Genetic Programming for Multiclass Object Classification
Will Smart · 2005
This report describes the use of Genetic Programming (GP) to solve multipleclass object classification problems. Objects is reduced to feature vectors containing four features. GP is then used to evolve classifiers for the feature vectors. Objects are taken from four image datasets of increasing difficulty. Three research directions are discussed in this report. Three classification strategies are compared, including two new ones, Centred Dynamic Range Selection (CDRS) and Slotted Dynamic Range Selection (SDRS). The new strategies were found to improve the performance of the system, over the standard SRS. This is especially true for harder problems with classes in an arbitrary order. Gradient-descent search on individual programs is introduced to GP in this report. GP is still used as a global beam search, but another, local gradientdescent search is made on programs. The subject of the search is the numeric terminals of the programs. The use of gradient-descent markedly improves the