The Geometry of Tartarus Fitness Cases

Created by W.Langdon from gp-bibliography.bib Revision:1.3973

@InProceedings{Ashlock3:2008:cec,
  author =       "Daniel Ashlock and Elizabeth Warner",
  title =        "The Geometry of Tartarus Fitness Cases",
  booktitle =    "2008 IEEE World Congress on Computational
                 Intelligence",
  year =         "2008",
  editor =       "Jun Wang",
  pages =        "1309--1316",
  address =      "Hong Kong",
  month =        "1-6 " # jun,
  organization = "IEEE Computational Intelligence Society",
  publisher =    "IEEE Press",
  isbn13 =       "978-1-4244-1823-7",
  file =         "EC0339.pdf",
  DOI =          "doi:10.1109/CEC.2008.4630965",
  abstract =     "Tartarus is a standard AI task for grid robots in
                 which boxes must be moved to the walls of a virtual
                 world. There are 320,320 fitness cases for the standard
                 Tartarus task of which 297,040 are valid according to
                 the original statement of the problem. This paper
                 studies different schemes for allocating fitness trials
                 for Tartarus using an agent-based metric on the fitness
                 cases to aid in the design process. This agent-based
                 metric is a tool that permits exploration of the
                 geometry of the space of fitness cases. The information
                 gained from this exploration demonstrates why a scheme
                 designed to yield a superior set of training cases in
                 fact yielded an inferior one. The information gained
                 also suggests a new scheme for allocating fitness
                 trials that decreases the number of trials required to
                 achieve a given fitness of the best agent. This scheme
                 achieves similar fitness to a standard evolutionary
                 algorithm using fewer fitness cases. The space of
                 fitness cases for Tartarus is found, relative to the
                 agent-based metric, to form a hollow sphere with a
                 nonuniform distribution of the fitness cases within the
                 space. The tools developed in this study include a
                 generalisable technique for placing an agent-based
                 metric space structure on the fitness cases of any
                 problem that has multiple fitness cases. This metric
                 space structure can be used to better understand the
                 distribution of fitness cases and so design more
                 effective evolutionary algorithms.",
  keywords =     "genetic algorithms, genetic programming",
  notes =        "WCCI 2008 - A joint meeting of the IEEE, the INNS, the
                 EPS and the IET.",
}

Genetic Programming entries for Daniel Ashlock Elizabeth Warner

Citations