Recursively partitioning neural networks for radar target recognition
T. Kerstetter, S. Massey, J. Roberts · 2003
Conflicting information in the training data is responsible for most of the problems experienced by the backpropagation algorithm during network training. The self-partitioning neural network (SPNN) approach has been shown to be effective in overcoming the ill effects of learning conflict that exists among the patterns of a given class. Intra-class learning conflict present in the training patterns is reduced by using a divide-and-conquer approach. A cluster seeking algorithm is used to partition training patterns according to a conflict metric. Simulation studies have verified that the SPNN has distinct advantages over the conventional backpropagation network. However for the large training sets mandated by theater missile defense performance constraints, calculation of the conflict metrics used in the SPNN is computationally intractable. Furthermore, cluster seeking algorithms are not guaranteed to converge. The recursively partitioning neural network (RPNN) was created to remove these two shortcomings of the SPNN. Two computationally efficient methods for measuring conflict and an efficient partitioning scheme are developed. Simulation results are presented to demonstrate the RPNN's capabilities.