Nonlinear identification of a two link robotic system using dynamic neural networks

The aim of this work is the identification of a two link robotic system using dynamic neural networks. The identified system was a mechanical leg formed by two revolute links. The theoretical model of the system corresponds to a two-link kinematic chain. However, a number of nonlinearities which are present in the real system are difficult to include in the theoretical model. An empirical model of the system was obtained instead using dynamic neural network. New training and validation techniques that assure a good performance of the empirical model have been applied. The consist of including the initial stae of the hidden neurons in the decision vector associated with the optimisation problem that is solved for training the network. Once a network has been suitably trained, the initial states of the hidden neurons are also optimised based on the validation data.