A Variant Program Structure in Tree-Based Genetic Programming for Multiclass Object Classification

Abstract

This chapter describes an approach to the use of genetic programming for multiclass object classification. Instead of using the standard tree-based genetic programming approach, where each genetic program returns just one floating point number that is then translated into different class labels, this approach invents a new program structure with multiple outputs, each for a particular class. A voting scheme is then applied to these output values to determine the class of the input object. The approach is examined and compared with the standard genetic programming approach on four multiclass object classification tasks with increasing difficulty. The results show that the new approach outperforms the basic approach on these problems. A characteristic of the proposed program structure is that it can easily produce multiple outputs for multiclass object classification problems, while still keeping the advantages of the standard genetic programming approach for easy crossover and mutation. This approach can solve a multiclass object recognition problem using a single evolved program in a single run.