Learning about dynamical systems via unfalsification of hypotheses

This paper examines the problem of learning behaviours of a dynamical system from experimental data via unfalsification of hypotheses within the behavioural approach to system theory of Willems. Behaviours of the dynamic systems are postulated as hypotheses and then tested against experimental data. A simple and concise condition for falsification of hypotheses by experimental data in terms of a kernel is presented. The approach is applicable both to learning models for a plant and to adapting controllers to satisfy performance and robustness goals. Copyright © 2004 John Wiley & Sons, Ltd.