Title: Expanding Causal Metamodeling to Non-Markovian Queues

Abstract: Metamodels serve as approximations for discrete-event simulations, allowing researchers to estimate system behavior without incurring the high computational costs of running full simulations. While previous research established modular dynamic Bayesian networks (MDBNs) as a powerful class of metamodels capable of addressing a variety of probabilistic and causal queries (PCQs) through a single trained model, this methodology was previously restricted to Markovian systems. This study introduces an extension of MDBNs to non-Markovian queues, utilizing phase-type distributions to approximate non-exponential distributions. This adaptation introduces several new challenges, such as determining the optimal number of phases to balance accuracy with tractability, efficiently learning the metamodel’s parameters, and selecting the appropriate sampling interval to convert a continuous-time simulation into a discrete-time MDBN. We offer preliminary solutions to these issues, marking the first application of causal metamodeling techniques to non-Markovian systems. Our experiments on a G/M/1 queue demonstrate that the MDBN approach yields accurate PCQ responses while achieving inference speeds that are orders of magnitude faster than direct simulation.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC