A STUDY OF CONTROL VIA ON-LINE SIMULATION USING STOCHASTIC PETRI NETS

Complex systems such as flexible manufacturing sys-tems and traffic systems typically evolve with alter-nating periods of transient and nearly steady-state behavior; such systems often show suboptimal per-formance. Thus, it is desirable to optimize the sys-tem’s performance on-line by adjusting the system’s parameters properly before a performance drop is to occur. To this end, the system’s future evolution is assessed in advance repeatedly by means of on-line simulation. However, there are several problems ac-companying this approach, particularly the demand of real-time decisions, that have not been sufficiently solved yet. Aiming at studying the dynamics of on-line control as well as its impact on the system’s operation, we built a stochastic Petri net model that simulates on-line control of a simple open queueing network as it performs by means of on-line simulation. The sys-tem under control is easy to study since it has known properties and can be considered as part of a man-ufacturing system; jobs arriving at the system have to be dispatched to one of two machines, each pro-viding a queue for jobs waiting to be processed. The processing times of the machines are deterministic or stochastic, while the jobs ’ arrival times are stochas-tic. With on-line simulation, the system’s future per-formance is assessed by virtually dispatching a new job to either of the machines, based on the system’s current state; the results are compared and thus lead to the real decision concerning to what machine the new job should be dispatched in order to minimize the work in progress. In this work, we compare the quality of on-line con-trol with that of other policies such as random choice and join the shortest queue