Linear GP for multi-class classification problems (Dissertation)

Abstract: Multi-class image classification is an important research topic within computer science and artificial intelligence. A wide range of images of different types need to be classified, and the techniques developed for multi-class image classification can often be applied to other forms of multi-class classification. Genetic programming has had some success with these problems in the past, however multi-class image classification is still a difficult task and the resulting classifiers are often very hard for humans to understand. This project develops a methodology using linear genetic programming for multi-class image classification. A new fitness function is developed to improve this methodology and the standard tree-based genetic programming methodology. Two heuristics are found to guide initial decisions on a linear genetic programming configuration and to aid in comparing tree-based genetic programming and linear genetic programming configurations. The resulting methodology is compared to the standard genetic programming approach to multi-class image classification. The methodology developed outperforms the standard genetic programming methodology on all six of the varying tasks and does so significantly on five of them. A hill climbing algorithm is also developed and this algorithm augments the powerful evolutionary beam search linear genetic programming uses.

  @misc{fogelberg05,
  author       = "Christopher Fogelberg",
  title        = "Linear Genetic Programming for Multi-class Classification Problems",
  howpublished = {{BSc (Honours) Research Project Report, School of Mathematics, Statistics and Computer Science,
                   Victoria University of Wellington, New Zealand.}},
  month        = "November",
  year         = "2005",
  note = 	 "Available from {\tt{http://syntilect.com/cgf/pubs/dissertation}}"
}

Available here.