Enhancing Genetic Programming for
Predictive Modeling
RIKARD KÖNIG
Computer Science
Örebro Studies in Technology 58 I
ÖREBRO 2014ÖREBRO STUDIES IN TECHNOLOGY 58 2014
RIK
AR
D K
ÖN
IG
En
ha
nc
ing G
en
eti
c P
ro
gr
am
m
ing f
or P
red
ic
tiv
e M
od
eli
ing
rikard könig (b. 1977) is a member of the CSL@BS research group and a lecturer in informatics at the University of Borås. His doctoral studies have been a collaboration between the universities of Örebro, Skövde and Borås. The research focus is evolutionary computation and data mining, more specifically algorithmic improvement of predictive modelling techniques. Traditional predictive modeling techniques are most often optimized using a local greedy strategy and are tied to specific representations and optimization criteria, leading to potentially suboptimal solutions with regard to both predictive performance and comprehensibility. However, using genetic programming (GP), both the model representation of the solution and the optimization criteria can be designed specifically for the problem at hand and the model is optimized globally. On the other hand, GP has disadvantages that traditional techniques do not have, i.e., it is inherently inconsistent, struggles with large search spaces, and produces programs that often contain introns. Furthermore, its advantages are rarely fully exploited in the few predictive modeling frameworks that incorporate GP.
Hence, this thesis aims to enhance GP for predictive modeling, by exploit-ing the advantages and counteractexploit-ing the deficiencies. A set of criteria for frameworks of predictive modeling based on GP are suggested and several novel techniques that improve the accuracy or comprehensibility of predictive models produced using GP are presented. Finally, a GP framework for pre-dictive modeling, i.e., G-REX, has been implemented. G-REX fulfills most of the suggested criteria, realizes many of the presented techniques, and is available to the public on www.grex.se.
issn 1650-8580 isbn 978-91-7529-001-0
