Layered genetic programming for feature extraction in classification problems

Abstract

Genetic programming has been proven to be a successful technique for feature extraction in various applications. In this thesis, we present a Layered Genetic Programming system which implements genetic programming-based feature extraction mechanism. The proposed system uses a layered structure where instead of evolving just one population of individuals, several populations are evolved sequentially. Each such population transforms the input data received from the previous population into a lower dimensional space with the aim of improving classification performance. The performance of the proposed system was experimentally tested on 5 real-world problems using different dimensionality reduction step sizes and different classifiers. The proposed method was able to outperform a simple classifier applied directly on the original data on two problems. On the remaining problems, the classifier performed better using the original data. The best solutions were often obtained in the first few layers which implied that increasing the size of the system, i.e. adding more layers was not useful. However, the layered structure allowed control of the size of individuals.