Convergence to satisfactory minima of the extended Kalman filter algorithm for supervised learning

Present training algorithms for feedforward artificial neural networks do get trapped in local minima. Some of these minima are satisfactory in terms of desired performance but many are not. When the weights converge to an unsatisfactory local minimum, the choice usually is to restart the algorithm from a different initial condition, hoping to achieve a better solution. We suggest practical ways and techniques to solve the problem of convergence to unsatisfactory local minima without the inconvenience of restarting the algorithm. A comparison of the performance of the improved algorithm with the original one is presented through computer simulations of region classification problems.