確率パラメータをもつサンプル値系の学習制御

抄録

In recent years many literatures about controls have been presented. But scientific forms are not yet completed. This paper treats the problem about a learning control which minimizes the conditional expectation of quadratic form of state vector and control, when correlated random parameters are involved in the control systems and the control has no constraint.
Two cases are considered as follows.
(1) Random parameter takes two values.
(2) The variation of random parameter is Gaussian.
We analyze the first case by using Bayesian theorem and the second case by using a mean square error theorem.
By predicting future fluctuations of random parameters from their measured past values, we can get the proper control policies.
Numerical example shows that the value of performance index with learning is better than the one without learning.