All-purpose controllers in stochastic problems of optimal control of linear stationary objects (Q1281032)

The author considers the problem of optimal control of linear stationary objects in the presence of a stochastic stationary action \[ \dot x=Ax+Bu+f^0\varphi(t),\quad y=Cx,\quad t\geqslant 0, \] under the assumptions, that the matrix of spectral densities \(S(\theta)\) for external actions is unknown. It is only assumed that the upper bound \(\sigma(\theta)\) of its norm \((\|(S(\theta)\|\leqslant\sigma(\theta))\) is known and also that \(\sigma(\theta)\) decreases rapidly (for example, exponentially) when \(\theta\to\infty.\) A controller that minimizes a given quadratic functional is required to be found. It is shown that an optimal controller generally does not exist in the natural class of feasible controllers. Because of this, the problem of finding an \(\varepsilon\)-optimal controller (with any given \(\varepsilon > 0\)) for which the quality functional differs from the lower bound by no more than \(\varepsilon\) is stated. The functional itself and its lower bound throughout the class of feasible controllers are dependent on the unknown spectral density \(S(\theta).\) Therefore, it would be natural to expect that, in the general case, the \(\varepsilon\)-optimal controller depends on \(S(\theta),\) too. It is shown, however, that, in this problem, among an aggregate of \(\varepsilon\)-optimal controllers (dependent on \(S(\theta)\)), an ``all-purpose'' \(\varepsilon\)-optimal controller can be found under certain assumptions which is independent of \(S(\theta),\) and that this assertion is valid for any \(\varepsilon > 0.\) This controller can be employed in cases where \(S(\theta)\) is unknown. When an object is subject to any external action from the wide class mentioned previously (in this case, the value of the quality functional is, in general, dependent on \(S(\theta)\)), such a controller gives an ``almost'' optimal value to the quality functional. Apart from a number of natural and technical conditions, the major condition for an all-purpose \(\varepsilon\)-optimal controller to exist is the requirement that the number of (scalar) independent inputs of a controller be no less than the number of (scalar) external actions. If this condition is not satisfied, then, as a rule, the``all-purpose'' controller mentioned above does not exist. This situation is much different from the Wiener case where the function \(S(\theta)\) is a rational one. In the Wiener case, an all-purpose controller (or an ``all-purpose'' controller with any \(\varepsilon > 0\)) does not exist, as a rule.