Multi-classifier Fusion by Genetic Programming

B F Buxton and W B Langdon
Department of Computer Science, University College London,
Gower Street, London WC1E 6BT, UK.

Genetic programs and algorithms are often used to solve difficult optimization problems in science, engineering and commerce. One such problem, where the optimal solution requires a function to be found is the combination of classifiers to produce an improved decision system. There are no general rules as to how a number of classifiers should best be combined in a multi-classifier system, although it is known that weak classifiers should be combined additively and independent classifiers multiplicatively. In this talk, we will describe how effective, non-trivial, classifier combinations can automatically be generated by genetic programming (GP), in particular when a robust performance measure based on the area under classifier receiver-operating-characteristic (ROC) curves is used as a fitness measure. It will be shown how this performance measure may be used in a GP to facilitate evolution of multi-classifier systems that outperform their constituent individual classifiers. The approach will be illustrated by results obtained on test data sets from the multi-classifier combination and machine learning literature and by application to publicly available Landsat data and to pharmaceutical data of the kind used in one stage of the drug design process. The development of this approach to classifier fusion and, in particular its application to the drug design process, was carried out in collaboration with Galxo-SmithKline R&D, Harlow, in a collaborative EPSRC supported research project within the INTErSECT Faraday Partnership.

For some details and list of publications see: