The role of the interactor matrix in multivariable stochastic adaptive control (Q799644)

Many adaptive control laws and algorithms have been described by several authors. The early work on these algorithms only treated SISO systems. For this case, the system time delay emerges as a key variable. The notion of delay in the MIMO case was intimately related to the notion of the system interactor matrix. If the interactor matrix is assumed known, then it is a relatively straightforward matter to develop a globally convergent stochastic adaptive control algorithm. However, requiring knowledge of the interactor matrix is a severe practical limitation. To remove this restriction and to develop a corresponding adaptive control law is the strong motivation in this paper. The basic idea is the use of a multistep optimization horizon. The system model considered in this paper is the following ARMAX model: \[ A(q^{-1})y(t)=B(q^{- 1})u(t)+C(q^{-1})\omega (t), \] or rewritten in transfer function form \[ Y(t)=T(q)u(t)N(q)\omega (t). \] It is assumed that the system is strictly proper and that the transfer function, T(z), has full rank. Thus, there exists a matrix \(\xi\) (z), known as the interactor matrix, which is a polynomial matrix and has the following structure: \(\xi (z)=H(z)D(z)\) where \(D(z)=diag(z^{f_ 1},...,z^{f_ m})\) and H(z) is a lower triangular unimodular matrix. The author presents a suboptimal stochastic control law which does not directly depend upon knowledge of the interactor matrix for its derivation.