Neural Processing Letters 7: 1–4, 1998. 1c 1998 Kluwer Academic Publishers. Printed in the Netherlands.Characterization of the Sonar Signals BenchmarkJUAN MANUEL TORRES MORENO AND MIRTA B. GORDONDepart´ ement de Recherche Fondamentale sur la Matier` e Condensee´ , CEA/Grenoble – 17, Av. desMartyrs – 38054 Grenoble Cede´ x 9. FranceE-mail: Mirta.Gordon@cea.frKey words: classification problems, generalization, learning, perceptrons, sonar targetsAbstract. We study the classification of sonar targets first introduced by Gorman & Sejnowski(1988). We discovered that not only the training set and the test set of this benchmark are bothlinearly separable, although by different hyperplanes, but that the complete set of patterns, trainingand test patterns together, is also linearly separable. The distances of the patterns to the separatinghyperplane determined by learning with the training set alone, and to the one determined by learningthe complete data set, are presented.It has become a current practice to test the performance of learning algorithms onrealistic benchmark problems. The underlying difficulty of such tests is that in gen-eral these problems are not well characterized: given a solution to the classificationproblem, it is impossible to decide whether a better one exists.The sonar signals benchmark [1] has been widely used to test learning algorithms[2–10]. In this problem the classifier has to discriminate if a given sonar returnwas produced by a metal ...