204 CHAPTER 6. NEURAL CONTROL
When the inverse dynamics of a plant exist, one can try to control the plant
by modeling its inverse dynamics.
6.3 Neural networks in direct neural control
As suggested in Section 6.2, suppose we have a plant whose inverse dynamics
exists but does not have a closed form. Approximating this inverse dynamics
provides a way to control the plant. Of course, approximations of functions
(control laws) can be done in many ways. When we use neural networks for
modeling inverse dynamics, we are designing direct neural controllers.
Direct designmeans that a neural network directly implements the control-
ler ñ that is, the controller is a neural network (see Figure 6.1). The network
must be trained as the controller according to some criteria, using either nu-
merical input-output data or a mathematical model of the system.
Figure 6.1. Direct designA natural question that arises in this type of neural control is the selec-
tion of the type of neural network needed for the controller. We have seen
from the previous chapter on neural networks that there are several types of
neural network architectures. Multi-layered perceptron (MLP) neural networks
are composed of configurations of simple perceptrons in a hierarchical structure
forming a feedforward network. They have one or more hidden layers of per-
ceptrons between the input and output layers. It is permissible to have any
prior layer nodes connected to subsequent layer nodes via a corresponding set
of weights. Different learning algorithms can be used for MLPs, but the most
common ones have been the delta rule and error-backpropagation algorithms
discussed previously. These algorithms do work fairly well but they tend to be
slow. Faster and more efficient algorithms have been developed [8, 20, 32, 37],
and ongoing research is continually discovering further improvements.
6.4 Example: temperature control
In Section 2.7.2, we discussed the development of the classical proportional-
integral-derivative (PID) control parameters for a temperature control problem.
Here we extend this example to incorporate a neural controller. In a conven-
tional PID control system, the gains are allfixed; while with neural networks,
they can change.