Using Factorial Experiments to Evaluate the Effect of Genetic Programming Parameters
by R. Feldt and P. Nordin
PDF
Statistical techniques for designing and analysing experiments are used to evaluate the individual and combined effects of genetic programming parameters. Three binary classification problems are investigated in a total of seven experiments consisting of 1108 runs of a machine code genetic programming system. The parameters having the largest effect in these experiments are the population size and the number of generations. A large number of parameters have negligible effects. The experiments indicate that the investigated genetic programming system is robust to parameter variations, with the exception of a few important parameters.
Bibtex
@InProceedings{Feldt2000FactorialGP,
author = "Robert Feldt and Peter Nordin",
title = "Using Factorial Experiments to Evaluate the Effect of Genetic Programming Parameters",
year = "2000",
editor = "Riccardo Poli and Wolfgang Banzhaf and William B. Langdon and Julian F. Miller and Peter Nordin and Terence C. Fogarty",
pages = "271--282",
notes = "EuroGP Conference 2000",
volume = "1802",
series = "Lecture Notes on Computer Science",
address = "Edinburgh",
publisher = "Springer-Verlag",
publisher_address = "Berlin",
keywords = "Evolutionary algorithms; Genetic programming; Parameter setting",
keywords = "Design of Experiment; Factorial experiment; Performance evaluation",
url = "http://www.cse.chalmers.se/~feldt/publications/feldt_2000_factorial_exp_gp_params.html",
url = "http://www.cse.chalmers.se/~feldt/publications/feldt_2000_factorial_exp_gp_params.pdf",
}