An Accelerated Error Back - Propagation Learning Algorithm

We propose a method for learning in multilayer perceptrons (MLPs). It includes new self-adapting features that make it suitable for dealing with a variety of problems without the need for parameter re-adjustments. The validity of our approach is benchmarked for two types of problems. The first benchmark is performed for the topologically complex parity problem with a number ofbinary inputs ranging from 2 to 7. We reduce the learning times by two to three orders of magnitude compared with conventional error back-propagation (EBP). The second problem type occurs when a high accuracy in separating example classes is needed. With classical EBP techniques and even for a one-dimensional input the learning time sharply increases with decreasing interclass Euclidean distances e according to a law of the form l/e2• Our algorithm yields substantially shorter learning times that behave as 10g(l/e). We demonstrate satisfactory learning for problems combining topological and accuracy difficulties and for which conventional EBP is practically useless ..