« All Events

Applied Probability and Queues

Oct 5, 09:30 - Oct 7, 16:30

We are pleased to announce a three-day workshop on Applied Probability and Queueing Theory to be held at TU/e, Eindhoven, The Netherlands, from October 5 to 7, 2026.

The program will bring together long-standing collaborators of EURANDOM alongside both established and emerging researchers in applied probability and queueing theory. It will provide a forum for the exchange of recent theoretical advances, methodological developments, and applications across a broad range of stochastic models, including queues, stochastic networks, and related systems. The workshop aims to foster new connections, strengthen existing collaborations, and stimulate cross-fertilization between different generations and perspectives within the field.

The second day of the workshop will be dedicated to mark the retirement of Offer Kella, a long term visitor and collaborator of EURANDOM. This day will feature talks by colleagues, collaborators, and former students to celebrate his profound contributions to applied probability, Lévy processes, queueing theory, and stochastic networks, and to EURANDOM.

Speakers

E-mails on travels and/or hotels will only come from Eurandom. Please ignore e-mails from GT OPS-travel or other travel companies.

Program (tentative): Program_APQ_ v31072026

On behalf of the organizing committee:

 

Registration for participants:

You can register for the workshop using the form below. Coffee / tea and lunch are included in the registration fee.

After registering, you will immediately receive an automatic confirmation of your registration (check your spam folder!). Your proof of payment will be sent separately, in the week following the workshop.

Registration closes on Monday 28 September 2026.













Total €

 

=============================================================

Abstracts: 

Andreas Löpker

Title: The Defective Geometric Distribution and Minification

Abstract: Geometrically distributed random variables with extra mass at zero and infinity have some nice properties. For example, their probability generating functions can be represented by 2×2 matrices with determinant

1 in a way that matrix multiplication is associated with a random sum (or, in other words, a branching mechanism). It turns out that the pgf of the minimum of such an extended geometric random variable and any other N_0-valued random variable also has a tractable form. The talk is based on joint work with Offer Kella and Christian Weiß.

 

Antonio Castellanos 

Title: What You Say Versus When You Say It: Efficiently Predicting Service Completions with LLMs and Stochastic Processes

Abstract: Accurately determining whether an ongoing service conversation will continue or has effectively ended is central to operational control in chat-based service systems. Closing too early can trigger costly customer recontacts; waiting too long wastes agent capacity. Using large-scale data from a food-delivery service organization, we study two prediction tasks: short-term conversation continuation and end-of-service completion. We compare stochastic self-exciting point processes, metadata-based neural networks, and large language models (LLMs) using full conversation text. The results reveal a sharp task distinction. For short-term continuation, lightweight Hawkes models capture temporal dynamics effectively, delivering competitive accuracy at minimal computational cost. For end-of-service prediction, semantic information is crucial: a fine-tuned Transformer text classifier outperforms temporal models, whereas zero-shot prompted LLMs perform poorly despite their scale. A cost-accuracy analysis shows that large models are justified only for selected high-stakes decisions. We therefore propose a hybrid approach combining real-time Hawkes monitoring with selective LLM queries.

Joint work with Andrew Daw and Galit B. Yom-Tov.

Arnoud den Boer

Title: How long does it take to sell your products?

Abstract: A price-setting seller offers q units of inventory over a finite selling horizon of length T. Prices are set either according to the optimal static policy, or the optimal dynamic policy (as in Gallego and Van Ryzin, 1994). As T grows large, what is the distribution of the time it takes to sell k items? Is static pricing systematically slower or faster than dynamic pricing? And how does this all depend on characteristics of the willingness-to-pay distribution? In this joint work with Tarek Abdallah of Northwestern Univ., we address these questions.

Bernardo D’Auria 

Title: Exit Problems for Markov-Modulated Additive-Increase Multiplicative-Decrease Processes

Abstract: We study one-sided exit problems for a class of additive-increase multiplicative-decrease (AIMD) processes evolving in a finite-state Markovian environment. The process increases linearly at a regime-dependent rate and undergoes downward multiplicative jumps at the arrival times of a Markov-modulated Poisson process. For both upward and downward crossing times, we derive explicit expressions for the Laplace-Stieltjes transforms of the hitting times jointly with the modulating state at exit. The analysis leads to matrix-functional equations for the transforms, which are solved through an iterative construction involving regime-indexed products and a matrix series representation. The resulting formulas provide a tractable characterization of exit distributions and extend classical fluctuation-type identities to an AIMD setting. The techniques developed here may provide analytical tools for the study of systems with linear workload growth and multiplicative reductions, should such dynamics arise in queueing or related models.

Joint work with Zbigniew Palmowski.

Caroline Jagtenberg 

Title: Spatial Queues with Moving Servers: Closing the Gap Between Hypercube and Simulation

Abstract: Spatial queueing systems with moving servers remain an open problem. In emergency medical services (EMS), the standard tool – Larson’s Hypercube model – captures exact dependence between servers but scales poorly beyond a handful of vehicles. Its widely used scalable approximation instead replaces exact joint state probabilities with independence-based correction factors, an assumption known to degrade when servers are co-located at the same base. Moreover, there are no transient versions of these approaches, even though time-varying arrival rates make transient behavior central to EMS operations.

We present a computational model based on QPLEX, that captures within-base vehicle dependence and transient dynamics directly, without relying on this independence assumption, while remaining computationally tractable as the number of bases and vehicles per base grows. We benchmark QPLEX against discrete-event simulation (DES) and the Hypercube model along two dimensions: computational cost and realism. QPLEX matches Hypercube’s tractability at large scale while retaining much of DES’s realism, at a fraction of DES’s computational cost.

Across a range of realistic instances, QPLEX tracks DES’s transient busy-fraction dynamics closely while running far faster, and keeps Hypercube-level cost as the problem size grows. This speed also makes QPLEX practical to embed inside an optimization loop; we report early experiments exploring how many vehicles to station at each base.

Efrat Perel

Title: “Exact Analysis of Join the Shortest Queue with Serve the 2-Longest Queue Policy”

Abstract: We study a two-queue polling system where arriving customers join the shorter queue while a single server switches to the other queue when the queue length difference reaches 2. We derive the probability generating functions for the steady-state distribution and calculate key performance measures. Numerical comparisons with the previously studied threshold of 1 demonstrate substantial improvements: significant reduction in expected queue lengths and 15-30\% reduction in switching overhead. That is, tolerating larger queue imbalance before switching outperforms immediate response.

Joint work with Nir Perel

Eliran Sherzer

Title: Making matrix-geometric great again with a novel PH/MAP fitting method.

Matrix-analytic methods provide some of the most powerful tools for the exact analysis of queueing systems, but their applicability depends on obtaining accurate Phase-Type (PH) distributions and Markovian Arrival Processes (MAPs) that faithfully represent real data. In practice, fitting such models remains one of the main obstacles to applying these techniques to empirical systems.

In this talk, we present a gradient-descent framework for learning PH distributions and MAPs directly from data. The approach employs an unconstrained differentiable parameterization that guarantees valid stochastic models while allowing optimization with standard deep-learning tools. Unlike likelihood-based methods, the framework directly optimizes distributional shape and, for MAPs, temporal dependence, enabling accurate representations of both marginal behavior and arrival correlations.

We demonstrate that the proposed PH fitting method substantially improves fitting accuracy over existing expectation-maximization approaches while maintaining practical runtimes. We further show how the same optimization framework extends naturally to MAP fitting. Together, these methods provide a practical pipeline for converting empirical traces into analytically tractable stochastic models that can be immediately used in matrix-geometric and other matrix-analytic algorithms for the performance evaluation of queueing, inventory, reliability, and communication systems. Matrix-analytic methods provide some of the most powerful tools for the exact analysis of queueing systems, but their applicability depends on obtaining accurate Phase-Type (PH) distributions and Markovian Arrival Processes (MAPs) that faithfully represent real data. In practice, fitting such models remains one of the main obstacles to applying these techniques to empirical systems.

In this talk, we present a gradient-descent framework for learning PH distributions and MAPs directly from data. The approach employs an unconstrained differentiable parameterization that guarantees valid stochastic models while allowing optimization with standard deep-learning tools. Unlike likelihood-based methods, the framework directly optimizes distributional shape and, for MAPs, temporal dependence, enabling accurate representations of both marginal behavior and arrival correlations.

We demonstrate that the proposed PH fitting method substantially improves fitting accuracy over existing expectation-maximization approaches while maintaining practical runtimes. We further show how the same optimization framework extends naturally to MAP fitting. Together, these methods provide a practical pipeline for converting empirical traces into analytically tractable stochastic models that can be immediately used in matrix-geometric and other matrix-analytic algorithms for the performance evaluation of queueing, inventory, reliability, and communication systems.

Joint work: Yehezkel Resheff and Miklos Telek.

Fiona Sloothaak:

Title: Threshold-based routing policies in skill-based queues

Abstract:Threshold-based routing policies are among the simplest and most practical control mechanisms in queueing systems, yet many fundamental questions remain surprisingly open. In this talk, we consider two such questions in the context of skill-based queues. First, we study deterministic waiting-time thresholds, where additional compatibility edges become available only after customers have waited sufficiently long. Although it is intuitively clear that such thresholds should not affect stability, proving this turns out to require a careful Piecewise Deterministic Markov Process formulation together with a Lyapunov-based analysis that explicitly accounts for the threshold structure. Second, we consider queue-length thresholds as a routing mechanism to balance the trade-off between matching quality and delays. We present threshold-based priority policies that achieve routing rates arbitrarily close to an optimal linear-program benchmark while maintaining explicit queue-length guarantees, matching the optimal heavy-traffic scaling in the models considered. A central technical challenge is that our routing induces Markov-modulated arrival processes, leading to state-dependent drift terms that cannot be analyzed directly. We show how these terms can be controlled by augmenting the Lyapunov function with a correction obtained from a Poisson equation, allowing the state-dependent dynamics to be replaced by their steady-state averages up to a controllable error.

Galit Yom-Tov:

Title: Using Hawkes processes to model service processes: Theoretical development, empirical evidence, and operational implications

Abstract: Services are characterized by the interaction between two (or more) people: together, the customers and agents co-produce the service. Specifically, the success of a service interaction relies on contributions from the agent and the customer, and these two sides of the interaction rely on one another to complete the service. The customer depends on the superior technical knowledge, authority, or resources available to the agent, and the agent draws upon the customer’s ability to clearly articulate their problem and provide relevant information throughout. In a series of papers, we show that marked multi-variant Hawkes cluster processes fit very well a variety of service processes, such as telecommunication contact center interaction that lasts around an hour and a complex court case that progress over several years.

This modeling of the service process allows us to (a) dynamically predict agent workload to improve routing procedures; (b) investigating behavioral aspects of how actors and activities influence the service process dynamics; and (c) asking strategic questions about the system design.

Joint work with Andrew Daw, Antonio Castellanos, Jamol Pender, and Galit Kadzelashvily

Jan-Pieter Dorsman

Title: Modelling skill-based queueing systems using pass-and-swap queues and Whittle networks

Abstract: In recent years, we have witnessed a surge of interest in skill-based queueing systems with product-form stationary distributions. One may, for example, think of a variety of redundancy systems or matching systems, where the pairing of customers with servers is subject to compatibility constraints. Under appropriate assumptions, the stationary distributions in such models have indeed been shown to have product forms.

In this presentation, we study how the recently developed pass-and-swap queue and the Whittle network, as well as combinations thereof, can be used to create a framework that captures a multitude of skill-based queueing systems with product forms.

This presentation is based on joint work with Céline Comte (LAAS-CNRS) and Kristy Gardner (Amherst College).

Karl Sigman

Title:  A Queueing Theory Approach to the Massive Access Problem (MAP) in Telecommunications

Abstract: We give an overview of a recently proposed technique for addressing the Massive Access Problem (MAP), an issue in telecommunications that arises when too many devices transmit packets to a gateway in quick succession. The technique, the Adaptive-Quasi-Deterministic Transmission Policy (AQDTP), involves a special case of “traffic shaping” that delays some packets at the points of origin to alleviate congestion at the routers, along with formal aspects of queueing theory. One nice feature of AQDTP is that it loses no packets and allows an infinite buffer. To clarify the approach within a general queueing-theory framework and to move beyond the original telecommunications application, we frame these potential delays as the time customers spend at a café (a pleasant endeavor!) before proceeding to potential congestion at a service facility (an unpleasant endeavor!) Both sample-path results and a general stationary stochastic setting are presented. In the iid case, a Harris-recurrent Markov process is shown to exist, along with its interesting regenerative properties.

Martijn Gösgens

Title: Tail index estimation with controlled exceedance rate

Abstract: To design systems that are protected against events much rarer than the observational record, extreme value theory is needed to extrapolate distribution tails. For example, dams are designed to withstand a certain water level that is likely to occur only once in several millennia, while observations are only available for the last few centuries. To bridge this gap, tail index estimators (e.g., the Hill estimator) are used. In this setting, conservative estimators are desirable: whenever the estimated tail index exceeds the true tail index, then the probability of exceedance events is underestimated by an order of magnitude. To this end, we derive the large and moderate deviations of the Hill estimator and use this to design estimators for which the probability of exceeding the true index is vanishingly small. In the large-deviations regime, we show that a simple rescaled version of the Hill index achieves an optimal balance between bias and exceedance rate.

Michel Mandjes

Title: A storage process with alternating Lévy input

Abstract: In this talk I discuss a queueing process for which the dynamics are changed once the workload in the queue exceeds a predefined threshold, and these new dynamics stay in force until the queue is emptied, at which point the previous dynamics are again reinstalled. For general spectrally-positive Lévy processes in each case, I derive expressions for the workload at an independent exponentially distributed random time horizon, and study in detail properties of the workload in stationarity. Then I work out explicit formulas as well as expressions for the optimal changing threshold for special cases of the underlying process dynamics and a chosen set of involved switching, holding and service cost functions.

Joint work with H. Albrecher, O. Boxma, and O. Kella.

Nikki Levering

Title: Decentralized optimization in cancel-on-complete redundancy systems

Abstract: Redundancy scheduling, whereby copies of the same job are sent to a subset of the servers, has the potential of reducing system latency by exploiting the variability in both the queue lengths and the server speeds. However, if not operated properly, redundancy scheduling may also increase latency by misusing the server resources. We focus on the cancel-on complete redundancy variant, where multiple copies of the same job may be processed at the same time, and all copies are cancelled whenever one of them is completed. Our aim is to characterize the optimal scheduling policy in each server so as to minimize the average number of jobs. Since computing and implementing an optimal centralized policy is infeasible in large-scale systems, except in very specific configurations such as nested systems, we develop a fully decentralized approach whereby each server independently determines its own scheduling policy. In our framework, servers are intelligent in that they leverage redundancy to infer system dynamics through the departure of local copies triggered by the departure of non-local ones. We present conditions under which, as the number of arrival streams per class and the number of jobs in service grows, the performance of the decentralized approach is locally or even globally asymptotically optimal.

This presentation is based on joint work with Urtzi Ayesta (CNRS-IRIT, Ikerbasque), Céline Comte (LAAS-CNRS) and Maaike Verloop (CNRS-IRIT).

Noa Zychlinski

Title: An Operational View on Managing Mass Trauma Events

Abstract: Mass Trauma Events (MTEs), such as wars, natural disasters, and terror attacks,
present significant operational challenges for mental health systems. Affected
populations require both immediate and prolonged support, complicating response
efforts and placing substantial strain on already limited resources. Post-traumatic
Stress Disorder (PTSD) can have lasting effects and imposes a considerable
economic burden, making timely and effective intervention critical. These
challenges highlight the need for practical, analytically grounded policies to
manage surges in demand and support recovery over time.
We study the coordination of group and individual therapy channels in a multiserver
queueing setting. Group therapy can alleviate immediate workload but may
lead to increased follow-up demand for individual treatment. Our model captures
this trade-off and the interdependence between therapy channels, while
accounting for key mental health system features such as patient no-shows and
dropouts. Using a fluid approximation, we derive index-based policies tailored to
the Surge, Recovery, and Long-Term Phases of MTEs, integrating time-varying,
transient, and steady-state dynamics. Drawing on data from previous MTEs, we
demonstrate that coordinated policies can significantly reduce congestion,
accelerate recovery, and improve overall system performance relative to
benchmark approaches.

Rami Atar

Title: From load balancing to the Atlas model and free boundary problems

Abstract: I will discuss recent connections between mean-field limits of load balancing systems, the Atlas model for Brownian particles on the line, and free boundary problems for the heat equation.

Joint work with Amarjit Budhiraja and Gershon Wolansky

Refael Hassin

Title: Sequential subgame-perfect equilibrium in a single-server system with waiting and lateness costs

Abstract: N players sequentially decide when to enter a queue. They wish to obtain service as early as possible, but they also incur costs related to waiting. In this paper we compute the Sequential Arrival Equilibrium path of this game.

Royi Jacobovic

Title: Nonparametric inference from possibly unstable M/G/1 workload observations

Abstract: Hansen and Pitts (2006) introduced the problem of nonparametric estimation of the service-time distribution of an M/G/1 queue observed through its workload process at discrete times t=0,1,\ldots,n. Despite its seemingly simple formulation, obtaining an estimator with sharp risk guarantees for this observation model has remained an open challenge for nearly two decades.

In this paper, we construct such an estimator that attains nearly parametric convergence rate without assuming stationarity, stability, or knowledge of the arrival rate.

Our approach is based on a two-stage screening procedure that uncovers a hidden conditionally independent compound Poisson structure within the dependent workload observations. This probabilistic reduction transforms the original estimation problem into a classical decompounding problem, making it possible to leverage existing nonparametric estimation techniques despite the complex dependence induced by the reflected workload process.

More broadly, we hope that the proposed screening methodology will provide a useful framework for statistical inference from dependent stochastic systems beyond the classical assumption of stability.

This presentation is based on joint work with my student Binyamin Kobzantsev and is dedicated to my former PhD supervisor, Offer Kella.

 Søren Asmussen

Title: Topics in Markov additive processes

Abstracts:Markov additive processes are popular in applications and are often termed Markov-modulated in queueing and insurance and regime-switching in finance. Two set of problems are considered:

1) Denseness properties of MAPs and some subclasses in the space $D[0,\infty)$, involving both the $J_1$- and the $M_1$-topology.

2) Different approaches to Wiener-Hopf factorization  for MAPs with phase-type jumps, in particular

  1. a) the Kella-Whitt martingale and its MAP extension,
  2. b) a matrix-analytic approach developed by the author and illustrated by examples from queueing and finance

Uri Yechiali: 

Title: A-Synchronic Inclusion Process (ASIP) in Random Environment*

Abstract: The single-environment A-Synchronic Inclusion Process (ASIP) is an n-site stochastic tandem network, where single particles (customers, cars, products, ships, molecules, etc.) arrive randomly at the first site and move stochastically, unidirectional, in accumulating clusters, from one site to the next, until exiting the system. Each site is comprised of an unlimited-size buffer and a gate in front of it. When gate j (j=1,2, …, n-1) opens, all particles present in its buffer move simultaneously, as a batch, to the next site and join the batch of particles already there. When gate n opens, all particles in its buffer leave the system. The A-synchronic Inclusion process (Reuveni, Eliazar and Yechiali [2011]) bridges between the well-known Tandem Jackson Network (TJN) and the A-synchronic Exclusion Process (ASEP).
We extend the scope of the family of ASIP systems and study an ASIP network operating in a 2-phase random environment, where the entire system alternates stochastically between two phases, residing in phase k (k = 1, 2) an exponentially distributed time with mean 1/η(k). When in phase k, the arrival process is Poisson with rate λ(k), and the gate in front of site j (j = 1, 2, …, n) opens, independently of all other gates, every exponentially distributed time with mean 1/μ(j, k).
We analyze this complex ‘ASIP in Random Environment’ system, and obtain the following: (i) the multi-dimensional probability generating function (PGF) of site occupancies; (ii) the mean occupancy of each site; (iii) the mean traversal time of a particle through the network; (iv) the PGF of the overall load of the first m sites (m = 1,2, . . ., n); (v) the probability that the first occupied site is site j (j = 1, 2, . . ., n); (vi) the mean length of a busy period; and (vii) the mean draining time (the time elapsing from the moment the arrival process stops until the first instant thereafter that the system becomes empty).

* Presentation in honor of my former student, Offer Kella, upon his retirement
** Joint work with Yaron Yeger

Willem van Jaarsveld

Title: Projected Inventory-Level Policies for Lost Sales Inventory Systems: Asymptotic Optimality in Two Regimes

Abstract: We consider the canonical periodic review lost sales inventory system with positive lead times and stochastic i.i.d. demand under the average cost criterion. We introduce a new policy that places orders such that the expected inventory level at the time of arrival of an order is at a fixed level and call it the projected inventory-level policy. We prove that this policy has a cost rate superior to the equivalent system where excess demand is back-ordered instead of lost and therefore, is asymptotically optimal as the cost of losing a sale approaches infinity under mild distributional assumptions. We further show that this policy dominates the constant-order policy for any finite lead time and therefore, is asymptotically optimal as the lead time approaches infinity for the case of exponentially distributed demand per period. Numerical results show that this policy also performs superior relative to other policies.

 

 

 

 

Details

  • Start: Oct 5, 09:30
  • End: Oct 7, 16:30

Venue

Comments are closed.

Powered by Nirvana & WordPress.