN ) is a local network that offers the accuracy of higher order networks and the Multi Layer Perceptron (MLP ) , with the computational simplicity of linear networks . A method to design a PLN is demonstrated and several clustering algorithms , used in the design procedure , are compared . The performance of the Self Organizing Map (SOM ) clustering algorithm has been found to be slightly better than the other clustering methods . Methods to determine the appropriate threshold in the Sequential Leader algorithm have been studied . A binary search based approach was found to be the most efficient in terms of the number of trials needed . Methods to delete extra clusters generated have been studied and compared to pruning . Pruning yields the best networks followed by deleting the smallest clusters . Methods of improving PLN pruning performance have been developed , including segregation of patterns by clusters , the use of partial distances , and redesign of only changed clusters . Results have been presented for several different data files . |