A sliding window solution for the on-line implementation of the Levenberg-Marquardt algorithm

The Levenberg-Marquardt algorithm is considered as the most effective one for training Artificial Neural Networks but its computational complexity and the difficulty to compute the trust region have made it very difficult to develop a true iterative version to use in on-line training. The algorithm is frequently used for off-line training in batch versions although some attempts have been made to implement iterative versions. To overcome the difficulties in implementing the iterative version, a batch sliding window with Early Stopping, which uses a hybrid Direct/Specialized evaluation procedure, is proposed and tested with a real system. © 2004 Elsevier Science Lta. All rights reserved.