Probability based genetic programming for multiclass object classification

This paper describes a probability based genetic programming (GP) approach to multiclass object classication problems. Instead of using predened multi-ple thresholds to form dierent regions in the program output space for dierent classes, this approach uses probabilities of dierent classes, derived from Gaus-sian distributions, to construct the tness function for classication. Two tness measures, overlap area and weighted distribution distance, have been developed. The approach is examined on three multiclass object classication problems of increasing diculty and compared with a basic GP approach. The results sug-gest that the new approach is more eective and more ecient than the basic GP approach. While the area measure was a bit more eective than the distance measure in most cases, the distance measure was more ecient to learn good program classiers