Combining neural and conventional paradigms for modelling,prediction and control

Abstract

Whereas some researchers in the control community see neural networks as a panacea for solving all realistic control-related problems, another school rejects the neural-network paradigm altogether in favour of well-established conventional schemes with a sound theoretical basis. Much of the promising research using neural networks for modelling, prediction and control, however, realizes the value of, and the need for, both paradigms, and exploits their mutual complementarity to address realistic problem situations. This work develops a general qualitative framework for identifying the possible ways of combining the universality of neural networks with the prior knowledge and experience embedded in the available physical models, model-based estimators, conventional controllers, as well as vague linguistic descriptions provided by human experts about the system behaviour and manual control strategies. The presented framework not only naturally leads to the previously proposed schemes in the literature, but also reveals several new possibilities.