Closed-Loop Control Learning - Neural Networks: Methodology and Applications - page 301

Fig. 5.9. Learning a closed-loop control law by the dynamics model inversion using
recurrent back-propagation
That control methodology spurred the development of numerous practical
applications. A detailed description of an application of that technique in order
to perform the automatic piloting of a land vehicle is displayed in [Rivals 1995].
In the linear domain, predictive control is also very powerful and is com-
monly used in process technology (oil industry, etc.). It has a natural neural
extension that gave very good results in process control [Grondin 1994].
5.2.4 Using Recurrent Neural Networks
We just saw the limits of building a control law by a straightforward inversion
of the system model. Model Reference method allows taking advantage of
available heuristic or analytical knowledge about system dynamics. When no
such knowledge is available, the dynamics of the system must be taken into
account during training. That leads to using recurrent neural networks and
dynamical backpropagation as described in Chaps. 2 and 4. The diagram that
is represented in Fig. 5.9 shows how to design such a recurrent neural.
In that case, the global network is made of the neural controller and of
the system model. The input of the global network is the current state, and
its output is the state at the next time. The global network is turned into a
recurrent network by adding the state feedback to close the loop.
5.2.4.1 Using a Recurrent Neural Network to Control a Partially
Observed Dynamical System
We will end this section on neural control design by describing a major work
by industry researchers that was mentioned in previously in the context of
Kalman filters and recurrent networks [Puskorius et al. 1994]. The authors
study several examples of stabilization of nonlinear controlled dynamical sys-
tems, where the state is not completely known (partial and noisy observa-
tion). That paper presents several examples selected from real industry prob-
lems (biochemical reactor with limit cycles, rotation speed of a car engine).