Evolution of Artificial Neural Networks Using a Two-dimensional Representation

Created by W.Langdon from gp-bibliography.bib Revision:1.4496

  author =       "Joao Carlos Figueira Pujol",
  title =        "Evolution of Artificial Neural Networks Using a
                 Two-dimensional Representation",
  school =       "School of Computer Science, University of Birmingham",
  year =         "1999",
  address =      "UK",
  month =        apr,
  email =        "pujol@urano.cdtn.br",
  keywords =     "genetic algorithms, genetic programming, PDGP",
  URL =          "http://www.cs.ucl.ac.uk/staff/W.Langdon/ftp/papers/PhD_Thesis_pujol.pdf",
  size =         "178 pages",
  abstract =     "The design of artificial neural networks is still
                 largely performed by an expert, with only a few
                 heuristics to guide a trial-and-error search. Recently,
                 new methods based on evolutionary computation (EC) have
                 been applied to the synthesis of artificial neural
                 networks with modest results. The basic limitation of
                 EC-based methods is that they do not take into account
                 the fact that artificial neural networks are
                 two-dimensional structures, and do not use specialized
                 evolutionary operators. In this work, a new method
                 based on a special form of evolutionary computation
                 called genetic algorithms is proposed for the evolution
                 of artificial neural networks. The method is a general
                 purpose procedure able to evolve feedforward and
                 recurrent architectures. It is based on a
                 two-dimensional representation, and includes operators
                 to evolve the architecture and the connection weights
                 simultaneously. The new approach has shown promising
                 results, and has fared better than previous methods in
                 a number of applications, including: binary
                 classification problems, design of neural controllers
                 and a complex navigation task of traversing a trail. An
                 extension of the two-dimensional representation is also
                 presented, which can be combined with other methods,
                 providing them with an alternative procedure to evolve
                 the weights of the connections.",

Genetic Programming entries for Joao Carlos Figueira Pujol