Authors:
Geoffrey Neumann
and
David Cairns
Affiliation:
University of Stirling, United Kingdom
Keyword(s):
Estimation of Distribution Algorithms, Feature Selection, Genetic Algorithms, Hybrid Algorithms.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Computational Intelligence
;
Evolutionary Computing
;
Genetic Algorithms
;
Hybrid Systems
;
Informatics in Control, Automation and Robotics
;
Intelligent Control Systems and Optimization
;
Soft Computing
Abstract:
This paper presents the results of applying the hybrid Targeted Estimation of Distribution Algorithm (TEDA) to feature selection problems with 500 to 20,000 features. TEDA uses parent fitness and features to provide a target for the number of features required for classification and can quickly drive down the size of the selected feature set even when the initial feature set is relatively large. TEDA is a hybrid algorithm that transitions between the selection and crossover approaches of a Genetic Algorithm (GA) and those of an Estimation of Distribution Algorithm (EDA) based on the reliability of the estimated probability distribution.Targeting the number of features in this way has two key benefits. Firstly, it enables TEDA to efficiently find good solutions for cases with very low signal to noise ratios where the majority of available features are not associated with the given classification task. Secondly, due to the tendency of TEDA to select the smallest and most promising init
ial feature set, it builds compact classifiers that are able to evaluate populations more quickly than other approaches.
(More)