Adaptive Modelling, Estimation and Fusion from Data: A Neurofuzzy Approach

Citation

Andrew, A.M. (2004), "Adaptive Modelling, Estimation and Fusion from Data: A Neurofuzzy Approach", Kybernetes, Vol. 33 No. 7, pp. 1219-1220. https://doi.org/10.1108/k.2004.33.7.1219.2

This is an account of a major development by a research group in Southampton University on the extension of adaptive techniques to non‐linear and non‐stationary environments. Adaptive control in the linear case is accepted as being completely solved. There have been previous attempts to extend the linear theory to the non‐linear case by locally‐based linearisation, but it is claimed in the Preface that these are not very useful in practice. The difficulty arises even with control situations that are readily observable and well understood and where a human operator may be successful with only experiential knowledge.

New approaches to non‐linear modelling and optimisation were opened up by work on artificial neural nets, and also by development of fuzzy techniques. The present book describes means of combining these two and so producing solutions that have been applied in fairly simple situations but that tend to become impractical on scaling‐up, an effect that has been termed the “curse of dimensionality”. Much of the book is concerned with attempts to overcome this limitation.

Both neural‐net and fuzzy techniques stem in the first place from biological observations, the former of phenomena at the single‐cell level, and the latter of human linguistic performance, where a reference to, for example, a tall man, is understood even though “tall” is an imprecise term. The combining of the two might appear to suggest that fundamental neural action had been linked to overall behaviour at the linguistic level, but of course neuroscience has not proceeded so far. The present book is not primarily concerned with neuroscience but with practical control algorithms and unquestionably demonstrates the value of the neurofuzzy approach there.

There may be interesting projects for psychologists and neuroscientists in looking for analogies with living systems, though it is easy to feel doubtful about such endeavour. The theory of artificial neural nets has moved rather far from the biological prototype, with introduction of such expedients as splines and radial basis functions, and fuzzy control theory is also rather far removed from its linguistic origins. The links to the original biological examples have become tenuous, even before the combination of these two streams under the heading of “neurofuzzy”.

The emphasis on “data fusion” refers to the ability of the adaptive controllers to combine data inputs from diverse sources. The methods are related to Kalman filtering and to support‐vector machines. There seems to be no doubt that this well‐presented book is indispensable for anyone concerned with difficult non‐linear problems of control. Reference is made to applications by the authors and their colleagues in the following areas:

• •

  submersible vehicle modelling and control;

• •

  gas turbine modelling and fault detection;

• •

  car driver and driving modelling;

• •

  obstacle detection and collision avoidance for cars;

• •

  missile tracking and guidance laws;

• •

  car engine torque modelling;

• •

  ship tracking, guidance and control;

• •

  ship collision avoidance systems;

• •

  helicopter collision avoidance system; and

• •

  processing property relationships of aerospace materials such as Al alloys.