A First Course in FUZZY and NEURAL CONTROL

6.5. NEURAL NETWORKS IN INDIRECT NEURAL CONTROL 217

This is similar to standard control in that a mathematical model is needed, but
here the mathematical model is a neural network.
Indirect neural control designs involve two phases. Thefirst phase consists
of identifying the plant dynamics by a neural network from training data ñ
that is, system identification. In the second phase, the control design can be
rather conventional even though the controller is derived, not from a standard
mathematical model of a plant, but from its neural network identified model.
Since the identified neural network model of the plant is nonlinear, one way
to design the controller is to linearize its identified neural network model and
apply standard linear controller designs. Another way is through ìinstantaneous
linearization,î as described in Section 6.5.3.

6.5.1 System identification.....................

Thesystemidentification phase can be carried out by using neural networks.
Recall that a controller and the plant that it controls bear an inverse relation-
ship, and that a primary goal in the development of an autonomous system is to
build a controller whose behavior bears an inverse relationship to the plant. In
particular, system identification is necessary for establishing a model on which
the controller design can be based. The general problem of system identifica-
tion of nonlinear systems is daunting and, lacking sufficient theoretical results,
much needs to be done on a case-by-case basis. It is one of the areas where the
application of neural networks is promising.
The followingfigure depicts the general problem of system identification of
nonlinear systems. The parameters of the identification model are estimated as
the model changes over time, so that the difference between the plant output
and the model output is minimized.

Figure 6.22. System identification model

Theideaisfortheidentification process to produce a model of the dynamical
system with no prior knowledge of the dynamics of the system. This is referred
to asblack box modeling. The learning algorithm that is used to train the
network is commonly a version of backpropagation, as discussed in Section 5.5.
Of course, the identification model may be reliably trained only for the data it
experiences and may not produce the desired output for new data. However,