Simulation-based predictive analytics for dynamic queueing systems

Simulation and simulation optimization have primarily been used for static system design problems based on long-run average performance measures. Control or policy-based optimization has been a weakness, because it requires a way to predict future behavior based on current state and time information. This work is a first step in that direction with a focus on congestion measures for queueing systems. The idea is to fit predictive models to dynamic sample paths of the system state from a detailed simulation. We propose a two-step method to dynamically predict the probability that the system state belongs to a certain subset and test the performance of this method on two examples.