Controller validation (Q2763360)

This paper is concerned with controller validation or, as otherwise called, controller unfalsification. The study of the unfalsified control provides a theoretical basis for directly validating closed-loop controller performance from open-loop experimental data, or even from closed-loop data obtained during the use of another controller in the feedback loop. The paper expands this theory through a study of the connections between model and controller validation. It is shown that controller validation is a feedback generalization of model validation and that it is less conservative to unfalsify (validate) control laws directly rather than unfalsify (validate) the plant models from which those control laws are constructed. The approach of the paper allows controller validation to be accomplished without making any prior assumptions about the plant or the noise. Thus the conservatism inherent in indirect two-step validation approaches that require one to first identify a plant model and uncertainty bounds from experimental data before analyzing control system robustness is avoided.