Discrete event-based cost criteria for controller parameter optimization

In the field of control engineering there are several methods to be adopted for iterative optimization of the controller parameters. The key issue in the optimization procedure is to define a suitable cost criterion which depends on the parameters. However, the metrics typically used in derivation of the cost criterion often require continuous, or at least, discrete measurements with a constant sampling rate. This paper presents cost criteria for controller parameter optimization which do not require continuous measurements from the response nor the control signal. Moreover, the approach does not assume any prior structure about the system. Only the measured output signal is modeled by assuming certain types of modes which can appear. The input vectors for the criteria consist of discrete events. The criteria are based on modeling the step response by training a hidden Markov model (HMM) by the discrete event vectors to evaluate the performance of the system. The states of the HMM are defined explicitly to correspond the abnormalities, for example, oscillation, which may occur in the response. The dependencies between the cost criteria and the controller parameters are studied by a simple case example.