Sensitivity analysis of discrete-event systems with nonMarkovian event processes

Abstract

The authors presume the availability of a sample realization (or sample path) of a discrete-event system, possibly in real time, under some nominal parameter value(s), and they seek to estimate the sensitivity of some performance functional with respect to specified parameter(s). They do this by constructing a sample path of a system embodying perturbations of the structural parameters of the underlying transition structure of the system. Thus it is possible to evaluate the effect of various forms of supervisory control. The authors give dynamic equations describing the probability law of the process. From the sample-path perspective, the probability law between events is fully characterized by information available at the start of the corresponding interval. Moreover, several states may share the same underlying law. This fact and several generalizations allow portions of the nominal sample path to be used to construct perturbed sample paths efficiently. An algorithm is presented for such a construction, and a practically important class of systems for which it is effective is established. >