Modeling and on-line control of software and hardware systems
Nejib Ben Hadj-Alouane · 1995
Statistics show that the cost of control software for computer-integrated manufacturing systems and other discrete-event systems is on the rise. Some of the important reasons behind these facts are: non-reusability and inflexibility. In prior work, we developed a methodology for designing control software that overcomes the above problems. It is based on assemblages of software/hardware components, formal models, and generic controllers. Rule-based models are used for specification purposes, and event-based models are used for control purposes. Generic controllers compute the control policy on-line, while allowing for changes in the models. This thesis expands the above concepts. First, a mechanism for assembling rule-based models is provided: Models of assembly components are given in terms of constituents' models, thus reducing the specification effort. Second, the supervisory control strategy of Ramadge and Wonham is adopted. Given an event-based model of the system, its behavior is restricted within a given specification under the presence of uncontrollable and unobservable events. We develop and formally study an on-line supervisory control scheme which avoids the pitfalls of off-line methods. This scheme involves three main algorithms. The first, VLP-S, assumes total behavior observation and efficiently generates the optimal behavior (the supremal controllable sublanguage of the specification). The second algorithm, VLP-PO, assumes only partial behavior observation. Under this condition, an optimal solution (a supremal) may not exist and off-line computations involve exponential complexities. Using a priority scheme, VLP-PO generates maximal observable and controllable sublanguages with a linear on-line computational effort. By varying the event priorities, different maximals and maximals containing the supremal controllable and normal sublanguage (a benchmark) can be generated. The third, DI-VLP-PO, a distributed version of VLP-PO, computes the control action on-line using several communicating agents. This reduces the computational effort by orders of magnitude, and with reasonable communication bounds. If the system has special structure, DI-VLP-PO recovers the policies of its sequential counterpart. Extensions of the algorithms to allow for limited-lookahead, where only partial models are available, are provided (to handle infinite systems and time-constrained computation).