Conference 2026

Abstracts

Plenary presentations

Merve Bodur

Methodological Advances in Two-stage Stochastic Programming

Stochastic programming is a useful framework for dealing with uncertainty and integrality requirements in optimization problems. In this talk, we focus on two-stage stochastic integer programs (2SIPs), in particular their sample average approximations which yield computationally challenging mixed-integer programming (MIP) models. For various classes of 2SIPs, we review commonly used decomposition approaches (most notably logic-based Benders decomposition variants), and present some important enhancement strategies such as incorporation of general purpose cutting planes from the MIP literature. We also introduce several recently proposed techniques, such as the uses of

decision diagrams to convexify the second-stage problems,
machine learning to learn the recourse value functions as well as cut selection/generation, and
MIP-based methodology for scenario reduction.

For the presented methods, we provide numerical results on problems from a variety of application domains.

Recent Advances in Solving Multistage Stochastic Mixed-integer Programs

Multistage stochastic mixed-integer programs (MSMIPs) can model complex sequential decision-making problems under uncertainty and appear in many applications. However, due to both the stochastic and integer components, their inherent computational challenges require sophisticated solution methodologies. In this talk, we will review recent advances in solving MSMIPs, in particular for the scenario-tree-based approaches. We will discuss exact methods, bounding ideas, partially extended formulations, and various policy classes, including aggregation-based policies, decision-rule-based policies, and interpretable policies.

Antonis Economou

Queueing Games: The role of information

After a brief introduction in the area of Queueing Games (also known as Rational Queueing), we will focus on the impact of information that is provided to the customers on their strategic behavior and associated social welfare and system revenue. We will compare various information structures and discuss their pros and cons.

Queueing Games: The role of customers’ heterogeneity

We will present some classical and recent queueing models with strategic customers who are heterogeneous in some respects. We will discuss the implications of heterogeneity on the economic performance of the systems under study and on the optimization problems faced by a social planner or a monopolist.

Matthew O. Jackson

Lecture 1: Multiplexing in Networks and Diffusion

Social and economic networks are often multiplexed, meaning that people are connected by different types of relationships—such as borrowing goods and giving advice. We make three contributions to the study of multiplexing. First, we document empirical multiplexing patterns in Indian village data. Second, we examine how these layers and their overlap affect information diffusion in a field experiment. This leads to our third contribution: developing a model and theoretical results about diffusion in multiplex networks. Multiplexing slows the spread of simple contagions, such as diseases or basic information, but can either impede or enhance the spread of complex contagions, such as new technologies, depending on their virality. Finally, we identify differences in multiplexing by gender and connectedness.

Lecture 2: Using AI to Categorize Games and Better Understand Human Behaviors in Strategic Situations

By varying prompts to a large language model, we can elicit the full range of human behaviors in a variety of different scenarios in classic economic games. By analyzing which prompts elicit which behaviors, we can categorize and compare different strategic situations, which can also help provide insight into what different economic scenarios induce people to think about. We discuss how this provides a step towards a non-standard method of inferring (deciphering) the motivations behind human behaviors. We also show how this deciphering process can be used to categorize differences in the behavioral tendencies of different populations.

Madeleine Udell

Lecture 1: Hunting the Hessian: Randomized Methods

Low rank structure is pervasive in real-world datasets. This talk shows how to accelerate the solution of fundamental computational problems, including eigenvalue decomposition, linear system solves, composite convex optimization, and stochastic optimization (including deep learning), by exploiting this low rank structure. We present a simple method based on randomized numerical linear algebra for efficiently computing approximate top eigendecompositions, which can be used to replace large matrices (such as Hessians and constraint matrices) with low rank surrogates that are faster to apply and invert. The resulting solvers for linear systems (NystromPCG), composite convex optimization (NysADMM), and stochastic optimization (SketchySGD and PROMISE) demonstrate strong theoretical and numerical support, outperforming state-of-the-art methods in terms of speed and robustness to hyperparameters.

Lecture 2: Hunting the Hessian: Online Methods

We introduce a framework to accelerate the convergence of gradient-based methods with online learning. The framework learns to scale the gradient at each iteration through an online learning algorithm and provably accelerates gradient-based methods asymptotically. In contrast with previous literature, where convergence is established based on worst-case analysis, our framework provides a strong convergence guarantee with respect to the optimal stepsize for the iteration trajectory. We obtain the first convergence guarantee for the widely-used hypergradient descent method (HDM) and for new variants with heavy-ball momentum (HDM-HB) and Nesterov momentum. HDM-HB significantly outperforms other adaptive first-order methods, and often matches the performance of L-BFGS using less memory and cheaper iterations.

LNMB / NGB Seminar

Durk Kingma

Generative Models: Past, Present, Future

In the past decade, the field of generative models has seen rapid progress in capabilities, with large language models (LLMs) and diffusion models now used routinely and productively in daily life. This progress in capabilities was caused in equal parts by significant increases in scale, and improved algorithms. In this talk, we will review some of the main machine learning techniques for learning generative models, including some of our own work while working at university (UvA) and industry (OpenAI, Google, Anthropic). We will discuss current LLM-powered productivity tools, and the short- and medium-term capabilities we can expect.

Wouter Kool

When OR Meets Generative AI — From Exact Models to Real-World Complexity

The success of operations research has always depended on our ability to model reality — but real-world problems are rarely as clean as our equations. Generative AI, and in particular large language models, open new possibilities for turning ambiguous requirements, messy data, and human intent into structured problems solvable by OR. In this talk, we explore this emerging connection, outline how recent AI developments complement traditional optimization, and consider both opportunities and limitations. The session will set the stage for a day of ideas at the intersection of mathematics, AI, and decision-making.

Bernardino Romera Paredes

Evolving Superhuman Algorithms for Long-Horizon Planning

Automated planning in complex domains like autonomous manufacturing and container port logistics faces a dual challenge: classical solvers cannot scale, and manual algorithm design is prohibitively expensive. While LLMs offer a potential solution, they historically fail at the systematic reasoning needed for long-horizon tasks.

In this talk, we introduce the Hive, a novel framework that automates algorithm design by combining the generative creativity of LLMs with the rigorous selection pressure of evolutionary computation. We validate this approach using the Airbus Beluga competition, where the Hive autonomously evolved a solver for a massive aircraft production scheduling problem. Given only the problem description and a feedback signal, our system generated strategies that outperformed rival expert-engineered submissions.

We conclude by discussing how this approach democratizes algorithm design for novel domains where combinatorial explosion renders classical methods ineffective and domain expertise is scarce.

Madeleine Udell

AI and the Future of Optimization

Optimization problems are pervasive in sectors from manufacturing and distribution to healthcare. However, most such problems are still solved heuristically by hand rather than optimally by state-of-the-art solvers, as the expertise required to formulate and solve these problems limits the widespread adoption of optimization tools and techniques.

As a glimpse of the future, this talk will introduce OptiMUS, a Large Language Model (LLM)-based agent designed to formulate and solve MILP problems from natural language descriptions. OptiMUS can develop mathematical models, write and debug solver code, develop tests, and check the validity of generated solutions. More broadly, we discuss the potential for LLMs in domains where accuracy and fidelity to real-world data is critical and strategies to augment and safeguard their performance.

PhD presentations

Monday, 12 January 2026

Room: Steyl

Anne-Fleur Dijkhorst, Erasmus University

Title: Resource allocation in hospitals and ambulances to decrease time to care in EMS systems in LMICs

Abstract: Emergency medical services (EMS) in low- and middle-income countries (LMICs) are often fragmented, with multiple healthcare providers each operating a small fleet of ambulances and sometimes their own emergency hotlines. This decentralized system reduces efficiency and increases response times (""time to scene"") compared to the more coordinated EMS systems in high-income countries. Furthermore, when emergency cases require specialized care, patients in LMICs are less likely to find the required care at the nearest healthcare facility or in the nearest ambulance. Healthcare facilities and ambulances in these regions often vary greatly in terms of available equipment and level of care, which can result in extended delays before patients receive appropriate treatment ("time to care").

In several LMICs, organizations have addressed the lack of coordination of the decentralized EMS by introducing an ambulance platform to decrease the time to scene of medical emergencies. An example of such a platform is our industry partner Flare, which operates in Kenya. However, to also address the time to care, it is necessary to look beyond the availability of ambulances and dispatch policies and investigate how long it takes for patients to receive the specialized care that their emergency cases require. The time to care can decrease by investing in new equipment — for example, AEDs, diagnostic equipment, or surgical kits ¬— in the existing hospitals and ambulances.

In this study, we develop two mathematical models to investigate the decision-making process of investing in equipment in hospitals and ambulances to decrease the time to care; a continuous approximation model and a data-driven optimization model. Both models are used to first evaluate the time to care for emergency cases to identify types of emergency cases that often have a long time to care, or locations from which the time to care after an incident is long. Then, the models are used to find the optimal allocation of investment in new equipment to reduce time to care. Given a limited budget, we investigate the trade-off between adding the new equipment to healthcare facilities or ambulances.

The continuous approximation model is used to derive simple decision rules for adding equipment which can be generalized to different areas. This is done based on the density of equipped and unequipped ambulances and hospitals in the area, such that no detailed data on exact locations of ambulances and hospitals is necessary for this model. An example of such a decision rule would be: if the density of equipped ambulances is larger than the density of equipped hospitals in a region, then it is optimal to invest in equipping more ambulances. The data-driven model is applied to a specific case study from Flare in Kenya. This model requires more detailed input data, such as the exact locations and equipment available in the ambulances and hospitals. The optimal solution from the model shows in which specific ambulances or hospitals we should invest. We verify whether these are in line with the general decision rules obtained from the continuous approximation model .

Giulia Montagna, Erasmus University

Title: Information design for volunteer emergency response

Abstract: Community first responder systems expand the traditional emergency services with trained volunteers that can be dispatched by an app. Volunteers can accept or decline alerts to provide first aid at an emergency until professional services arrive. The decision to accept may depend on the expected impact on patient health outcomes. Usually, volunteers are only informed about their distance to the incident. However, additional information, such as the number of other volunteers closer, might influence the acceptance probability.

This study investigates how providing more information to volunteers could lead to the most valuable volunteers accepting the request. The value of a volunteer indicates how much the patient would benefit if they accept. It will be measured as the expected reduction in time to CPR. This value depends on the location of all available volunteers, their acceptance probabilities and the response time of the emergency services.

Given the relationship between the believed value of a volunteer and their acceptance probability, we explore how sharing more information can nudge the highest value volunteer to accept the alert. How would sharing more information with volunteers impact the response time? How much information would you need to share? How does this depend on the relationship between the acceptance probability and the expected value?

Roby Cremers, Erasmus University

Title: Wanted, dead or alive: Leveraging deceased donors for enhanced patient outcomes in kidney exchange

Abstract: We introduce a novel constraint learning mechanism for Constraint Programming (CP) solvers that integrates cutting planes reasoning into the conflict analysis procedure. Our approach, named Lazy Linear Generation (LLG), can derive implied linear integer inequalities to prune the search space, rather than propositional clauses, as is the standard in current CP solvers. This combines the strengths of constraint programming (strong propagation through global constraints) with the cutting-planes proof system, which is possibly stronger than clausal proof systems.

Room: St. Janzaal

Denise Graafsma, University of Twente

Title: Refining the Complexity Landscape of Speed Scaling

Abstract: We study the computational complexity of scheduling jobs on a single speed-scalable processor with the objective of capturing the trade-off between the (weighted) flow time and the energy consumption. This trade-off has been extensively explored in the literature through a number of problem formulations that differ in the specific job characteristics and the precise objective function. Nevertheless, the computational complexity of four important problem variants has remained unresolved and was explicitly identified as an open question in prior work.

We settle the complexity of these variants. More specifically, we prove that the problem of minimizing the objective of total (weighted) flow time plus energy is NP-hard for the cases of (i) unit-weight jobs with arbitrary sizes, and (ii) arbitrary-weight jobs with unit sizes. These results extend to the objective of minimizing the total (weighted) flow time subject to an energy budget and hold even when the schedule is required to adhere to a given priority ordering. In contrast, we show that when a completion-time ordering is provided, the same problem variants become polynomial-time solvable. The latter result highlights the subtle differences between priority and completion orderings for the problem.

Cédric Roy, Eindhoven University of Technology

Title: A framework for handling and exploiting symmetry in Benders' decomposition

Abstract: Benders' decomposition (BD) is a framework for solving optimization problems by removing some variables and modeling their contribution to the original problem via so-called Benders cuts. While many advanced optimization techniques can be applied in a BD framework, one central technique has not been applied systematically in BD: symmetry handling. The main reason for this is that Benders cuts are not known explicitly but only generated via a separation oracle.

In this work, we close this gap by developing a theory of symmetry detection within the BD framework. To this end, we introduce a tailored family of graphs that capture the symmetry information of both the Benders master problem and the Benders oracles. Once symmetries of these graphs are known, which can be found by established techniques, classical symmetry handling approaches become available to accelerate BD. We complement these approaches by devising novel techniques for the separation and aggregation of symmetric Benders cuts by means of tailored separation routines and extended formulations. Both substantially reduce the number of executions of the separation oracles. In a numerical study, we show the effect of both symmetry handling and cut aggregation for bin packing and scheduling problems.

Maarten Flippo, Delft University of Technology

Title: Exploiting cutting planes reasoning in constraint programming solvers

Tuesday, 13 January 2026

Room: Steyl

Victor R. Ploeg, University of Twente

Title:When pandemic allocation meets regular care: stability analysis of a multi-hospital regular care queueing network via fluid limits

Abstract:Pandemic patient-allocation policies usually aim to balance infectious disease load across hospitals, yet their impact on regular non-pandemic care, often overlooked during crises, remains insufficiently understood. This work develops a fluid limit framework that captures the long-term interaction between pandemic patient allocation and regular care waiting lists in a regional network of hospitals. By deriving a system of ordinary differential equations that describes the fluid evolution of hospital backlogs under a general allocation rule, we assess whether such policies preserve the natural order of waiting lists, avoid oscillatory behaviour, and maintain system-wide stability. The approach enables rigorous evaluation, ensuring that no hospital accumulates disproportionate backlog even under fluctuating pandemic pressure. We illustrate the framework using a recently proposed allocation rule designed to minimise the impact of pandemic admissions on regular-care queues. The results provide a mathematical foundation for assessing allocation mechanisms not only in pandemic preparedness, but also in broader scenarios involving healthcare strain or emergency-induced shifts in patient inflow.

Sten Wessel, Eindhoven University of Technology

Title: Fair policies for sequential resource allocation

Abstract: In sequential resource allocation, we are given an amount of supply, an ordering of n clients, and for each client a probability distribution of their demand. The supply needs to be allocated to the clients in an online fashion, i.e., only at arrival at the client the demand of that client becomes known. As typically the total demand of the clients exceeds the available supply, the goal is to maximize the minimum fill rate, where the fill rate of a client is defined as the fraction of allocated amount and realized demand. We apply the notion of individual fairness to sequential resource allocation. We introduce a new policy based on linear programming that is individually fair. Further, we provide an extensive computational study to evaluate existing policies and our new policy with respect to ex-post and ex-ante fairness. Using a fine-tuned implementation of existing methods, we are able to evaluate significantly larger problem instances than those previously considered in the literature. As a consequence, we are able to assess how the performance of various polices scales with the number of clients. We are also able to show to what extent the position of a client in the sequence affects their allocation. We clarify to what extent applying individual fairness impacts the minimum fill rate. Also, our computational evaluation of several allocation policies gives insight into the quality of their allocations from various angles: (i) from a fairness perspective, and (ii) when number of clients increases. These findings are relevant for any organization that aims to distribute supply over clients in an online fashion while pursuing a fairness objective.

Ugur Guler, Maastricht University

Title: A fair and efficient way of cost sharing in collaborative VRP

Abstract: In modern logistics, competition and market fragmentation often lead to inefficiencies, as companies design their delivery routes independently. Collaborative Vehicle Routing Problem has emerged as a promising approach to address this issue by allowing companies to form coalitions, share depots and customers, and jointly optimize their routes to reduce total transportation costs. While collaboration yields significant efficiency gains, it also raises the challenge of allocating these savings fairly among participants. In this work, we propose an algorithm that simultaneously minimizes the total routing cost and determines a fair cost allocation. We prove that the proposed method satisfies three key fairness properties from cooperative game theory. Furthermore, we benchmark the algorithm against established solution concepts to highlight its practicality and efficiency. Finally, we present empirical results on Vehicle Routing Problem instances to demonstrate the performance and applicability of our approach.

Room: Angola

Danny Zhu, Erasmus University

Title: Real-time rolling stock rescheduling with dynamic and uncertain disruption information

Abstract: On the day of operations, disruptions can occur on a railway network, which require the rolling stock plan to be adjusted. In this paper, we consider the problem of rolling stock rescheduling with disruptions for which information about the duration, precise location and severity is uncertain and becomes available dynamically. In current practice, it is typical that the most recent disruption information is naively used to reschedule the rolling stock, which can lead to unnecessary cancelations or decreased passenger satisfaction in case the disruption turns out different than expected. We propose two approaches for real-time rolling stock rescheduling with dynamic and uncertain disruption information, which have the ability to anticipate changes in the disruption information. The approaches iteratively create interim rolling stock schedules by solving a two-stage optimization problem with a variety of optimization scenarios. Computational experiments are conducted on instances of the Dutch railway network. Compared to the naive rolling stock rescheduling approach, the developed approaches successfully create rolling stock schedules with fewer cancelations and lower rescheduling costs for disruptions with changing information, in short computation times.

Philip de Bruin, Utrecht University

Title: Integrated airline fleet and crew recovery through local search

Abstract: Airline operations are prone to delays and disruptions, since the schedules are generally tight and depend on a lot of resources. When disruptions occur, the flight schedule needs to be adjusted for the operation to continue. This needs to be done as close to real-time as possible, posing a challenge with respect to computation time. Moreover, to limit the impact of disruptions, we need a solution with minimal cost and passenger impact. Airline operations include many interlinked decisions, making the use of sequential methods suboptimal. For this reason, we study an integrated approach, specifically looking at the aircraft and crew schedules. Resolving these disruptions in an integrated way is a complex problem.

To solve this, we developed a fast simulated annealing approach. Our approach is compared with traditional approaches, and an experimental study is done to evaluate different neighbour generation methods, and to investigate different recovery scenarios and strategies. We use real world data from KLM Royal Dutch Airlines. We show that our approach resolves disruptions quickly and in a cost-efficient manner, and that it outperforms traditional approaches. Compared to naive delay propagation, our method saves 40% in non-performance costs. Moreover, while most airlines use tools that consider resources separately, our approach shows that integrated disruption management is possible within 30 seconds.

Gijs Bakker, University of Groningen

Title: A dynamic overbooking model for parcel locker systems

Abstract: In this study, we examine the reservation management problem in parcel locker systems, which offer a sustainable and efficient last-mile delivery solution. A key challenge in managing locker capacity arises from the stochastic nature of parcel pickup times. Most existing studies address this problem under the simplifying assumption of deterministic pickup behaviour, leading to conservative acceptance policies and suboptimal utilization. Through collaboration with European logistics providers, we gained practical insights into operational decision-making and observed that capacity overbooking is sometimes applied in practice. Despite its relevance in practice and in other domains, the concept of capacity overbooking has been underexplored in the context of parcel locker systems with these specific operational characteristics. To mitigate this, we model the process as an infinite-horizon semi-Markov decision process and proof that the optimal policy is monotone. We propose a newsvendor approach to derive near-optimal solutions. Our results show that average utilization can increase by up to 30% compared to benchmark methods without compromising service reliability. We further demonstrate that overbooking is beneficial only under specific operational conditions.

Room: Congo

Jiayun Wang, University of Twente

Title: Dynamic resource flow optimization in hubs for circularity (H4C)

Abstract: Industrial Symbiosis Networks (ISNs) are becoming increasingly important as a strategy for sustainable resource management. They are clusters of industries interacting in a way that one’s waste, e.g., materials and energy, can be used as input of another, potentially after treatment. The hub for circularity (H4C) extends the traditional boundaries of ISNs by facilitating resource exchanges not only within industrial clusters but also with surrounding urban and rural areas. Within this context, Waste Management Center (WMC) plays a central role in coordinating production, storage, and logistics for treatment plants. In this study, we aim to develop decision-making tools for WMC to support its daily operations, taking into account uncertainties in waste availability and resource demand. We formulate a Markov decision process (MDP) where a set of interconnected treatment plants dynamically decide on production planning, inventory levels, and resource exchanges. As this MDP is intractable to solve to optimality due to state and action spaces that grow exponentially with the network size, we leverage state-of-the-art approximate dynamic programming (ADP) methods to compute near-optimal policies and dual bounds. Our results highlight the value of collaboration and provide insights into production and inventory strategies in relation to the evolution of uncertainties.

Maike de Jongh, University of Twente

Title: Preventing large-scale avalanches in the two-dimensional Abelian sandpile model through strategic interventions

Abstract: The Abelian sandpile model is a classic example of a system in which critical behavior emerges independently of external parameters, a phenomenon known as self-organized criticality. In this model, sand grains are added randomly to vertices of a graph. Whenever a vertex collects too many sand grains, it topples and redistributes its mass to its neighbors, a mechanism that may trigger large cascading events. Such avalanches are frequently used to model catastrophic phenomena in real-world systems, such as earthquakes, financial market crashes, and forest fires. Motivated by these applications, we study intervention strategies aimed at mitigating avalanche sizes in the Abelian sandpile model. We present theoretical results on how removing sand grains at designated locations affects the expected avalanche size. In addition, we evaluate several heuristic strategies and investigate their long-term impact on the evolution of the system. Our findings demonstrate the potential of strategic interventions in stabilizing self-organized critical systems.

Britt van Veggel, University of Amsterdam

Title: Heatmaps for maximum coverage facility location optimization

Abstract: We propose an efficient method to generate so-called heatmaps for the maximum coverage facility location problem. Heatmaps are geospatial representations in which each region is colored according to the expected value or distribution of a given variable, assuming a facility were to be placed within this region. We discuss three applications of this methodology. The first application is that such a heatmap gives the user insight in other (almost) optimal locations for the facility for both (almost) highest coverage and mean traveling distance. Secondly, the heatmaps can be used as a post-processing improvement method for solutions found using a coordinate-based input. Lastly, the heatmaps can be used as a pre-processing method to improve the initial discrete set of potential facility locations, and with that the quality of the solution.

We develop two variants of the method: one can be used in case the distances are Euclidean, and one for the case that the distance is calculated via the roads. The methodology is tested on a case that aims to find optimal locations for water wells in West Darfur.

The numerical results show that the pre-processing stage allows the user to reach full coverage with 15\% less facilities (7707 rather than 9098).

Room: Steyl

Alban Kryeziu, University of Groningen

Title: A unified variation error bounds for expected value functions with application to two-stage recourse models

Abstract: In this paper we study the expected error incurred when replacing the nonconvex second-stage value in the two-stage stochastic mixed integer programs by a convex approximation. We develop a variation-potential framework that yields distribution-aware certificates depending on the joint bounded-variation (BV) regularity of the underlying distribution. The analysis uses BV Gauss-Green identity with two boundary elimination mechanisms, and it extends from smooth to BV and locally integrable densities via mollification. The results hold in any dimension and extend beyond boxes to general Lipschitz domains through bi-Lipschitz charts, and the bounds have explicit domain constants. In addition, we construct a convex piecewise-linear surrogate via the minimax (Chebyshev) fitting that enforces slice-mean-zero condition and provides a uniform error radius that enters the certificate directly. Furthermore, we provide a complementary Wasserstein-based bounds and identify regimes where BV and transport type guarantees are respectively tighter. Overall, the framework proposes verifiable and uniform error bounds that capture dependencies in the joint law and integrate seamlessly with standard discretization and decomposition schemes.

Sina Shahri Majarshin, Eindhoven University of Technology

Title: A robust age-based replacement problem

Abstract: We study the age-replacement problem for components under distributional uncertainty, where only the first and second moments of the lifetime distribution are known. Adopting a distributionally robust optimization framework, we minimize the long-run average cost per unit time against the worst-case distribution within the moment-constrained ambiguity set. We prove that the worst-case cost is attained by a two-point distribution, reducing the problem to a tractable finite-dimensional optimization. Our analysis refines classical conditions for failure-only replacement and provides an explicit preventive replacement policy when optimal. Numerical experiments reveal a non-monotonic relationship between lifetime standard deviation and replacement timing, offering practical insights for maintenance in uncertain environments such as manufacturing and healthcare.

Kjell Raaijmakers, Eindhoven University of Technology

Title: The speed–accuracy trade-off: optimizing congestion, testing speed, and diagnostic precision

Abstract: We study a diagnostics model in which an agent must determine the type of arriving customers through imperfect tests. Testing takes time, during which additional customers arrive, leading to congestion costs. To manage congestion, the agent can adjust the test rate at an increased cost. Misclassifying a customer also incurs a cost, creating a trade-off between testing accuracy, waiting costs, and diagnostic speed costs. We formulate this problem as a Markov Decision Process (MDP) and establish the existence of a search interval: When the agent’s belief about a customer’s type falls within this interval, testing continues; otherwise, classification occurs. We show that the search interval shrinks as congestion increases. Moreover, when congestion is high or when the agent’s belief nears the decision threshold, the optimal policy prescribes a higher service rate despite the increased cost.

Room: Angola

Busra Gultekin, Wageningen University & Research

Title: Reverse stress testing for supply chains

Abstract: This study introduces reverse stress testing for supply chains, designed to identify the minimal deviations from normal operations that would drive supply chains to a predefined performance failure. First, we present a framework for reverse stress testing with the purpose of assessing supply chain vulnerabilities. The framework involves six steps: selecting risk variables, defining baselines, formalizing the policy, formalizing key performance indicators (KPIs), determining critical thresholds, and identifying the critical scenario. Building on this framework, we propose an optimization model that identifies the critical scenario by uncovering the smallest risk variable deviations that stress the supply chain beyond critical KPI thresholds. We also show that the optimization model has desirable theoretical properties, proving that a stress scenario always exists under mild assumptions and that the model is computationally tractable. Two numerical examples demonstrate the framework’s applicability: a serial supply chain model and the service guarantee model by Simchi-Levi et al. (2018). Results show that the framework effectively uncovers where disruptions become most harmful.

Tjum van Dijck, Tilburg University

Title: Inventory planning in capacitated high-tech assembly systems under non-stationary demand

Abstract: We study inventory planning in high-tech low-volume manufacturing supply chains driven by the semiconductor market. This problem is characterized by the challenging combination of time-correlated trending demand and stringent capacity constraints. Here, anticipating demand fluctuations, while balancing inventories across nodes at each echelon of the assembly network, remains a fundamental challenge. We introduce a tractable non-stationary demand model that can be directly fitted to strategic market outlooks and that reflects key demand dynamics observed in practice.

To solve the resulting capacitated inventory planning problem, we first develop a new two-way balancing base-stock policy that balances inventory with upstream nodes and with restricted future production of downstream nodes, thereby substantially outperforming state-of-the-art base-stock policies. We further propose a scalable deep reinforcement learning (DRL) approach that decomposes the multi-dimensional action into single-node sub-decisions with a shared output layer in the neural network, and we enable cross-learning between sub-decisions through standardized node-specific features.

This DRL approach, initiated from the two-way balancing base-stock policy, attains near-optimal performance, with optimality gaps of 0.1% on average across several smaller instances. An extensive case, motivated by the setting of our industry partner ASML, highlights that two-way inventory balancing, whether achieved through analytical methods or learned policies, is essential for effective inventory control in capacitated assembly systems. Furthermore, we observe that DRL offers additional performance gains in non-stationary environments by effectively anticipating the risk of structural demand increases under capacity restrictions. This is achieved by strategically pre-building additional safety stocks at nodes that carry an elevated risk of becoming a bottleneck for system output, as well as at the most downstream node.

Casper Loman, Utrecht University

Title: A two-stage stochastic programming algorithm for creating robust freight train schedules

Abstract: Freight trains play an important role in the transportation of goods across large distances and international borders. However, freight train transport is prone to disturbances. In ca- pacity limited networks such as the Netherlands, it is therefore important that schedules for freight trains include room to handle these disturbances. We propose a two-stage stochastic programming algorithm that creates a planning for a set of freight trains, while considering a set of delay scenarios for these trains. The first stage of our problem consists of creating an initial planning for the freight trains, where binary variables are used to model the possible routes of all the trains. Then in the second stage scenarios, the actual release dates of the trains become available. In each scenario, some of the trains will be delayed and therefore need to be replanned; these new routes of these delayed trains are also modelled with binary variables. Our algorithm starts with a progressive hedging procedure to limit the number of binary route variables in the first stage. After this, a full Integer linear program is solved with the restricted set of initial phase variables, and all of the corresponding second stage variables for all scenar- ios. To control the desired robustness level of our solution, we include a Γ-robustness constraint that limits how often a train is allowed to not fit in a delay scenario. We test our algorithm with real-life data from ProRail. We evaluate the schedules created by different settings of our algorithm through discrete event simulation. Here we compare the simulated robustness of the schedule, as well as the deterministic quality of the schedule. We show which parameters of the algorithm have the most influence on the robustness of the schedule, and the trade-off that can be achieved for robustness against the quality of service. Our initial results suggest that our method is successful in creating schedules for freight trains that are robust against disruptions, and easily adjustable to attain a desirable level of robustness.

Room: Congo

Ekaterina Kosarevskaia, Leiden University

Title: Two-product make-to-stock systems: strategic joining and optimal inventory levels

Abstract: This talk analyzes a make-to-stock queueing system where a single production facility produces two products with independent Poisson customer arrivals. Customers make strategic join-or-balk decisions without observing inventory levels. The analysis examines Nash equilibria in joining strategies and characterizes optimal base-stock levels from both profit-maximizing and welfare-maximizing perspectives.

Joint work with Odysseas Kanavetas

Bas Verseveldt, Tilburg University

Title: Robust randomized pricing with purpose

Abstract: How can firms set prices that are socially responsible when demand is uncertain? We study monopoly pricing when the firm knows only the mean and support of the demand and wishes to balance profit with consumer surplus. Considering three formulations—mixed objectives, Nash bargaining, and Rawlsian fairness—we analyze both fixed and randomized pricing, and demonstrate how the seller can improve his strategy by choosing a pricing distribution over a posted price. We also compare randomized pricing mechanisms to direct truthful mechanisms and show that for social objectives the two are in general not equivalent, and can in fact have vastly different optimal strategies. Using max–min optimization and saddle-point arguments, we show that for a distributionally robust seller a stronger concern for consumer surplus generally leads to lower and more inclusive pricing, though the speed of adjustment varies greatly across objectives and mechanisms, offering firms strategic guidance for socially responsible pricing under uncertainty.

Camiel Koopmans, Leiden University

Title: To control or not to control: reducing average cost when control comes at a price

Abstract: In standard queueing control systems, it is assumed that control is a permanent feature of the system. In practice, this does not always happen. One can consider human controllers who need to do different tasks as well, or a specific real-life situation with sensor-based control for rush hour lanes with fallible sensors. Therefore, we argue that it is time to research queueing control systems with temporary control dynamics.

In this talk, we consider a Markovian model with control periods of service rate control with a fixed service rate after control and observation is lost. The next control period can be scheduled at any time point but comes at a price. We will identify in which cases the use of control can actually reduce the average cost and we give a lower bound of the average cost reduction for these cases.

This talk will not only be of interest through this new model and the found results, but also due to our combined methodology using theory about Discrete Time Markov Decision Processes and Value Iteration, Continuous Time Stochastic Processes with a vanishing discount, Renewal Reward Processes and notions like Total Variation distance and Tauberian theorems.