A recursive self-learning pattern classification technique

The problem of classifying samples from a population into one of N classes is considered. Several classification algorithms are investigated, including a nearest mean method and one involving a constant learning parameter. However, the most effective method consists of using a Kalman filter to recursively estimate the vector of probabilities of occurrence for the N classes. The probabilities for the classes conditioned only on the current observation are used for the classification decision and also serve as pseudo-measurements into the Kalman filter. Extensive numerical results are presented for automatic identification of six classes of feed grain.