Netherlands OML Conference 2024​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​

Abstracts

Session 1: Transportation and Traffic Management

Ilke Bakir (RUG) – Optimizing Drone-Assisted Last-Mile Deliveries: The Vehicle Routing Problem with Flexible Drones

We study a hybrid delivery system in which (multiple) trucks and (multiple) drones operate in tandem. In particular, we introduce the vehicle routing problem with flexible drones (VRPFD), which seeks to find a set of delivery routes for a fleet of trucks and drones operating in synchronization, with the goal of minimizing the makespan. We formulate the VRPFD as a mixed integer linear program on a time-space network, and present an efficient optimization algorithm based on a dynamic discretization discovery approach. We demonstrate the benefits of drone flexibility and the efficiency of the proposed solution approach through a detailed computational study performed on a newly generated set of benchmark instances. Our findings suggest that the flexible use of drones facilitates higher drone utilization and therefore results in makespan improvements. In clustered geographies, drone flexibility reduces the makespan by up to 12.12%, with an average of 5.39%. Our proposed solution approach is able to efficiently solve the VRPFD instances by making use of an intelligent lower bounding mechanism while keeping its subproblem small. Computational experiments reveal that it is able to reduce the solution time by up to a factor of 6.5 when compared to solving the VRPFD using a commercial solver.

Marco Rinaldi (TUD) – A hybrid data-driven/optimization heuristic for dynamic boundary perimeter control in network traffic management

Cities are experiencing uninterrupted growth, accompanied by ever increasing demand for transport, and accompanying issues such as congestion, pollution, and reduced livability. This is leading policymakers to seek for operational and tactical approaches aimed, for example, at reducing car usage and car ownership. Maintaining accessibility (as defined by e.g. average travelling time) while promoting sustainable mobility remains a challenging endeavor.

From an operations perspective, Intelligent Transportation Systems have become a very promising tool to help addressing this challenge, empowered by EU-wide efforts in harmonized data collection (e.g. the European Mobility Data Space), attention towards automation (CCAM initiative) etc. A well-studied approach in transportation literature to help manage excessive demand towards e.g. Central Business Districts (CBD) is Perimeter Control. This has however found little application in practice, due to expensive implementation, and inherent static nature, making it insufficiently responsive to disturbances proper to transportation networks. In this work, we combine a data-driven approach to estimate spatio-temporal distribution of congestion in urban networks, based on a Gaussian Mixture Model, with an optimization approach expressing qualitative policy criteria to help redistribute traffic away from congested areas, in a state-responsive fashion. We present results based on microscopic traffic flow simulation, carried out in DLR SUMO.

Dennis Huisman (EUR) – OR in Railways: Achievements and Challenges

Operations Research (OR) models and algorithms play a major role in railway planning and operations. Since 1990, many achievements have been made to improve the railways, not only in the Netherlands but also around the world. In this talk, we will present some examples of successful OR applications at Netherlands Railways (NS). We conclude the presentation with some important challenges that are still unsolved and need further research.

Session 2: Data Analytics

Derya Demirtas (UT) – Unlocking the Value of Extensive Data: Estimating spatial cardiac arrest risk to guide resource allocation decisions

Out-of-hospital cardiac arrest (OHCA) is a significant public health problem with notably low survival rates. Early defibrillation is crucial for survival, highlighting the importance of nearby automated external defibrillators (AEDs).  Current AED placement strategies often rely on historical OHCA data, which are limited in availability.  Publicly available demographic/socioeconomic data are often easily available and shown to have correlations with OHCA risk. This study aims to 1) estimate spatial cardiac arrest risk using demographic/socioeconomic data alone 2) compare AED location models based solely on estimated risk with those incorporating historical OHCA data to inform demand. Machine learning techniques were applied to a comprehensive dataset spanning multiple municipalities. Predicted OHCA incidence of each district were used to optimize AED locations, alongside AED optimization models that used smoothed out historical cardiac arrest data as demand. Results on several municipalities underscore the value of an OHCA registry. Nonetheless, in its absence, machine learning models leveraging demographic and socioeconomic data offer a viable means to substantially enhance coverage.

Müge Tekin (RSM) – Estimation using marginal competitor sales information

We tackle the estimation problem under a market-share model, focusing on the hotel industry, and develop methodologies to overcome the following significant challenges: (i) competitor’s data is aggregated across multiple LOS with distinct demands (ii) we do not observe no-purchasers, i.e. those who purchase neither ours nor the competitor’s products, and finally, (iii) the competitor makes private sales to groups before the retail sales period; thus even the competitor’s capacity is unobservable. 

Jan Fransoo, Department of Information Systems & Operations Management, School of Economics and Management, Tilburg University - Econometrics for data analytics in novel environments

Over the past five years, Jan Fransoo and his team have been conducting extensive empirical research on nanoretailing in developing countries, mainly in Latin America and Africa. In this talk, Jan will share a brief introduction to the context, the strategies in acquiring data with companies, and how to stay up to date on the rapidly evolving econometric methods used to analyze these data.

Session 3: Decision Support for Maintenance Logistics

Collin Drent (TU/e) - Condition-based production for stochastically deteriorating systems

Production systems deteriorate stochastically due to usage and may eventually break down, resulting in high maintenance costs at scheduled maintenance moments. This deterioration behavior is affected by the system's production rate. While producing at a higher rate generates more revenue, the system may also deteriorate faster. Production can (and should) thus be controlled dynamically to trade-off deterioration and revenue accumulation in between the moments at which maintenance is performed on the system. We study systems for which the relation between production and deterioration is known and the same for each system as well as systems for which this relation differs from system to system and needs to be learned on-the-fly. For systems with a known production-deterioration relation, we cast the decision problem as a continuous-time Markov decision process and prove that the optimal policy has intuitive monotonic properties. We also present sufficient conditions for the optimality of so-called bang-bang production policies. For systems that exhibit variability in their production-deterioration relations, we propose a Bayesian procedure to learn the unknown deterioration rate under any production policy. Numerical studies demonstrate the profitability of our proposed approach in practical situations compared to existing state-of-the-art methods.

El-Houssaine Aghezzaf (UGent): Integration of production and maintenance planning

In this talk we discuss the problem of integrating production and preventive/predictive maintenance plans in a multimachine shop-floor, where machines are subject to degradation. We will first discuss the cases where failure rate distributions of the machines are known and that preventive maintenance is essential to keep the system’s yield at an adequate level. Next, we examine the situations in which machines failure data is available but the underlying distributions are unknown, and then discuss possible robust models and related solution methods for these later cases.

Ruud Teunter (RUG) – Data-driven condition-based maintenance

With the rise of Industry 4.0, smart industries and the use of (sensor) data, research on condition based maintenance is also getting more and more attention. Surprisingly, though, the literature on maintenance optimization by and large ignores the data analytics part, and simply assumes perfect information on the deterioration process and on what deterioration level implies a system failure. In response, we have developed new approaches that are fully data driven. I will present these and discuss their performance.

Session 4: Advances in Retail Operations

Maxi Udenio (KULeuven) – Global vs local ML models in retail forecasting

Accurate sales forecasts play a crucial role in supporting operations, and one notable recent development has been the heightened enthusiasm for global forecasting methods. These approaches have demonstrated remarkable competitiveness when applied to real-world datasets. By fitting a single method over multiple time series, global methods have access to more training data seemingly leading to better generalization than traditional local forecasting methods, which estimate a single method per time series. This prompts the question of whether pooling time series should be consistently practiced when groups of time series are available, as this would constitute a significant departure from established practice. This paper assesses the performance of (non-) linear local and global forecasting methods on both homogeneous and heterogeneous datasets by conducting extensive simulations and examining a real-world retailing case study. Our results show that heterogeneity complicates parameter estimation for global methods, making correctly specified local methods hard to outperform. Simply providing more training data and lags is not sufficient to perform competitively. However, when training data is limited, global methods mostly outperform their local counterpart. We find that non-linear methods handle heterogeneity better than their linear counterpart out-of-the-box. This difference in forecast accuracy can be offset by including inputs based on problem-specific feature engineering. While it shows that linear global methods are also able to successfully deal with heterogeneity, the required feature engineering is hard to implement for …

Rene Haijema (WUR) – Optimal dynamic discounting of fresh products at a supermarket

Consumer tend to not pick products that are close to or at the expiration date, either because these are considered to be less fresh, or because they offer less planning flexibility. By putting a discount on the oldest products, a retailer entices consumer to pick an old discounted product over younger products. This so-called expiration date based pricing strategy is being used by supermarkets to reduce food waste without knowing the precise impact on the retailer’s profit. In practice, often all products that expire by the end of the day gets discounted by a fixed percentage (e.g. 35%) of the sales price. We show that this is often far from optimal. In our study we determine optimal discount levels for a product with a fixed expiration date of say m days. We consider the case of discounting only at the last day, i.e., at the expiration date, as well as the case of discounting at the last two days using two different discounting levels. By Stochastic dynamic programming we solve an underlying Markov Decision Problem of determining the optimal discounts at the last two expiration dates. Results show, discount levels depend on the actual number of products in stock as well as whether a discount increases the total demand. Often times, it is optimal to not discount all at, especially when discounting is only changing the picking order of products but not the total demand. Furthermore, optimal discount levels turn out to be much lower than those in practice, and average discount levels are lower when discount at the last two days.

Bram de Moor (KULeuven) - Reward shaping to improve the performance of deep reinforcement learning for perishable inventory management

Deep reinforcement learning (DRL) has proven to be an effective, general-purpose technology to develop ‘good’ replenishment policies in inventory management. We show how transfer learning from existing, well-performing heuristics may stabilize the training process and improve the performance of DRL in inventory control. While the idea is general, we specifically implement potential-based reward shaping to a deep Q-network algorithm to manage inventory of perishable goods that, cursed by dimensionality, has proven to be notoriously complex. The application of our approach may not only improve inventory cost performance and reduce computational effort, the increased training stability may also help to gain trust in the policies obtained by black box DRL algorithms.

Session 5: Energy

Fiona Sloothaak, Agnieszka Janicka, Maria Vlasiou and Bert Zwart - Emergence of scale-free blackouts: the effect of network topology

The occurrence of blackouts stemming from cascading line failures within transmission grids presents a critical challenge in understanding complex system dynamics and interdependencies. This study delves into the phenomenon, investigating why blackouts often manifest as scale-free events despite their perceived rarity. We explore the hypothesis that this emergence is intricately linked to heavy-tailed power demands within the grid. Interestingly, this hypothesis implies there are only a select few likely cascading line failure sequences that could occur for a large-scale blackout to occur. We consider these cascade sequences for several network topologies, providing initial insights on how network topology affects the robustness and reliability of power grids.

Adrian Esteban Perez, Yashar Ghiassi-Farrokhfal and Derek Bunn - Load forecasting and demand flexibility via inverse optimization

A method to forecast the demand and flexibility level of consumers of electricity is presented. The price-response model is defined by an optimization program whose defining parameters are represented by time series of prices, generation and minimum and maximum load flexibility levels. These parameters are, in turn, estimated from observational data by exploiting an approach based on duality theory. The proposed methodology is data-driven and exploits information from covariates via Kernel regression functions, such as price, and weather variables, to account the non-linearity for changes in the parameter estimates. The resulting estimation problem is a tractable mixed-integer program with conic constraints. Furthermore, we illustrate the benefits of the proposed model to forecast the demand of customers and the flexibility levels in a real dataset.

Jasper Bakker, Jose A Lopez and Paul Buijs - A network design perspective on the adoption potential of electric road systems in early development stages

The electrification of heavy freight trucks presents a significant challenge in the global push to decarbonize the transport sector. This study explores the deployment of Electric Road Systems (ERS) as a potential solution. We propose a novel methodology, grounded in rich empirical data, and apply it to the potential deployment of ERS infrastructure in the Netherlands. The primary aim is to analyze the adoption potential of different ERS network designs during the early stages of ERS development, specifically contrasting dense infrastructures with longer ERS corridors. The results show that corridors offer superior performance in terms of the distance traveled over the ERS infrastructure while dense networks attract more, but shorter trips. These differences in performance indicators have important implications for policymakers when considering the financial viability and environmental impact of different ERS network designs.

Session 6: Sustainable Operations

Ece Gülserliler (TiU) - Business Model Choice under Right-to-Repair: Economic and Environmental Implications

Right-to-Repair (RTR) regulations require producers to supply necessary information and parts for consumers to independently undertake repairs. While these regulations aim to prolong product lifetimes through repairs, increase secondhand use, and reduce waste; the ease of access to proprietary information and spare parts can inadvertently facilitate intellectual property (IP) rights infringement by third parties. We investigate potential producer strategies to respond to this risk. In particular, we analyze when business model choice between ownership and non-ownership can be a viable strategy for producers to protect their IP and preempt competition with third parties and secondary markets. Considering this choice, we examine the consequences of RTR for producers, consumers, and the environment.

Melvin Drent (TU/e) – Efficient emission reduction through dynamic supply mode selection

Reducing the carbon footprint of global supply chains is a challenge for many companies. Governmental emission regulations are increasingly stringent, and consumers are increasingly environmentally conscious. Companies should therefore integrate carbon emissions in their supply chain decision making. In this paper, we study the inbound supply mode and inventory management decision making for a company that sells an assortment of products. Stochastic demand for each product arrives periodically and unmet demand is backlogged. Each product has two distinct supply modes that differ in terms of their carbon emissions, speed, and costs. The company needs to decide when to ship how much using which supply mode such that total holding, backlog, and procurement costs are minimized while the emissions associated with different supply modes across the assortment remains below a target level. We formulate this decision problem as a mixed integer linear program that we solve through decomposition. We benchmark our decision model against two state-of-the-art approaches in a large test-bed based on real-life carbon emissions data. Relative to our decision model, the first benchmark lacks the flexibility to dynamically ship products with two supply modes while the second benchmark makes supply mode decisions for each product individually. Our computational experiment shows that our decision model can outperform the first and second benchmark by up to 15 and 40 percent, respectively, for moderate carbon emission reduction targets.

Paul Buijs (RUG) - Beyond the Green Label: Life Cycle Assessment of Cargo Bikes in Last-Mile Delivery

This talk discusses the carbon footprint impact of cargo bikes in last-mile delivery, considering all life cycle stages, including production, operation, and disposal. Previous life cycle studies often confine their analysis to direct comparisons between sustainable vehicles and their conventional counterparts, assuming similar operational deployment. However, cargo bikes often demand a distinct operational approach. Our study responds to this need by employing a systemic methodology that specifically accounts for these operational distinctions. This approach not only refines existing methods of life cycle assessment but also deepens the understanding of 'green' labels within urban freight transport.

Session 7: Healthcare Management

Gréanne Leeftink (UT) - Predicting next week's bed census: combining medical expertise with data

(Authors: Hayo Bos, Stef Baas, Richard Boucherie, Erwin Hans, Gréanne Leeftink)

Bed census predictions play a key role in hospital capacity management decisions, such as ward dimensioning, staffing decisions, surgery scheduling and discharge planning. Although there exists much literature on predicting a patient’s Length-of-Stay (LoS), based on a variety of individual and contextual features, the census predictions in practice are typically based on the healthcare professional’s estimate of the Expected Due Date (EDD). In this talk we propose two probabilistic models to combine the EDD with LoS distributions. Using Poisson Binomial distribution and probabilistic convolution, we obtain a full census distribution. We show the accurate working of our method with real hospital data, demonstrate the benefits of combining LoS and EDD information, and discuss the model’s implementation in practice.

Cynthia Kong (EUR) - When Does Distribution Matter? The Impact of Last-Mile Vaccine Delivery Systems on Vaccine Uptake

The last-mile distribution of vaccines is crucial in pandemic control, impacting roll-out speed and vaccination uptake. To design vaccine distribution systems during varying pandemic stages, it is key to understand how its operational attributes – appointment flexibility, appointment delay, in-facility waiting time, travel time, and familiarity with personnel – influence vaccination uptake. COVID-19 provided a unique opportunity to examine this. Utilizing discrete-choice experiments across three European countries with distinct vaccination uptake rates (the Netherlands, Italy and Poland), we found that the distribution system significantly impacts vaccine uptake, and the magnitude of this impact strongly changes over time. In 2021 – during the 2021 epidemic phase – the effect was minor (up to 2.4%), supporting a focus on efficiency. Conversely, operational attributes substantially affected uptake (up to 55.9%) in 2023, when COVID-19 transitioned into the endemic phase, emphasizing the need to optimize the distribution system to meet people’s preferences. The most important attributes are travel time, appointment delay, and in-facility waiting time. These findings also hold for two key priority populations – the elderly and those in poor health. Our insights contrast with many countries’ last-mile distribution systems during COVID-19 and provide actionable insights for inevitable future mass vaccination campaigns.

Sebastian Kraul (VU) - Optimizing physician schedules with resilient break assignments

In my talk, I will present a novel model for building biweekly rosters for physicians according to the regulations of a German teaching hospital while also ensuring the viability of breaks. Currently, rosters are manually prepared by experienced physicians with basic spreadsheet knowledge, leading to significant costs and time consumption because of the complexity of the problem and the individual working conditions of the physicians. Unfortunately, manually generated rosters frequently prove to be non-compliant with labor regulations and ergonomic agreements, resulting in potential overtime hours and employee dissatisfaction. A particular concern is the inability of physicians to take mandatory breaks, which negatively affects both employee motivation and the hospital service level. To address these challenges, we propose a data-driven formulation of an operational physician scheduling problem, considering overstaffing and overtime hours as primary cost drivers and integrating shift preferences and break viability as ergonomic objectives. Given the limitations of standard solvers in producing high-quality solutions within a reasonable timeframe, we adopt a Dantzig-Wolfe decomposition to reformulate the proposed model. Furthermore, we develop a branch-and-price algorithm to achieve optimal solutions and introduce a problem-specific variable selection strategy for efficient branching. To assess the algorithm's effectiveness and examine the impact of the new break assignment constraint, we conducted a comprehensive computational study using real-world data from a German training hospital. Using our approach, healthcare institutions can streamline the rostering process, minimize the costs associated with overstaffing and overtime hours, and improve employee satisfaction by ensuring that physicians can take their legally mandated breaks. Ultimately, this contributes to better employee motivation and improves the overall level of hospital service."

Session 8: Applications of Reinforcement Learning

Stephan Meisel (UT): An Online Reinforcement Learning Approach for Energy Storage Management

We solve an energy storage management problem with intermittent energy production and with access to a short-term energy market. Both prices and production are exogenous processes that evolve randomly over time. We consider the case where models of these processes are not available, such that deriving a good policy by offline reinforcement learning is not an option. Instead, we propose a Bayesian online learning approach that uses concurrent simulation to solve the problem by online reinforcement learning with performance guarantees.

Rob Basten (TU/e) & Joost ter Haar (KU Leuven): Industrializing Deep Reinforcement Learning for ASML’s Service Network

We investigate the use of deep reinforcement learning (DRL) for operational spare parts planning. Operational spare parts planning can be characterized by large-scale networks with a need for quick decisions, making it difficult to identify good decisions. DRL models can theoretically be trained to take near-optimal decisions for such complex problems almost instantaneously. However, training DRL models for industrial-scale service networks remains an open challenge. We therefore propose a novel DRL approach that combines reward shaping, action space decomposition and global learning. We demonstrate the effectiveness of our approach on a stylized version of ASML's service network, and obtain insights into its performance. Our results show that our approach can be used to train a DRL model for a fully connected service network with 10,000 SKUs and 60 locations. We thereby extend the applicability of DRL to industrial-scale service networks, and show its usefulness for operational planning.

Stella Kapodistria (TU/e) - t.b.a.

Session 9: Empirical research in Supply Chain Management

Xichen Sun (Tilburg University) - The Effectiveness of Science-Based Carbon Targets

Although goal-setting theory predicts that a challenging and specific goal can improve performance, the evidence regarding the effectiveness of an environmental goal in organizations is mixed. Using a panel data set consisting over 700 global firms across multiple industries, we apply a Difference-in-Differences (DiD) strategy and empirically analyze the effect of having science-based carbon emissions targets (SBTs) on firms’ carbon performance. The findings in this study contribute to the debate regarding the necessity of setting SBTs.

Dick den Hertog (UvA & TU) - Analytics for a Better World

In this talk, I will describe two Analytics applications that contribute to one or more of the 17 Sustainable Development Goals (SDGs) of the United Nations. The first application is an optimization model to optimize the steering of the boats of The Ocean Cleanup to faster remove plastic from the oceans. The second application is an optimization model to optimize healthcare facility locations in Timor-Leste and stroke center locations in Vietnam. This project is carried out in collaboration with the World Bank. If time permits, he will shortly describe the current and future activities of the recently initiated Analytics for a Better World Institute.

Jan van Dalen (RSM) - Disrupted supply chains, violated labor conditions: empirical analysis of the associations between supply chain glitches and employment violations

Companies faced with supply chain disruptions have been found to report a weaker financial performance, to experience adverse social media responses, as well as to cope with operational deficiencies, like production halts and increased operations risks. Different stakeholders in these companies will respond differently to the organizational challenges caused by these disruptions, and management will have to make a dedicated effort to rebalance disrupted processes. Employment-related measures, like hiring freezes, promotion stops, and restructuring, are often-used instruments, possibly affecting the legitimate interests of personnel. The aim of the present research is to empirically explore the relation between supply chain glitches and the subsequent pattern of employment violations. Focus will be on S&P100 firms, using financial data from Compustat North America, news items about supply chain glitches from Wall Street Journal and the Dow Jones Newswire, employment violations information from Violations Tracker and Lex Machina, corporate ownership data from Thomson Reuters, and board composition data from BoardEx USA; observation period, 2000 – 2021. Apart from discussing motivation and presenting findings, the presentation will emphasize the use of matched samples, specifically the choice of preferred matching method, and the consequences of these choices for inference.