Neural and statistical classifiers. Can such approaches be complementary?
Résumé
Neural networks are efficient in certain pattern recognition sub-problems, especially in feature extraction and classification. In many cases neural and statistical techniques are seen as alternatives. Our aim is to verify if these approaches can give complementary responses in order to consider the implementation of fusion methods. The comparison is applied to three examples belonging to mobile robot localization: (i) laser range finder modeling, (ii) feature extraction from ultrasonic range finder data and (iii) localization by a stereoscopic camera. In each case the solution of the problem is based partly on a classifier. The paper compares the performances of a multilayer perceptron (MLP) known as an efficient classifier and three statistical methods-quadratic discriminant analysis (QDA), linear discriminant analysis (LDA) and Bayesian. The performances of the classifier are estimated by classical criteria such as success and misclassification percentages and the study is completed by a sharp analysis where the method results are crossed two by two to evaluate the success percentage of a method applied to the misclassified set of another one. Experiments show the set of patterns misclassified by the different classifiers does not completely overlap.