Abstract:
Simulation is a powerful tool for prescriptive analysis of queueing models. With ample data and expert knowledge of the underlying system structure, a good model can be constructed and used to predict impact of various interventions. However, such manual construction is both time- and skill-demanding. Moreover it is somewhat subjective – if the expert failed to note an important feature of the system (e.g. different customer types receiving different service priorities), the model will not be accurate.
As an alternative, we propose a data-driven representation of system building blocks, justified by the G-computation results from causal inference literature. We describe the queueing data generation process with structural equations and apply machine learning models to fit the equations to the data. Through numerical experiments, we show that this approach can replace the explicit queueing dynamics of the simulator. Our model is shown to capture the intervention effect in multi-type M/G/c queues with independent hyper-exponential service time and first-come-first-serve queueing discipline.