Unsupervised Problem Decomposition using Genetic Programming

Created by W.Langdon from gp-bibliography.bib Revision:1.4420

  author =       "Ahmed Kattan and Alexandros Agapitos and 
                 Riccardo Poli",
  title =        "Unsupervised Problem Decomposition using Genetic
  booktitle =    "Proceedings of the 13th European Conference on Genetic
                 Programming, EuroGP 2010",
  year =         "2010",
  editor =       "Anna Isabel Esparcia-Alcazar and Aniko Ekart and 
                 Sara Silva and Stephen Dignum and A. Sima Uyar",
  volume =       "6021",
  series =       "LNCS",
  pages =        "122--133",
  address =      "Istanbul",
  month =        "7-9 " # apr,
  organisation = "EvoStar",
  publisher =    "Springer",
  keywords =     "genetic algorithms, genetic programming",
  isbn13 =       "978-3-642-12147-0",
  DOI =          "doi:10.1007/978-3-642-12148-7_11",
  abstract =     "We propose a new framework based on Genetic
                 Programming (GP) to automatically decompose problems
                 into smaller and simpler tasks. The frame-work uses GP
                 at two levels. At the top level GP evolves ways of
                 splitting the fitness cases into subsets. At the lower
                 level GP evolves programs that solve the fitness cases
                 in each subset. The top level GP programs include two
                 components. Each component receives a training case as
                 the input. The components' outputs act as coordinates
                 to project training examples onto a 2-D Euclidean
                 space. When an individual is evaluated, K-means
                 clustering is applied to group the fitness cases of the
                 problem. The number of clusters is decided based on the
                 density of the projected samples. Each cluster then
                 invokes an independent GP run to solve its member
                 fitness cases. The fitness of the lower level GP
                 individuals is evaluated as usual. The fitness of the
                 high-level GP individuals is a combination of the
                 fitness of the best evolved programs in each of the
                 lower level GP runs. The proposed framework has been
                 tested on several symbolic regression problems and has
                 been seen to significantly outperforming standard GP
  notes =        "Part of \cite{Esparcia-Alcazar:2010:GP} EuroGP'2010
                 held in conjunction with EvoCOP2010 EvoBIO2010 and

Genetic Programming entries for Ahmed Kattan Alexandros Agapitos Riccardo Poli