Dynamics of genetic programming and chaotic time series prediction

  • Authors:

  • Brian S. Mulloy;Rick L. Riolo;Robert S. Savit

  • Affiliations:

  • University of Michigan, Ann Arbor, MI;University of Michigan, Ann Arbor, MI;University of Michigan, Ann Arbor, MI

  • Venue:
  • GECCO '96 Proceedings of the 1st annual conference on Genetic and evolutionary computation
  • Year:
  • 1996

Quantified Score

Hi-index 0.00

Visualization

Abstract

An investigation into the dynamics of Genetic Programming applied to chaotic time series prediction is reported. An interesting characteristic of adaptive search techniques is their ability to perform well in many problem domains while failing in others. Because of Genetic Programming's flexible tree structure, any particular problem can be represented in myriad forms. These representations have variegated effects on search performance. Therefore, an aspect of fundamental engineering significance is to find a representation which, when acted upon by Genetic Programming operators, optimizes search performance. We discover, in the case of chaotic time series prediction, that the representation commonly used in this domain does not yield optimal solutions. Instead, we find that the population converges onto one "accurately replicating" tree before other trees can be explored. To correct for this premature convergence we make a simple modification to the crossover operator. In this paper we review previous work with GP time series prediction, pointing out an anomalous result related to overlearning, and report the improvement effected by our modified crossover operator.