Closed-Loop Control Learning - Neural Networks: Methodology and Applications - page 297

Fig. 5.3. Trajectory of the stabilized system with a noise factor of 0.5. (Straightfor-
ward model inversion control learning with larger weighting of the speed deviation
in the cost function)
are necessary in order to obtain satisfactory results. We are going to review
some of those improvements.
5.2.2 Model Reference Adaptive Control
In the method called MRAC, for model reference adaptive control, prior
knowledge of the system is used to build the control law [Rivals et al. 2000]. In
that method, the cost function does not push the system directly to the desired
target, but it is designed to force the closed-loop controlled system to track
a reference trajectory. The choice of that reference trajectory is made from
prior knowledge of the controlled system, especially on the actuator abilities.
Actually, there is always an implicit reference model; in the straightforward
inverted control of the previous section, the reference model is just a single
time lag.
Figure 5.5 shows the general organization of the training process of a neural
controller with a reference model.
The Reference Model method turns out to be valuable in numerous appli-
cations to real world problems. It is generally used to improve the performance
of classically controlled dynamical systems. Whenever possible, the reference
trajectory is chosen as the trajectory of a linear system with a critical damping