Benjamin Legros
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Benjamin Legros has a total of 35 co-authors. Here are the top co-authors by number of shared publications:
- Oualid Jouini
Université Paris‐Saclay, CentraleSupélec, Laboratoire Génie Industriel Gif‐sur‐Yvette France
15 shared publications - Ger Koole
Department of Mathematics, Vrije Universiteit Amsterdam, Amsterdam, The Netherlands
6 shared publications - Jan C. Fransoo
Tilburg University
5 shared publications Oualid Jouini
Laboratoire Genie Industriel, Université Paris-Saclay, Gif-sur-Yvette, France
5 shared publications- Ali Devin Sezer
Middle East Technical University
2 shared publications - Bingnan Lu
Univ Minnesota, Dept Ind & Syst Engn, Minneapolis, MN 55455 USA
2 shared publications
Latest Publications
- 12 October 2026
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in Operations Research
Convexity of Inverse Blocking Curves in Loss Systems
Why Larger Service Systems Benefit More from Scale How much extra demand can a service system handle when capacity is increased while keeping the same target level of customer blocking? In the paper “Convexity of Inverse Blocking Curves in Loss Systems,” Benjamin Legros establishes an exact economies-of-scale result for loss systems. The paper proves that the maximum supportable demand is an increasing and globally convex function of the number of servers. In practical terms, as a system becomes larger, each additional server can support progressively more additional demand while maintaining the same blocking target. Equivalently, the extra capacity required to accommodate additional demand decreases as the system grows. The result provides a rigorous mathematical foundation for resource pooling in capacity-constrained services. It applies to the classical Erlang-B loss model and extends to systems with a finite waiting room, with potential relevance for telecommunications, healthcare, emergency response, and other service operations.
- 12 October 2026
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in Operations Research
Convexity of Inverse Blocking Curves in Loss Systems
Why Larger Service Systems Benefit More from Scale How much extra demand can a service system handle when capacity is increased while keeping the same target level of customer blocking? In the paper “Convexity of Inverse Blocking Curves in Loss Systems,” Benjamin Legros establishes an exact economies-of-scale result for loss systems. The paper proves that the maximum supportable demand is an increasing and globally convex function of the number of servers. In practical terms, as a system becomes larger, each additional server can support progressively more additional demand while maintaining the same blocking target. Equivalently, the extra capacity required to accommodate additional demand decreases as the system grows. The result provides a rigorous mathematical foundation for resource pooling in capacity-constrained services. It applies to the classical Erlang-B loss model and extends to systems with a finite waiting room, with potential relevance for telecommunications, healthcare, emergency response, and other service operations.
- 9 October 2026
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in INFORMS Journal on Computing
Stationary Coupled-Flow Approximation for Time-Varying Retrial Queues
Although stationary queueing models, such as Erlang-C and Erlang-A, are widely used for performance evaluation in service systems, they do not account for time dependence and retrials, which limits their applicability. Building on stationary [Formula: see text] formulas and point-wise stationary approximations (PSAs), this paper proposes the stationary coupled-flow (SCF) approximation. The method partitions the time horizon into intervals, models each as an [Formula: see text] in steady state, and iteratively determines the arrival rate, accounting for new customers and retrials. This yields an approximation of long-run performance metrics of retrial queues with stationary or time-varying arrivals. We prove that the SCF approximation converges to a finite limit for constant and cyclic arrival rates. Extensive numerical experiments demonstrate the high accuracy of the SCF approximation. Furthermore, SCF outperforms PSA, particularly during peak periods. Therefore, SCF provides a fast, transparent tool for planning and staffing in services with retrials. It remains reliable when the exact distribution of retrials is uncertain, enabling day-to-day scheduling and scenario testing under time-varying demand. We find that the retrial probability plays a stronger role than the retrial-time distribution. The impact of retrials in time-varying queues is driven by two effects: the state at the abandonment epoch and the cycle length-retrial time interaction. We posit that as retrial times increase, the first effect dampens congestion feedback, whereas the second effect can amplify or smooth fluctuations. Under sufficiently high variability in retrial times, the cycle length-retrial time interaction becomes negligible, leaving the system governed by the abandonment-epoch state. History: Accepted by Shane Henderson, Area Editor for Simulation, Stochastic Models, & Stochastic Optimization. Funding: This work was supported by the Deutsche Forschungsgemeinschaft [Grant 522848863]. Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information ( https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2025.1626 ) as well as from the IJOC GitHub software repository ( https://github.com/INFORMSJoC/2025.1626 ). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/ .
- 9 October 2026
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in INFORMS Journal on Computing
Stationary Coupled-Flow Approximation for Time-Varying Retrial Queues
Although stationary queueing models, such as Erlang-C and Erlang-A, are widely used for performance evaluation in service systems, they do not account for time dependence and retrials, which limits their applicability. Building on stationary [Formula: see text] formulas and point-wise stationary approximations (PSAs), this paper proposes the stationary coupled-flow (SCF) approximation. The method partitions the time horizon into intervals, models each as an [Formula: see text] in steady state, and iteratively determines the arrival rate, accounting for new customers and retrials. This yields an approximation of long-run performance metrics of retrial queues with stationary or time-varying arrivals. We prove that the SCF approximation converges to a finite limit for constant and cyclic arrival rates. Extensive numerical experiments demonstrate the high accuracy of the SCF approximation. Furthermore, SCF outperforms PSA, particularly during peak periods. Therefore, SCF provides a fast, transparent tool for planning and staffing in services with retrials. It remains reliable when the exact distribution of retrials is uncertain, enabling day-to-day scheduling and scenario testing under time-varying demand. We find that the retrial probability plays a stronger role than the retrial-time distribution. The impact of retrials in time-varying queues is driven by two effects: the state at the abandonment epoch and the cycle length-retrial time interaction. We posit that as retrial times increase, the first effect dampens congestion feedback, whereas the second effect can amplify or smooth fluctuations. Under sufficiently high variability in retrial times, the cycle length-retrial time interaction becomes negligible, leaving the system governed by the abandonment-epoch state. History: Accepted by Shane Henderson, Area Editor for Simulation, Stochastic Models, & Stochastic Optimization. Funding: This work was supported by the Deutsche Forschungsgemeinschaft [Grant 522848863]. Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information ( https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2025.1626 ) as well as from the IJOC GitHub software repository ( https://github.com/INFORMSJoC/2025.1626 ). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/ .
- 3 September 2026
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in INFORMS Journal on Computing
Code and Data Repository for Stationary Coupled-Flow Approximation for Time-Varying Retrial Queues
The software and data in this repository are a snapshot of the software that was used in the research reported in the paper Stationary Coupled-Flow Approximation for Time-Varying Retrial Queues by Ö. Schmiel, B. Legros, and R. Stolletz.
- 3 September 2026
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in INFORMS Journal on Computing
Code and Data Repository for Stationary Coupled-Flow Approximation for Time-Varying Retrial Queues
The software and data in this repository are a snapshot of the software that was used in the research reported in the paper Stationary Coupled-Flow Approximation for Time-Varying Retrial Queues by Ö. Schmiel, B. Legros, and R. Stolletz.
- 3 May 2026
- Published by SAGE Publications in Production and Operations Management
- Vol. 35 (11) , 4096-4110
When more capacity creates more congestion: The role of risk aversion
A fundamental principle in operations management holds that increasing the number of servers reduces delays in service systems. To date, no mechanism has been identified that could reverse this effect. We propose, however, that risk aversion can cause increased service capacity to intensify congestion. We study an unobservable M / M / s queue where risk-averse customers choose whether to join based on their anticipated waiting time. We show that, in equilibrium, demand, expected waiting time, expected sojourn time, and the probability of waiting all increase with the number of servers, and that these effects are stronger for more risk-averse customers. We further uncover the mechanism behind this phenomenon: adding servers makes delays less risky (in the sense of second-order stochastic dominance), which increases the sensitivity of demand to capacity as customers become more risk-averse. These patterns are more prevalent in small systems and fade as the system grows. They can also persist when customers differ in their degree of risk aversion, when capacity is increased by raising service speed, and when the system is observable. Our findings reveal a novel trade-off created by customer risk aversion: expanding capacity attracts more customers, but also exacerbates congestion. A manager aiming to reduce waiting times may therefore prefer to de-pool service capacity instead of following the standard approach of pooling, while increasing capacity in the resulting smaller systems to preserve total throughput. When the objective is to maximize profitability, our results further suggest that the cost of additional servers may be offset by the associated increase in revenue when customers are sufficiently risk-averse.
- 3 May 2026
- Published by SAGE Publications in Production and Operations Management
- Vol. 35 (11) , 4096-4110
When more capacity creates more congestion: The role of risk aversion
A fundamental principle in operations management holds that increasing the number of servers reduces delays in service systems. To date, no mechanism has been identified that could reverse this effect. We propose, however, that risk aversion can cause increased service capacity to intensify congestion. We study an unobservable M / M / s queue where risk-averse customers choose whether to join based on their anticipated waiting time. We show that, in equilibrium, demand, expected waiting time, expected sojourn time, and the probability of waiting all increase with the number of servers, and that these effects are stronger for more risk-averse customers. We further uncover the mechanism behind this phenomenon: adding servers makes delays less risky (in the sense of second-order stochastic dominance), which increases the sensitivity of demand to capacity as customers become more risk-averse. These patterns are more prevalent in small systems and fade as the system grows. They can also persist when customers differ in their degree of risk aversion, when capacity is increased by raising service speed, and when the system is observable. Our findings reveal a novel trade-off created by customer risk aversion: expanding capacity attracts more customers, but also exacerbates congestion. A manager aiming to reduce waiting times may therefore prefer to de-pool service capacity instead of following the standard approach of pooling, while increasing capacity in the resulting smaller systems to preserve total throughput. When the objective is to maximize profitability, our results further suggest that the cost of additional servers may be offset by the associated increase in revenue when customers are sufficiently risk-averse.
- 2 April 2026
- Published by SAGE Publications in Production and Operations Management
- Vol. 35 (10) , 3750-3769
EXPRESS: Waiting Time Externality Awareness in Socially Responsible Queueing
In service systems, customers tend to overutilize resources, resulting in high congestion. We study a noncoercive way to alleviate congestion based on customers’ awareness of the negative waiting externalities they impose. To this end, we revisit Naor’s rational-queue model by adding a term that captures the cost a customer associates with the waiting time generated for other customers encountered during their sojourn. First, we show that this modification preserves the threshold structure of customers’ joining strategy. Using lattice-path counting techniques, we derive a closedform expression for the expected generated waiting time, conditional on the number of customers observed upon arrival. The generated wait is increasing and concave in the queue length at arrival. Viewing expected waiting time as allocated across arrivals, the queue length associated with an arbitrary unit of expected generated waiting time is first-order stochastically lower than the queue length associated with a unit of expected personal waiting time. We prove that the maximum generated wait is increasing and convex in the joining threshold and provide tight bounds with exact light/heavy traffic asymptotics. Internalizing social awareness (weight on generated wait) lowers the joining threshold and reduces waiting; in high-demand regimes, it can also increase social welfare. When the operator can also influence the weight on personal wait, that lever reduces congestion more strongly than waiting externality awareness. Finally, we quantify the awareness level needed to reach Naor’s social threshold and show that it is lowest at a non-extreme demand level.
- 2 April 2026
- Published by SAGE Publications in Production and Operations Management
- Vol. 35 (10) , 3750-3769
EXPRESS: Waiting Time Externality Awareness in Socially Responsible Queueing
In service systems, customers tend to overutilize resources, resulting in high congestion. We study a noncoercive way to alleviate congestion based on customers’ awareness of the negative waiting externalities they impose. To this end, we revisit Naor’s rational-queue model by adding a term that captures the cost a customer associates with the waiting time generated for other customers encountered during their sojourn. First, we show that this modification preserves the threshold structure of customers’ joining strategy. Using lattice-path counting techniques, we derive a closedform expression for the expected generated waiting time, conditional on the number of customers observed upon arrival. The generated wait is increasing and concave in the queue length at arrival. Viewing expected waiting time as allocated across arrivals, the queue length associated with an arbitrary unit of expected generated waiting time is first-order stochastically lower than the queue length associated with a unit of expected personal waiting time. We prove that the maximum generated wait is increasing and convex in the joining threshold and provide tight bounds with exact light/heavy traffic asymptotics. Internalizing social awareness (weight on generated wait) lowers the joining threshold and reduces waiting; in high-demand regimes, it can also increase social welfare. When the operator can also influence the weight on personal wait, that lever reduces congestion more strongly than waiting externality awareness. Finally, we quantify the awareness level needed to reach Naor’s social threshold and show that it is lowest at a non-extreme demand level.
- 31 March 2026
- Published by Agence Bibliographique de l'Enseignement Supérieur
Optimization of multi-channel and multi-skill call centers
Optimisation dans les centres d'appels multi-compétences et multi-canaux: Les centres d’appels connaissent un grand succès depuis leur introduction dans les entreprises de service. Ils sont le principal point de contact avec les clients, et une composante essentielle de la majorité des entreprises. L’émergence des centres d’appels à grande échelle a suscité de nombreuses problématiques de management. Dans cette thèse, nous considérons des problématiques de management orientées sur les centres multi‐canaux et multi‐compétences. L’objectif de notre travail est de trouver des résultats qualitatifs et quantitatifs utiles pour le management. Dans la première partie, nous considérons les architectures de centres multi‐compétences à flexibilité limitée. Le contexte est celui de centres d’appels avec des paramètres asymétriques : charge de travail non équilibrée, différents temps de services, prédominance d’une catégorie de clients, taux d’abandons variables et couts élevés de la multi‐compétence. Les architectures les plus connues avec flexibilité limitée comme chaining ne résistent pas à de telles asymétries. Nous proposons une nouvelle architecture, appelée Single Pooling avec seulement deux compétences par agent et nous démontrons son efficience dans diverses situations d’asymétrie. Dans la seconde partie, nous nous intéressons aux problèmes de routage dans les centres d’appels multi‐canaux. Dans le premier projet, nous considérons un centre avec des appels arrivant au fil du temps et des emails présents en nombre illimité. Le service des appels se fait en trois étapes dont la seconde est une pause pour l’agent. Nous cherchons à optimiser le routage des emails. L’objectif est de maximiser le débit d’emails traités sous contrainte de temps d’attente pour les appels. De nombreuses recommandations sont proposées au manager. En particulier, nous démontrons que pour obtenir un routage optimal il est nécessaire de fixer à une valeur extrême au moins l’un des deux paramètres définissant le routage des emails. Dans le second projet, nous étudions une politique de seuil de réservation d’agents pour les appels en réception. Nous considérons un cas général de modèle non stationnaire ou le processus d’arrivée des appels est Poisson non homogène. Le problème d’optimisation est la maximisation du débit de taches en émission sous contrainte de qualité de service sur les appels en réception. Nous proposons une méthode efficiente et facile à implémenter de changement adaptatif de seuil. Cette politique est évaluée en comparaison avec les performances optimales trouvées dans le cas particulier de taux d’arrivée constant, et en comparaison avec d’autres méthodes intuitives de changement adaptatif de seuil dans le cas général non stationnaire. Dans le troisième projet, nous considérons un modèle de centre avec option de rappels. Cette option permet de transformer un appel en réception en un appel en attente d’émission. Le problème d’optimisation est la minimisation du temps d’attente des appels en émission sous contrainte de qualité de service pour les appels en réception. Nous proposons une politique de routage à deux seuils, un sur la réservation d’agents pour les appels en réception et un sur le nombre d’appels en attente d’émission. Nous déterminons une courbe optimale entre ces deux seuils
- 31 March 2026
- Published by Agence Bibliographique de l'Enseignement Supérieur
Optimization of multi-channel and multi-skill call centers
Optimisation dans les centres d'appels multi-compétences et multi-canaux: Les centres d’appels connaissent un grand succès depuis leur introduction dans les entreprises de service. Ils sont le principal point de contact avec les clients, et une composante essentielle de la majorité des entreprises. L’émergence des centres d’appels à grande échelle a suscité de nombreuses problématiques de management. Dans cette thèse, nous considérons des problématiques de management orientées sur les centres multi‐canaux et multi‐compétences. L’objectif de notre travail est de trouver des résultats qualitatifs et quantitatifs utiles pour le management. Dans la première partie, nous considérons les architectures de centres multi‐compétences à flexibilité limitée. Le contexte est celui de centres d’appels avec des paramètres asymétriques : charge de travail non équilibrée, différents temps de services, prédominance d’une catégorie de clients, taux d’abandons variables et couts élevés de la multi‐compétence. Les architectures les plus connues avec flexibilité limitée comme chaining ne résistent pas à de telles asymétries. Nous proposons une nouvelle architecture, appelée Single Pooling avec seulement deux compétences par agent et nous démontrons son efficience dans diverses situations d’asymétrie. Dans la seconde partie, nous nous intéressons aux problèmes de routage dans les centres d’appels multi‐canaux. Dans le premier projet, nous considérons un centre avec des appels arrivant au fil du temps et des emails présents en nombre illimité. Le service des appels se fait en trois étapes dont la seconde est une pause pour l’agent. Nous cherchons à optimiser le routage des emails. L’objectif est de maximiser le débit d’emails traités sous contrainte de temps d’attente pour les appels. De nombreuses recommandations sont proposées au manager. En particulier, nous démontrons que pour obtenir un routage optimal il est nécessaire de fixer à une valeur extrême au moins l’un des deux paramètres définissant le routage des emails. Dans le second projet, nous étudions une politique de seuil de réservation d’agents pour les appels en réception. Nous considérons un cas général de modèle non stationnaire ou le processus d’arrivée des appels est Poisson non homogène. Le problème d’optimisation est la maximisation du débit de taches en émission sous contrainte de qualité de service sur les appels en réception. Nous proposons une méthode efficiente et facile à implémenter de changement adaptatif de seuil. Cette politique est évaluée en comparaison avec les performances optimales trouvées dans le cas particulier de taux d’arrivée constant, et en comparaison avec d’autres méthodes intuitives de changement adaptatif de seuil dans le cas général non stationnaire. Dans le troisième projet, nous considérons un modèle de centre avec option de rappels. Cette option permet de transformer un appel en réception en un appel en attente d’émission. Le problème d’optimisation est la minimisation du temps d’attente des appels en émission sous contrainte de qualité de service pour les appels en réception. Nous proposons une politique de routage à deux seuils, un sur la réservation d’agents pour les appels en réception et un sur le nombre d’appels en attente d’émission. Nous déterminons une courbe optimale entre ces deux seuils
- Open Access
- 8 April 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 281
Machine learning-based agent staffing under uncertainty: The case of a relay call center
- 8 April 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 281
Open AccessMachine learning-based agent staffing under uncertainty: The case of a relay call center
- 26 February 2025
- Published by SAGE Publications in Production and Operations Management
- Vol. 34 (8) , 2413-2429
Managing Reusable Resources With Usage Time Limits
In reusable resources like vehicle sharing, city parking, and hotel services, resource units are used by consecutive customers for a stochastic usage time and are made available for future customers upon return. We model such reusable resources as Erlang loss systems, where customers who find all units occupied are blocked. Customers are rational decision-makers, and their willingness to join the system depends on factors such as price, time limits, service value, and resource availability. When each customer is charged a fixed price, we examine whether it is beneficial to install time limits, which increase availability at the expense of reduced usage time. We formulate and solve a two-dimensional revenue maximization problem to optimize both the usage price and the time limit. We differentiate between on-demand resources, which are immediately reusable upon return, and reserved resources, which can only be reused after the time limit expires. With on-demand resources, setting a time limit is proven to be detrimental, making the entry price the sole tool for maximizing revenue. This conclusion remains valid when the usage time increases with the entry price. In contrast, managing reserved resources involves an interplay between time limits and entry prices. Additionally, we show that loss systems (where customers do not wait) should be managed differently from systems where customers do wait. Specifically, in systems with waiting, we prove that only the time limit should be optimized, while the entry price should capture the full market share. We prove our results by analyzing the behavior of the equilibrium arrival rate and revenue, which necessitates establishing novel properties for the blocking probability. These properties yield new sharp bounds for the blocking probability and enable us to prove the convergence of the optimized model for on-demand resources to the quality-and-efficiency-drivenregime as the number of resource units grows.
- 26 February 2025
- Published by SAGE Publications in Production and Operations Management
- Vol. 34 (8) , 2413-2429
Managing Reusable Resources With Usage Time Limits
In reusable resources like vehicle sharing, city parking, and hotel services, resource units are used by consecutive customers for a stochastic usage time and are made available for future customers upon return. We model such reusable resources as Erlang loss systems, where customers who find all units occupied are blocked. Customers are rational decision-makers, and their willingness to join the system depends on factors such as price, time limits, service value, and resource availability. When each customer is charged a fixed price, we examine whether it is beneficial to install time limits, which increase availability at the expense of reduced usage time. We formulate and solve a two-dimensional revenue maximization problem to optimize both the usage price and the time limit. We differentiate between on-demand resources, which are immediately reusable upon return, and reserved resources, which can only be reused after the time limit expires. With on-demand resources, setting a time limit is proven to be detrimental, making the entry price the sole tool for maximizing revenue. This conclusion remains valid when the usage time increases with the entry price. In contrast, managing reserved resources involves an interplay between time limits and entry prices. Additionally, we show that loss systems (where customers do not wait) should be managed differently from systems where customers do wait. Specifically, in systems with waiting, we prove that only the time limit should be optimized, while the entry price should capture the full market share. We prove our results by analyzing the behavior of the equilibrium arrival rate and revenue, which necessitates establishing novel properties for the blocking probability. These properties yield new sharp bounds for the blocking probability and enable us to prove the convergence of the optimized model for on-demand resources to the quality-and-efficiency-drivenregime as the number of resource units grows.
- 21 December 2024
- Published by Taylor & Francis in Journal of the Operational Research Society
- Vol. 76 (7) , 1449-1465
How should rejection be used to maximize congestion while preserving idle time?
- 21 December 2024
- Published by Taylor & Francis in Journal of the Operational Research Society
- Vol. 76 (7) , 1449-1465
How should rejection be used to maximize congestion while preserving idle time?
- 19 November 2024
- Published by Elsevier in Operations Research Letters
- Vol. 58
Pooled testing in the presence of congestion
- 19 November 2024
- Published by Elsevier in Operations Research Letters
- Vol. 58
Pooled testing in the presence of congestion