2015 年 51 巻 7 号 p. 451-457
This paper proposes a data-driven controller design method for minimum variance control and analyzes convergence properties of the method. Data-driven controller design methods tune control parameters by minimizing a criterion which is derived from a single set of input and output data without a process model. However, optimization problems in these methods are generally non-convex. The analytical results show that a gradient descent algorithm can converge from the set of initial parameter values to the global minimum if the maximum-phase difference between the initial controllers and the minimum variance controller is smaller than π/2 radians.