Genetic Programming with Gradient Descent Search for Multiclass Object Classification


Authors: Mengjie Zhang, Will Smart
Source: GZipped PostScript (770kb); Adobe PDF (321kb)

This paper describes an approach to the use of gradient descent search in genetic programming (GP) for object classification problems. In this approach, pixel statistics are used to form the feature terminals and a random generator produces numeric parameter terminals. The four arithmetic operators and a conditional operator form the function set and the classification accuracy is used as the fitness function. Gradient descent search is introduced to the GP mechanism and is embedded into the genetic beam search, which allows the evolutionary learning process to globally follow the beam search and locally follow the gradient descent search. Two different methods, an online gradient descent scheme and an offline gradient descent scheme, are developed and compared with the basic GP method on three image data sets with object classification problems of increasing difficulty. The results show that both the online and the offline gradient descent GP methods outperform the basic GP method in both classification accuracy and training time and that the online scheme achieved better performance than the offline scheme. This suggests that the GP method with gradient descent search is more effective and more efficient than without and that the online gradient descent algorithm is best suited to object classification problems. Although developed for object classification problems, this approach is expected to be able to be applied to general classification and prediction tasks.

[Up to Computer Science Technical Report Archive: Home Page]