RBFNN for Real-Time Process Identification and Control with Selective Forgetting

This work is concerned with the problem of the real-time identification and control of dynamic non-linear systems using a neural network approach. The chosen model was the Gaussian Radial Basis Function Neural Network (RBFNN) type, due to its universal approximation property and also to the fact that the parameters are linearly related to the outputs allowing linear learning algorithms. A hybrid learning technique is developed, using an adaptive learning rate with process monitoring, and taking advantage of the locality property of this type of networks, we apply a selective forgetting algorithm. This on-line learning is used both for identification and control. A novel technique is proposed for the on-line adaptive control. The potential of the proposed method is demonstrated by a simulation example applied to a theoretical model and to a real laboratory process.