On-line training of neural networks: a sliding window approach for the levenberg-marquardt algorithm

  • Authors:

  • Fernando Morgado Dias;Ana Antunes;José Vieira;Alexandre Manuel Mota

  • Affiliations:

  • Departamento de Engenharia Electrotécnica, Escola Superior de Tecnologia de Setúbal do Instituto Politécnico de Setúbal, Setúbal, Portugal;Departamento de Engenharia Electrotécnica, Escola Superior de Tecnologia de Setúbal do Instituto Politécnico de Setúbal, Setúbal, Portugal;Departamento de Engenharia Electrotécnica, Escola Superior de Tecnologia de Castelo Branco, Castelo Branco, Portugal;Departamento de Electrónica e Telecomunicações, Universidade de Aveiro, Aveiro, Portugal

  • Venue:
  • IWINAC'05 Proceedings of the First international work-conference on the Interplay Between Natural and Artificial Computation conference on Artificial Intelligence and Knowledge Engineering Applications: a bioinspired approach - Volume Part II
  • Year:
  • 2005

Quantified Score

Hi-index 0.00

Visualization

Abstract

In the Neural Network universe, the Backpropagation and the Levenberg-Marquardt are the most used algorithms, being almost consensual that the latter is the most effective one. Unfortunately for this algorithm it has not been possible to develop a true iterative version for on-line use due to the necessity to implement the Hessian matrix and compute the trust region. To overcome the difficulties in implementing the iterative version, a batch sliding window with Early Stopping is proposed, which uses a hybrid Direct/Specialized evaluation procedure. The final solution is tested with a real system.