A new formulation of the learning problem of a neural network controller

The authors consider the learning problem for a class of multilayer perceptrons, which is particularly relevant in control systems applications. By reformulating this problem, a criterion is developed which reduces the number of iterations required for the learning phase. A Jacobian matrix is proposed, which decreases the computational complexity of the calculation of derivatives. Experimental results showed that this approach also yields, in comparison with existing methods, a faster rate of convergence, therefore achieving a significant reduction in computing time.<
>