% NAME: Sonar, Mines vs. Rocks% % SUMMARY: This is the data set used by Gorman and Sejnowski in their study% of the classification of sonar signals using a neural network [1].  The% task is to train a network to discriminate between sonar signals bounced% off a metal cylinder and those bounced off a roughly cylindrical rock.% % SOURCE: The data set was contributed to the benchmark collection by Terry% Sejnowski, now at the Salk Institute and the University of California at% San Deigo.  The data set was developed in collaboration with R. Paul% Gorman of Allied-Signal Aerospace Technology Center.% % MAINTAINER: Scott E. Fahlman% % PROBLEM DESCRIPTION:% % The file "sonar.mines" contains 111 patterns obtained by bouncing sonar% signals off a metal cylinder at various angles and under various% conditions.  The file "sonar.rocks" contains 97 patterns obtained from% rocks under similar conditions.  The transmitted sonar signal is a% frequency-modulated chirp, rising in frequency.  The data set contains% signals obtained from a variety of different aspect angles, spanning 90% degrees for the cylinder and 180 degrees for the rock.% % Each pattern is a set of 60 numbers in the range 0.0 to 1.0.  Each number% represents the energy within a particular frequency band, integrated over% a certain period of time.  The integration aperture for higher frequencies% occur later in time, since these frequencies are transmitted later during% the chirp.% % The label associated with each record contains the letter "R" if the object% is a rock and "M" if it is a mine (metal cylinder).  The numbers in the% labels are in increasing order of aspect angle, but they do not encode the% angle directly.% % METHODOLOGY: % % This data set can be used in a number of different ways to test learning% speed, quality of ultimate learning, ability to generalize, or combinations% of these factors.% % In [1], Gorman and Sejnowski report two series of experiments: an% "aspect-angle independent" series, in which the whole data set is used% without controlling for aspect angle, and an "aspect-angle dependent"% series in which the training and testing sets were carefully controlled to% ensure that each set contained cases from each aspect angle in% appropriate proportions.% % For the aspect-angle independent experiments the combined set of 208 cases% is divided randomly into 13 disjoint sets with 16 cases in each.  For each% experiment, 12 of these sets are used as training data, while the 13th is% reserved for testing.  The experiment is repeated 13 times so that every% case appears once as part of a test set.  The reported performance is an% average over the entire set of 13 different test sets, each run 10 times.% % It was observed that this random division of the sample set led to rather% uneven performance.  A few of the splits gave poor results, presumably% because the test set contains some samples from aspect angles that are% under-represented in the corresponding training set.  This motivated Gorman% and Sejnowski to devise a different set of experiments in which an attempt% was made to balance the training and test sets so that each would have a% representative number of samples from all aspect angles.  Since detailed% aspect angle information was not present in the data base of samples, the% 208 samples were first divided into clusters, using a 60-dimensional% Euclidian metric; each of these clusters was then divided between the% 104-member training set and the 104-member test set.  % % The actual training and testing samples used for the "aspect angle