Optimization for pattern classification using biased random search techniques

We demonstrate how optimization problems arise in the field of pattern classification, in particular in using piecewise-linear classification and classification based on an optimal linear separator. We motivate the need in this area for a general purpose optimization approach. We discuss ALOPEX, a biased random search approach, from the point of view of this need. While ALOPEX itself failed to fulfil our need, a newly-introduced generalization of it (iterated ALOPEX) was found to be appropriate for the optimization problems of our particular concern. We conclude the paper with a brief critical evaluation of this approach as compared to our original aims.