Neural Networks: An Overview - Neural Networks: Methodology and Applications - page 76

The application requested the design and implementation of two control
systems that must fulfill two tasks,

• the control of the driving wheel, in order to keep the vehicle on the desired

trajectory: a neural controller was designed, that performs a maximum
lateral error of 40 cm, for curvatures up to 0.1 m − 1 , and lateral slopes
up to 30% in rough terrain; some elements of that controller used semi-
physical modeling;

• the control of the throttle and the brake, in order to comply with the

desired velocity profile.

All neural networks implemented within that application, whether models
or controllers, are very parsimonious (less than ten hidden neurons). Their
implementation on board did not require any special-purpose hardware: they
were implemented as software on a standard microprocessor board that was
also used for other purposes.

1.5 Conclusion

In the present chapter, we endeavored to explain why, and for what purposes,
neural networks can be advantageously used. Some typical applications were
presented (others are described in various chapters), so that model designers
can get an intuition of what they can expect from that technique.
Before proceeding to more mathematical topics, it may be useful to em-
phasize the main points that should always be kept in mind when designing
neural networks, i.e.,
•

Neural networks are machine learning tools, that allow to fit very general
nonlinear functions to sets of experimental data; just as for any statistical
method, the availability of appropriate data is mandatory.

• Neural networks with supervised learning are parsimonious approxima-

tors, that can serve as static models (feedforward neural networks) or as
dynamic models (recurrent neural networks).

• Neural networks with supervised learning can be high-quality classifiers,

whose performances can reach those of the theoretical Bayes classifier;
however, in the framework of classification for pattern recognition, the
representation of the patterns to be recognized is often crucial for the
performance of the whole system; in that context, neural networks with
unsupervised learning may provide very valuable information for designing
an e cient data representation.

• it is generally desirable, and often possible, to take advantage of all existing

mathematical knowledge on the process to be modeled or patterns to be
classified: neural networks are not necessarily black boxes.
The next chapters provide the mathematical background and the algorithmic
information that are necessary for an e cient design of neural network models.