Optimal Discrete-Event Supervisory Control in a Setting of Real-Time Artificial Intelligence

Abstract Using a (homogeneous) probabilistic finite-state automaton (PFSA) model of an unsupervised plant, this paper develops optimal supervisory control systems in a setting of real-time artificial intelligence (AI), based on a real signed measure of regular languages. The underlying concept allows on-the-fly analysis, synthesis, and execution of optimal discrete-event supervisory (DES) control policies by selective manipulation of controllable and observable events to elementwise maximize real signed language measure vectors of the supervised plant. In contrast to the seminal work of Ramadge and Wonham, this language measure allows total ordering on a set of partially ordered sublanguages of regular languages of the supervised plant for a (quantitative) performance evaluation and comparison of different supervisors; and reward/penalty characterization of the individual PFSA states is used to construct the performance cost functional to be maximized. The optimal DES control system yields superior performance in the sense that the controlled plant is more likely to reach good (i.e., desirable) states and less likely to terminate at bad (i.e., undesirable) states than other DES control systems. This optimal DES control algorithm requires at most (2n+1) iterations, and its computational complexity is polynomial in n, where n is the number of states in the plant PFSA model; this property suggests that the DES control concept would be ideally suited for real-time AI operations, such as those built upon a stimulus-response procedure. The proposed control concept has been validated on a discrete-event model of an airborne twin-engine unmanned autonomous vehicle (UAV).