Learning approaches in motion control using neural networks

Summary form only given. Motion control deals generally speaking with plant modeling and controller design. Learning approaches have been developed so far, which offer novel solutions to long-standing subproblems within this field. The area of neural network learning has regained considerable interest in recent times leading to intensive research and it has become very broad, so that reviews potentially suffer from being incomplete and/or obsolete. Neural networks can be trained from examples using supervised or unsupervised (self-organization) learning techniques. For both cases of learning a considerable number of algorithms exist. For this paper the author has made a choice of method and concentrates on the following network models for supervised learning due to their solid analytical basis and their potential use for applications in motion control: multilayer networks and networks of local basis functions, which resemble solutions of surface approximation from sparse data. The following training methods are presented: backpropagation and second-order methods from the first group, and radial basis functions and wavelets from the second group