Improving Support Vector Machine Performance on a Problem in Object Recognition

This paper is concerned with the problem of object recognition using a sonar sensor array. The sonar returns are preprocessed and fed to a support vector machine (SVM) classifier to carry out the recognition. We demonstrate that it is possible to gradually improve recognition performance by using a genetic algorithm to select a sequence of subsets from the available training data. Performance improvement is possible because the SVM finds a solution that lies at the centre of the largest inscribable hypersphere in version space. The optimum solution lies at the Bayes Point which is generally some distance away from the SVM solution.