Robust control using inverse dynamics neurocontrollers
Csaba Szepesvári and András Lõrincz
Nonlinear Analysis, Theory, Methods & Applications
30,
1669--1676
(1997)
Abstract
Neurocontrollers typically realise static state feedback control
where the neural network is used to approximate the inverse
dynamics of the controlled plant. In practice it is often
unknown a priori how precise such an approximation can be.
On the other hand, it is well known that in this control
mode even small approximation errors can lead to instabilities.
The same happens if one is given a precise model of the
inverse dynamics, but the plantīs dynamics changes. The
simplest example of this kind is when the robot arm grasps
an object that is heavy compared to the arm. This problem
can be solved by increasing the stiffness of the robot, i.e.,
if one assumes a "strong" controller. Industrial controllers
often meet this assumption, but recent interest has grown
towards "light" controllers, such as robot arms with air
muscles that can be considerably faster. There are well-known
ways of neutralising the effects of unmodelled dynamics, such
as the sigma-modification, signal normalisation, (relative)
dead zone, and projection methods, being widely used and
discussed in the literature. Here a novel architecture
that does direct identification of the inverse dynamics
and a new method that utilizes this inverse dynamics controller
in two copies are desribed. The result is robust controller
of high precision put on a firm mathematical bases. The
capabilities of the controller will be demonstrated on a
chaotic bioreactor. The attractive learning properties will
be discussed.
