An alternative neural network training algorithm for real time control

Due to the advances in ANN (artificial neural networks), it is already possible to use this type of intelligent system in real-time applications. Among the applications that would benefit from this technology is process control. The long time generally required for training ANN has been a critical problem for the utilization of this technology in real-time. The generalized delta rule with backward error propagation (backpropagation) has been the most used training algorithm for ANN in control applications. However, despite many attempts to improve the performance of this learning algorithm, there is still no efficient and reliable method to train a multilayer perception. The main problem is high nonlinearity of the error function. This paper presents an alternative method for load information into a multilayer perception. The idea is to use successive quadratic approximations for the error function, until getting into the proximity of the global minimum. This technique is applied to a control system in order to obtain adaptive behaviour