Accession Number : AD0657171
Title : CONVERGENCE PROOFS FOR GLUCKSMAN'S PROCEDURE.
Descriptive Note : Rept. for Dec 66-Aug 67,
Corporate Author : STANFORD RESEARCH INST MENLO PARK CALIF
Personal Author(s) : Duda,Richard O.
Report Date : AUG 1967
Pagination or Media Count : 31
Abstract : The standard error-correction procedures for training a linear machine terminate as soon as the training patterns have been separated. If the training set is small, the resulting hyperplane boundaries may be undesirably close to some of the patterns. To correct this situation, Glucksman has proposed a new training procedure which requires the boundaries to be at least some minimum distance from all of the training patterns. This report gives convergence proofs for two modifications of Glucksman's procedure. In both cases, the distance delta used during training must be less than the maximum possible distance delta m. For the tw-category case, we show that convergence is guaranteed for any delta < delta m. For the R-category case, we can guarantee convergence if delta is less than some fraction of delta m. The results of applying Glucksman's procedure to two- and ten-category problems using Highleyman's data are given. (Author)
Descriptors : (*LEARNING MACHINES, TRAINING), (*PATTERN RECOGNITION, ADAPTIVE SYSTEMS), THEOREMS, CONVERGENCE, DIFFERENCE EQUATIONS, ALGORITHMS
Subject Categories : Theoretical Mathematics
Distribution Statement : APPROVED FOR PUBLIC RELEASE