I denne forelesningen vil black-box modellering (f.eks. ved bruk av
nevrale nettverk) bli gjenstand for en kritisk vurdering. Kritikken
vil bli framfoert baade ifra et analytisk perspektiv og ifra et
anvendelsesperspektiv. Et alternativ til rene black-box modeller vil
bli diskutert, og noen anvendelser fra Daimler-Benz blir
gjennomgaatt. Mer info. nedenfor.
Moet opp!
Bjarne Foss
------------------------------------------------------------------
"APPLICATIONS OF CONSTRUCTIVE MODELLING AND CONTROL"
Dr Ken Hunt, Daimler-Benz AG, Systems Technology Research Berlin
In the realm of dynamic systems modelling and control,
neural networks may reasonably be viewed as `the ultimate
black-box model'. A great diversity of neural network
structures have been proposed, but in the main they can
be seen simply as very rich nonlinear parameterisations.
The major neural models (e.g., the multi-layer perceptron
and radial basis-function networks) are all supported,
under appropriate preconditions, by the uniform approximation
theorem. Given sufficiently rich training data and sufficient
computing power and time, it is sometimes very
easy to set up a model and to obtain reasonable modelling
results. Moreover, this can be achieved with very little in the
way of a priori knowledge about the specific problem being
studied, and without regard to existing systems theory.
These facts are frequently cited by proponents of the
ultimate black-box.
We argue, however, that this point of view can in general
be very misleading, particularly for problems involving
complex dynamics; analysis of the properties of trained neural
networks is often not possible and it is difficult to
make concrete analytical statements about stability and
robustness. Typical structures do not admit the direct
incorporation of a priori knowledge - such knowledge typically
abounds in engineering problems. We argue that in complex
problems success hinges on the ability to incorporate
a priori knowledge in various forms and to directly
analyse the resulting model. It is also our experience
that many neural solutions are strongly overparameterised
and that comparable results could be achieved by careful
application of `traditional' methods.
In this presentation we pursue transparent model architectures
which have the following features: they should allow smooth
integration of existing models and information, be
amenable to analysis, and have structures which can be directly
exploited as the basis of control designs. We focus
on "local model networks".
Local model networks are based on smooth interpolation of
locally valid models (of various types) and can be used
as the basis of an interpolating network of local
controllers.
These constructive techniques have been extensively
applied and implemented within Daimler-Benz on a wide
range of modelling and control problems. This includes
automotive problems (e.g., gearbox control, speed and
distance regulation, and autonomous steering),
aerospace problems (e.g., satellite attitude, helicopter
dynamics), process control (e.g., steel rolling, induction
motor control) and bioengineering (e.g., muscle modelling
and control). A selection of these problems will be
used to illustrate the key points of this presentation.