Gradient Range-Based Heuristic (GRBH) Method for Accelerating Neural Network Convergence
H. Y. Y. Sanossian, David John Evans · Integrated Computer-Aided Engineering · 1995
In this article a new learning algorithm, i.e., the gradient range-based heuristic (GRBH) method for accelerating the convergence of neural networks, is presented based on the heuristic allocation of different learning rates to different gradient ranges. Numerical results for the two intertwined sp irals problem are presented which clearly demonstrate the superiority of the new learning strategy over well-known algorithms such as back-propagation, steepest descent, and similar methods.