Deep Sequence Machine Learning Models for Predicting Performance Characteristics of Retrial Tandem Queueing Systems

We consider a retrial tandem queueing system with an arbitrary number of single-server stations, phase-type distributed service times, finite buffers, and a common retrial orbit, with a Markovian arrival process. Such systems model linear-topology networks with retransmission protocols for lost packets. Exact analysis is available only for the two-station case, so larger systems rely on computationally expensive discrete-event simulation. We propose sequence-based models that learn a fixed-dimensional representation of the system’s parameters and derive the performance characteristics from it, suiting naturally ordered tandem queues whose parameter count grows with size. Experiments on a synthetic dataset show the sequence models predict aggregated characteristics no worse than fixed-dimensional methods while retaining the ability to extrapolate to larger unseen networks.