Deep learning methods for transient performance analysis and optimization of non-Markovian nonstationary queueing systems

(English) Empirical studies show that real-world service queueing systems, such as contact centers and healthcare facilities, often exhibit non-Markovian behavior, including non-exponential service and abandonment times, as well as nonstationary dynamics driven by time-varying arrivals and staffing....

Descripción completa

Detalles Bibliográficos
Autor: Garyfallos, Spyridon
Tipo de recurso: tesis doctoral
Fecha de publicación:2026
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/457092
Acceso en línea:https://hdl.handle.net/2117/457092
https://dx.doi.org/10.5821/dissertation-2117-457092
Access Level:acceso abierto
Palabra clave:non-Markovian queues
nonstationary queues
optimal staffing in queues
transfer learning
deep learning
simulation
queueing theory
004 - Informàtica
Àrees temàtiques de la UPC::Informàtica
Descripción
Sumario:(English) Empirical studies show that real-world service queueing systems, such as contact centers and healthcare facilities, often exhibit non-Markovian behavior, including non-exponential service and abandonment times, as well as nonstationary dynamics driven by time-varying arrivals and staffing. These complexities make performance analysis and capacity planning particularly challenging. Traditional methods face notable limitations: Monte Carlo simulation, while flexible, can be computationally intensive for large systems or real-time decisions; and heavy-traffic analysis is a model-specific approximation that applies only to certain queueing structures, depends on the service-level metric of interest, and its accuracy may deteriorate for small-scale systems. To address these challenges, this thesis develops machine learning frameworks for analyzing and optimizing non-Markovian, nonstationary queueing systems. Trained on simulated data, the models learn the spatio-temporal structure of complex service systems and provide extremely fast inference, enabling real-time decision support. They are effective and robust across problem scales, and their design and training are universal, in contrast to heavy-traffic formulas that must be re-derived for each model and performance measure. By integrating accurate prediction with real-time optimization, this work advances queueing theory by introducing modern data-driven tools for dynamic capacity management in complex service operations. The thesis consists of three main contributions: (1) a neural network framework for performance analysis of non-Markovian nonstationary queueing systems with fixed capacity; (2) a deep learning framework based on a transfer learning approach that enables model adaptation to queueing systems with time-varying server capacity; (3) an optimization framework for staffing in non-Markovian nonstationary environments. All of these methods are accompanied by accuracy, robustness, and computational runtime experiments, demonstrating the effectiveness of the proposed approaches compared to traditional methods.