Networks

Special edition of the Dutch National Queueing Colloquium

On 19 February 2027, a special edition of the Dutch National Queueing Colloquium will take place at EURANDOM in Eindhoven, on the occasion of the 60th birthday of Sem Borst.

 

The colloquium has played a central role in fostering interaction and collaboration within the Dutch queueing community since 1978. By bringing together researchers from universities and research institutes across the Netherlands and beyond, the event offers a unique opportunity to exchange ideas and strengthen connections between various areas of queueing theory and stochastic networks. The program on 19 February features talks by six eminent speakers on a wide range of topics, reflecting the rich scope of recent advances and emerging directions in this rapidly evolving field.

 

Program

10.00-10.05 Opening
10.05-10.45 Céline Comte (CNRS and Toulouse)
10.45-11.15 Break
11.15-11.55 Rob van der Mei (CWI and Vrije Universiteit)
11.55-12.35 Johan van Leeuwaarden (Tilburg University)
12.35-14.00 Lunch break
14.00-14.40 Debankur Mukherjee (Georgia Tech)
14.40-15.00 Break
15.00-15.40 Jaron Sanders (Eindhoven University of Technology)
15.40-16.00 Break
16.00-16.40 Bert Zwart (CWI and Eindhoven University of Technology)
16.40-18.00 Reception

 

Registration

Participation is free of charge, but for catering purposes we kindly request that you register in advance. Registration includes coffee, tea, and lunch.

 

 

> Register now

 

Céline Comte

Fluid Value Functions for Policy-Gradient Algorithms

Baselines are a standard variance-reduction tool in policy-gradient algorithms, but the variance-minimizing baseline requires computing the value function—precisely what simulation-based methods aim to avoid. In this paper, we propose a fluid baseline: a computationally tractable alternative built from the value function of a deterministic proxy for the controlled Markov chain, defined by an ordinary differential equation (ODE) driven by the chain's drift. The fluid value at any state is obtained by integrating a single deterministic trajectory of this ODE and accumulating the cost along the way.

To theoretically establish the fluid baseline's effectiveness, we analyze a family of controlled chains indexed by the contraction rate $\alpha$ of the fluid model toward its fixed point, in the regime where $\alpha$ approaches 0. Our main result proves that the fluid baseline reduces the variance of the gradient estimator by a factor of order $\alpha$ relative to no-baseline, matching the order-of-magnitude reduction achieved by the optimal baseline.

Practically, the fluid baseline can be integrated into any policy-gradient scheme that approximates the action-value function. We analyze the approach of using the fluid approximation for the action-value term as well, showing that the resulting bias is small relative to the gradient away from optimality, and remains below the noise of a single gradient sample near the optimum. This yields an estimator that preserves the theoretical guarantees of exact methods while avoiding their computational burden.

Our numerical experiments on two queueing problems validate these theoretical contributions in terms of variance reduction. Moreover, we observe that a stochastic gradient descent algorithm using the fluid value function achieves similar performance as one using the value function, when the latter is computationally tractable.

This is joint work with Itai Gurvich (Northwestern University).

 

Rob van der Mei

A Deep Dive into the Wonderful World of Polling Models

 

Johan van Leeuwaarden

Prophets and Random Walkers

Consider a random walk observed one step at a time. A prophet, who sees the entire future path, can stop at its maximum. A gambler has to decide online and may know the underlying model only through a limited number of past sample paths. How much of the prophet’s reward can still be guaranteed? This talk discusses joint work with Pieter Kleer and Daan Noordenbos on prophet inequalities for random walks and related sample-based stopping problems. Classical fluctuation theory allows the effect of sample information to be isolated and quantified sharply. As the amount of sample information grows, the optimal guarantee approaches 1/e, about 36.8 percent: the loss caused by not knowing the model disappears, while the fundamental disadvantage of not knowing the future remains. Given the special occasion, the path will occasionally wander sideways, including reflections on earlier work and collaborations with Sem Borst.

 

Debankur Mukherjee

Higher-Order Approximations of Sojourn Times in M/G/1 Queues via Stein's Method

We discuss the stationary sojourn time distribution in an M/G/1 queue operating under heavy traffic. It is known that the sojourn time converges to an exponential distribution in the limit. Our focus is on obtaining pre-asymptotic, higher-order approximations that go beyond the classical exponential limit. Using Stein’s method, we develop an approach based on higher-order expansions of the generator of the underlying Markov process. The key technical step is to represent higher-order derivatives in terms of lower-order ones and control the resulting error via derivative bounds of the Stein equation. Under suitable moment-matching conditions on the service distribution, we show that the approximation error decays as a high-order power of the slack parameter 𝜀 = 1 − 𝜌. Error bounds are established in the Zolotarev metric, which further imply bounds on the Wasserstein distance as well as the moments. Our results demonstrate that the accuracy of the exponential approximation can be systematically improved by matching progressively more moments of the service distribution.

 

Jaron Sanders

Recovering semipermeable barriers from reflected Brownian motion

In this talk, I will introduce you to a new problem: that of recovering semipermeable barriers from the trajectory of a stochastic process in a planar domain. The considered process acts like Brownian motion when away from the barriers. It is reflected upon contact with a barrier until a sufficient but random amount of interaction has occurred, after which it passes through. The average amount of interaction required is determined by the permeability of a barrier. Given a sequence of samples, we wonder when one can determine the location and shape of the barriers.

For this problem, we identify different recovery regimes. For example, the observation period T dictates if the full barriers or only certain pieces can be recovered, and the sampling rate influences the convergence rate significantly. Further, the environment’s impact on the difficulty of the problem is quantified using interpretable parameters in the recovery guarantees. Our results are accompanied by explicit algorithms, which we have also applied to some real-life data.

This talk is based on joint work with Alexander Van Werde. A preprint can be found here: https://arxiv.org/pdf/2412.14740

 

Details
When: Friday February 19th, 2027
Where: EURANDOM, Eindhoven
Location

5metaforum01

 

EURANDOM

Den Dolech 2
5612 AZ  Eindhoven
The Netherlands