Abstract: Multi-class object classification is an important field of research in computer vision. In this paper basic linear genetic programming is modified to be more suitable for multi-class classification and its performance is then compared to tree-based genetic programming. The directed acyclic graph nature of linear genetic programming is exploited. The existing fitness function is modified to more accurately approximate the true feature space. The results show that the new linear genetic programming approach outperforms the basic tree-based genetic programming approach on all the tasks investigated here and that the new fitness function leads to better and more consistent results. The genetic programs evolved by the new linear genetic programming system are also more comprehensible than those evolved by the tree-based system. Bibtex entry:

@InProceedings{fogelbergZhang05,
  author =       "Christopher Fogelberg and Mengjie Zhang",
  title =        "Linear Genetic Programming for Multi-class Object Classification",
  booktitle =    "AI 2005: Advances in Artificial Intelligence: Proceedings of the 18th Australian Joint Conference on
                  Artificial Intelligence, Lecture Notes in Computer Science, Vol. 3809.",
  year =         "2005",
  month =        "December",
  editor =       "Shichao Zhang and Ray Jarvis",
  series =       "LNAI",
  volume =       "3809",
  address =      "Sydney, Australia",
  publisher =    "Springer Verlag",
  publisher_address = "Berlin",
  pages =        "369--379",
  keywords =     "linear genetic programming, evolutionary computation, multi-class classification",
  ISBN =         "3-540-30462-2",
  abstract =     "Multi-class object classification is an important field of research
                  in computer vision. In this paper basic linear genetic programming
                  is modified to be more suitable for multi-class classification and
                  its performance is then compared to tree-based genetic programming.
                  The directed acyclic graph nature of linear genetic programming is
                  exploited. The existing fitness function is modified to more
                  accurately approximate the true feature space. The results show that
                  the new linear genetic programming approach outperforms the basic
                  tree-based genetic programming approach on all the tasks
                  investigated here and that the new fitness function leads to better
                  and more consistent results. The genetic programs evolved by the new
                  linear genetic programming system are also more comprehensible than
                  those evolved by the tree-based system.",
  note = 	 "Available from {\tt{http://syntilect.com/cgf/pubs:aaai2005}}"
}

Available here.