New Learning Method for Neural Network by Tabu Search Method.
Shin MORISHITA, Soichiro Une, Chinmoy Pal, Ichiro Hagiwara · TRANSACTIONS OF THE JAPAN SOCIETY OF MECHANICAL ENGINEERS Series C · 1994
Avoiding wasted time for learning and trapping at a local optimum are two important problems in learning of neural network. The tabu search method with random waves, to solve global optimization of continuous variables is introduced, and a new method which combines this and the steepest descent method is proposed. The reliability and effciency of the tabu search method and the new method are examined with the help of the one-dimension test function. They outperform the steepest descent method in the search of the global optimum. Using these methods as learning algorithm of a neural network, they are applied to examples of an exclusive OR problem, off-line identification of a dynamic system and on-line identification of a multi-degree system (cantilever) by simulation. In each case, it is shown that the neural network can learn quick and effectiveIy.