Automatically Evolving Rule Induction Algorithms with Grammarbased Genetic Programming
Created by W.Langdon from
gpbibliography.bib Revision:1.7697
 @PhdThesis{Pappa:thesis,

author = "Gisele Lobo Pappa",

title = "Automatically Evolving Rule Induction Algorithms with
Grammarbased Genetic Programming",

school = "Computing Laboratory, University of Kent",

year = "2007",

address = "Canterbury, UK",

keywords = "genetic algorithms, genetic programming",

URL = "http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.445792",

abstract = "In the last 30 years, research in the field of rule
induction algorithms produced a large number of
algorithms. However, these algorithms are usually
obtained from the combination of a basic rule induction
algorithm (typically following the sequential covering
approach) with new evaluation functions, pruning
methods and stopping criteria for refining or producing
rules, generating many {"}new{"} and more sophisticated
sequential covering algorithms. We cannot deny that
these attempts to improve the basic sequential covering
~approach have succeeded. Hence, if manually changing
these major components of rule induction algorithms can
result in new, significantly better ones, why not to
automate this process to make it more costeffective?
This is the core idea of this work: to automate the
process of designing rule induction algorithms by means
of grammarbased genetic programming. Grammarbased
Genetic Programming (GGP) is a special type of
evolutionary algorithm used to automatically evolve
computer programs. The most interesting feature of this
type of algorithm is that it incorporates a grammar
into its search mechanism, which expresses prior
knowledge about the problem being solved. Since we have
a lot of previous knowledge about how humans design
rule induction algorithms, this type of algorithm is
intuitively a suitable tool to automatically evolve
rule induction algorithms. The grammar given to the
proposed GGP system includes knowledge about how humans
design rule induction algorithms, and also presents
some new elements which could work in rule induction
algorithms, but to the best of our knowledge were not
previously tested. The GG P system aims to evolve rule
induction algorithms under two different frameworks, as
follows. In the first framework, the GGP is used to
evolve robust rule induction algorithms, i.e.,
algorithms which were designed to be applied to
virtually any classification data set, like a
manuallydesigned rule induction algorithm. In the
second framework, the GGP is applied to evolve rule
induction algorithms tailored to a specific application
XVI domain, i.e., rule induction algorithms tailored to
a single data set. Note that the latter framework is
hardly feasible on a hard scale in the case of
conventional, manuallydesigned algorithms, since the
number of classification data sets greatly outnumbers
the number of rule induction algorithms designers.
However, it is clearly feasible on a large scale when
using the proposed system, which automates the process
of rule induction algorithm design and implementation.
Overall, extensive computational experiments with 20
VCI data sets and 5 bioinformatics data sets showed
that effective rule induction algorithms can be
automatically generated using the GGP in both
frameworks. Moreover, the automatically evolved rule
induction algorithms were shown to be competitive with
(and overall slightly better than) four wellknown
manually designed rule induction algorithms when
comparing their predictive accuracies. The proposed GGP
system was also compared to a grammarbased
hillclimbing system, and experimental results showed
that the GGP system is a more effective method to
evolve rule induction algorithms than the grammarbased
hillclimbing method. At last, a multiobjective version
of the GGP (based on the concept of Pareto dominance)
was also proposed, and experiments were performed to
evolve robust rule induction algorithms which generate
both accurate and simple models. The results showed
that in most of the cases the GGP system can produce
rule induction algorithms which are competitive in
predictive accuracy to wellknown humandesigned rule
induction algorithms, but generate simpler
classification modes, i.e., smaller rule sets,
intuitively easier to be interpreted by the user.",

notes = "Submitted in August
uk.bl.ethos.445792",
 }
Genetic Programming entries for
Gisele L Pappa
Citations