- author = "Umberto Armani and Vassili V. Toropov and Andrey Polynkin and Osvaldo M. Querin and Luis Alvarez",
- title = "Enhancements to a hybrid genetic programming technique applied to symbolic regression",
- booktitle = "Proceedings of the 8th ASMO UK / ISSMO conference on Engineering Design Optimization Product and Process Improvement",
- year = "2010",
- editor = "Fabian Duddeck and Osvaldo M. Querin and Johann Sienz and Vassili V. Toropov and M. Hasan Shaheed",
- address = "Queen Mary University of London, UK",
- month = jul # " 8-9",
- keywords = "genetic algorithms, genetic programming",
- isbn13 = "978-0-85316-292-6",
-
URL = "
http://www.asmo-uk.com/8th-asmo-uk/html/menu_page.html",
-
URL = "
http://www.asmo-uk.com/8th-asmo-uk/presentations/session1_presentation5.pdf",
-
abstract = "A major problem in genetic programming techniques is
premature convergence, which emerges during evolution
as a progressive loss of variability among individuals
in the population. Moreover, the mechanisms according
to which individuals are created, recombined and
evaluated have of course strong influence on the
chances of success. Increasing variability of the
population and expressivity of the genotype are then
major issues for genetic programming techniques.
The aim of this paper is to investigate if a hybrid, tree-based GP implementation written for symbolic regression purposes can be improved in terms of reliability and precision of the results both by several modifications of the standard GP components and by pre-processing the input data set.
In order to increase variability, the effect of a simple archive updating strategy and of a periodical killing of a large part of the population (with the insertion of new and composed individuals) is assessed. As a promising measure to preserve variation among individuals, a MinMax approach in the definition of the fitness function is also proposed and tested as an alternative to the plain aggregating approach.
With regard to expressivity, a simple solution consisting in the definition of a unary function that introduces a translation in the argument of the function itself is put forward.
Other experiments are performed to assess if the redefinition of the fitness function using a normalised error can have beneficial effects on the evolution, as an alternative to the common root mean square error.
Finally, the splitting of the input data set in two different subsets, respectively for parameter tuning and fitness evaluation, is investigated.",
-
notes = "SQP Rosenbrock KILLandFILL Kotanchek RatPol2D Hock
Branin-Hoo Salustowicz session1_presentation5.pdf is
slides
School of Civil Engineering, University of Leeds, LS2 9JT, UK",
