Towards understanding the Correlation of Problem Difficulty and Parameter Sensitivity in Genetic Programming

Created by W.Langdon from gp-bibliography.bib Revision:1.4549

  author =       "Alan Piszcz",
  title =        "Towards understanding the Correlation of Problem
                 Difficulty and Parameter Sensitivity in Genetic
  school =       "Department of Computer Science, University of Idaho",
  year =         "2006",
  address =      "Moscow, ID, USA",
  month =        nov,
  keywords =     "genetic algorithms, genetic programming",
  URL =          "",
  URL =          "",
  size =         "145 pages",
  abstract =     "Genetic programming is a machine-learning technique
                 inspired by biological evolution to synthesize programs
                 for a given computational task. It can produce results
                 that equal or exceed human-designed solutions in areas
                 such as electronic circuits, electronic filters,
                 electronic control systems, optics, antennas and
                 planning problems. Genetic programming algorithms use
                 control parameters to modify the simulation
                 environment. A significant problem confronting
                 researchers is determining the appropriate set of
                 values such as population size and mutation frequency.
                 The choices for parameter settings vary from default
                 values, to experience or rule of thumb techniques
                 discovered by the research community. Inappropriate
                 parameter settings often result in no solutions. We
                 examine mutation techniques, population size and
                 associated parameter values to explain how each
                 inhibits or promotes the evolution of solutions with
                 genetic programming. The experiments analyse the
                 performance of the genetic program using parameter
                 sweeps for mutation frequency and population size.
                 Analysis of the mutation experiments result in three
                 general findings. First, the selection method for
                 individuals to mutate (e.g. the best, the worst, the
                 average, etc.) is a critical factor. Second, we show
                 non-linear effects on fitness from three structure
                 altering mutation techniques. All three structure
                 altering techniques show a non-linear relationship
                 between the rate of mutation and the performance of the
                 GP. Third, a correlation exists between problem
                 complexity and mutation frequency. The results of the
                 population experiments led to four important, specific
                 discoveries. First, the determination of near optimal
                 parameter settings leads to an improvement in
                 computational efficiency. Second, improvement in the
                 computational efficiency that is obtained when optimal
                 or near optimal parameter values are used varies with
                 problem complexity. Third, problems with low complexity
                 show significant increases in the number of acceptable
                 solutions with changes in population size. Fourth, the
                 optimal population size narrows as problem difficulty
                 increases. The employment of mutation in GP research
                 papers needs improvement. We introduce an expanded
                 mutation description scheme to improve the role of
                 mutation in GP research.",
  notes =        "Supervisor: Terrance Soule

                 UMI Microform 3250627",

Genetic Programming entries for Alan Piszcz