Advancements in Genetic Programming for Data Classification

Abstract

This thesis aims to advance the state of the art in data classification using Genetic programming (GP). GP is an evolutionary algorithm that has several outstanding features making it ideal for complex problems like data classification. However, it suffers from a few limitations that reduce its significance. This thesis targets at proposing optimal solutions to these GP limitations. The problems covered in this thesis are: 1. Increase in GP tree complexity during evolution that results in long training time. 2. Lack of convergence to a single (optimal) solution. 3. Lack of methodology to handle mixed data-type without type transformation. 4. Search of a better method for multi-class classification. Through this work, we have proposed a method which achieves significant reduction in bloat for classification task. Moreover, we have presented a Particle Swarm Optimization based hybrid approach to increase performance of GP evolved classifiers. The approach offers better performance in less computational effort. Another approach introduces a new two layered paradigm for mixed type data classification with an added feature that uses data in its original form instead of any transformation or pre-processing. The last but not the least contribution is an efficient binary encoding method for multi-class classification problems. The method involves smaller number of GP evolutions, reducing the computation and suffers from fewer conflicts yielding better results. All of the proposed methods have been tested and our experiments conclude the efficiency of proposed approaches.