Training set optimization methods for a probabilistic neural network

Abstract

In a real-time probabilistic neural network (PNN), both speed and accuracy are important for classification. In this work, three methods for reducing the size of a training set are compared: learning vector quantization (LVQ), reciprocal neighbors (RN) and a general grouping method (GGM). Each method produced multiple reductions that were tested to see the effects on the speed and accuracy of the PNN. The reductions showed little effect on the classification, 85–90% correct, or time for detection of flaming fires but increased the time for detection of smoldering fires. The general grouping method worked best, reducing the training set by 50% with an average of less than 4-s delay. The LVQ method reduced the training set by 75% but with a delay of 30–45 s. The RN method was able to reduce the training set with a larger range, from 35% to 75%, but gave results with an average delay of 40–50 s.