Agenda

Overview:

• 15-01-21: Staff scheduling
• 22-01-21: Business Analytics
• 29-01-21: Public transport
• 12-02-21: Collaborative Logistics
• 19-02-21: Data-driven decisions in OR
• 19-03-21: ORBEL Award

Staff scheduling - Pieter Smet and Hatice Çalik

• Pieter Smet (KU Leuven)
  Title:
  Determining buffer sizes in staff rostering problems
  
  Abstract:
  Employee absences are inevitable in practice and often significantly disrupt employee shift rosters. At an operational level, these unforeseen events are addressed using repair methods which, despite their widespread use, often require last-minute changes to rosters, thereby negatively affecting employees' personal lives. Robust rosters are constructed in such a way that they are less sensitive to unforeseen absences. This work proposes a methodology to determine the size and position of different types of buffers in employee shift rosters. Integer programming is used to solve an optimization problem which finds the optimal trade-off between costs associated with the different buffer types and costs incurred by the repair methods. A computational study demonstrates the impact of various problem characteristics on this trade-off. While the robust rosters are less affected by disruptions, the computational effort required to find these solutions increases considerably.
  
  
• Mariana da Cunha (University of Lisbon)
  Title:
  Multi-objective workforce scheduling at the Portuguese emergency medical services
  
  Abstract:
  This work targets the multi-skill workforce schedule at emergency medical services (EMS). EMS provide specialized and critical medical aid on a 24/7 basis. Hence, the efficiency of its operations and the satisfaction of its personnel are considered of high importance. EMS often have scarce and difficult to retain personnel, together with a very high cost of understaffing, since it directly impacts the level of service provided. Furthermore, additional challenges arise from the heterogeneity of skills and dispersed locations. As a consequence, a good quality workforce schedule can be very beneficial for EMS. To solve this problem, a multi-objective model is proposed considering three objectives: 1) demand satisfaction, where overstaffing and understaffing are considered; 2) schedule quality, regarding shift patterns and team changes; and 3) social fairness, with respect to overtime, undertime and weekends-off. The proposed solution approach combines local search with a multi-objective MILP strategy and is applied to instances from the Portuguese EMS provider.
  
  
• Carlo Sartori (KU Leuven)
  Title:
  A constraint satisfaction algorithm for the truck driver scheduling problem with interdependent routes
  
  Abstract:
  The Truck Driver Scheduling Problem (TDSP) is a well-known problem found in vehicle routing applications for ensuring compliance with hours of service regulations of truck drivers. In this work, we tackle a variant denominated TDSP with Interdependent Routes (TDSP-IR), in which routes of multiple truck drivers require temporal synchronization. Due to this requirement, schedules must be produced simultaneously for all truck routes so that they respect the interdependence constraints, the truck drivers' regulations, and the time windows of all tasks. To solve the TDSP-IR, we propose an algorithm based on constraint satisfaction techniques and show that it can efficiently produce schedules for multiple, interdependent truck routes while minimizing their makespan. Computational experiments demonstrate the superiority of the proposed algorithm compared to a mixed-integer programming formulation of the problem.
  
  
• Hatice Çalik (KU Leuven)
  Title:
  Staffing with training requirements
  
  Abstract:
  The focus of this work is on introducing a novel staffing problem where personnel must be trained to operate certain groups of machines, whereas others can also be operated by interim workers hired at an additional cost. In order to achieve a feasible or a minimum-cost assignment for long-term planning, it may be necessary to cross-train employees on different machines. However, this requires switching to different machines frequently, which is not desirable for personnel. Therefore, each switch incurs a penalty cost. This specific characteristic makes the problem unique and complex when compared to well-known related problems in the literature. The problem requires assigning each machine to a minimum number of operators for each day of a given planning horizon while minimizing the total cost of hiring interim workers and switching machines. We provide integer programming formulations of the problem together with several valid inequalities. We further develop an iterated local search method to solve instances with longer planning horizons. A comparison of the introduced methods on randomly generated instances indicates that the problem is very challenging to solve with mathematical models whereas the iterated local search algorithm is capable of finding high quality solutions within a reasonable time limit.
  
  

Business Analytics - Stiene Praet

(22 January 2021)

• Tom Vermeiren (UAntwerpen - David Martens)
  Title:
  Explainable Image Classification with Evidence Counterfactual
  
  Abstract:
  Due to their complexity, state-of-the-art image classification models are used in a black-box way without the ability to explain the predictions to humans. As image classification is increasingly used for critical decisions, this lack of explainability becomes a major problem. Counterfactual explanations are put forward by legal scholars and data scientists as a promising avenue of research in the field of explainable artificial intelligence. A counterfactual explanation for image classification points at the parts of an image that, when removed, change the classification. In this talk, SEDC and SEDC-T are introduced as model-agnostic explanation methods to generate such counterfactuals. After a brief overview of the existing literature, concrete examples and large-scale experiments are discussed to show the ability of our approach to derive insights (i) to increase trust in model decisions and (ii) to get input for model improvement. In addition, our approach is benchmarked against existing model-agnostic explanation approaches.
  
  
• Felix Vandervorst ( VUB - Wouter Verbeke, Tim Verdonck)
  Title:
  A non-parametric approach to underwriting data misrepresentation using conditional density estimation
  
  Abstract:
  Premium fraud is the risk of adverse data misrepresentation committed with the intent to benefit from undue lower premium at underwriting of an insurance contract. In this presentation, we show how recent methods in non-parametric conditional density estimations can be used jointly with a pricing model to detect premium fraud at underwriting of an insurance contract, based on a set of validated contracts.
  
  
• Bram Janssens ( UGent - Matthias Bogaert, Dirk Van den Poel)
  Title:
  Evaluating the Influence of AirBnb listings’ descriptions on demand
  
  Abstract:
  Hosts list their accommodations on Airbnb aspiring to attract guests. Extant research on the drivers of guests’ booking behaviour has solely considered structured information on the Airbnb platform, thereby omitting the rich information provided in the unstructured textual listing description. This work adds to the stream of research on Airbnb demand determinants by identifying the latent topics used in these unstructured descriptions as drivers of listing demand. Both our empirical model and follow-up experimental study indicate that Airbnb guests value unique accommodation aspects of which hosts can convince their potential guests by using the textual description. Guests especially value enthusiastic home experiences and a unique local city guide accompanying the listing. However, when hotel-like properties are conveyed in the description, prospective guests are dissuaded.
  
  
• Toon Vanderschueren (KULeuven - Bart Baesens, Wouter Verbeke, Tim Verdonck)
  Title:
  An empirical evaluation of instance-dependent cost-sensitive learning
  
  Abstract:
  Traditionally, machine learning algorithms aim to minimize the number of errors. However, this leads to suboptimal results for many business applications where the actual goal is to minimize the cost, not the number of errors. In customer churn for example, the algorithm’s predictions should be especially reliable when dealing with highly valuable customers, even if this means wrongly predicting the churn for customers with insignificant value. Cost-sensitive learning aims to address this issue by incorporating costs in the learning algorithm. Recently, a number of cost-sensitive classifiers have been suggested that deal with cases where costs are instance-dependent. This work presents an empirical study comparing several of these instance-dependent cost-sensitive methods. Furthermore, the effects of incorporating costs at the instance-level are examined, as well as the influence of thresholding and regularization.
  
  

Public Transportation - Lissa Melis

• Soukaina Bayri (VUB - Yves Molenburgh)
  Title:
  Routing and scheduling algorithms for integrated mobility systems with dynamic and stochastic characteristics
  
  Abstract:
  Integrated mobility systems are gaining popularity in many Western countries. These integrated systems allow passengers to travel from rural to urban areas and vice versa by means of a combination of (1) demand-responsive transportation services with flexible routes and schedules and (2) timetabled public transport. A journey starts with a customer submitting a request to a service provider. This provider plans an integrated route that may consist of a combination of public transport and demand-responsive services, assuring both systems to be aligned with each other. The service provider minimizes the cost of the demand-responsive services, while guaranteeing a reliable planning and optimal transfers to/from public transport. The main goal of this research is to develop an efficient routing and scheduling algorithm that considers both dynamic and stochastic characteristics of the problem. In this talk, a research plan, a literature overview and a problem formulation will be presented.
  
  
• Dilay Aktas (KULeuven - Pieter Vansteenwegen, Kenneth Sörensen)
  Title: A demand responsive public bus system with short-cut return trips
  Abstract:
  Between a conventional public bus system and a complete on-demand system, a range of demand responsive options exists for which real-time information on the actual demand for transportation is available. In this study, we focus on the morning peak hours where the passenger flows towards a city center are much larger than the flows in the opposite direction. In a conventional system, a fixed number of vehicles drives back and forth serving all stops between a terminal station and the city center, based on a predetermined timetable. We introduce a system where short-cut return trips are allowed. Based on the expected demand, it is decided for a single line, whether a vehicle should visit all the stops ahead or skip some of them to take a shorter way in the return trip so that it can start serving the passengers towards the city center again, earlier. When optimizing this system, it is taken into account that some recovery period might be required, before the system can return to its conventional operation. We present a Mixed Integer Quadratic Program to mathematically model this problem. Due to its complexity, only small-sized problems can be solved optimally. Therefore, we also design a metaheuristic algorithm based on Large Neighborhood Search that finds high-quality solutions within reasonable time for realistic instances. We analyze the effects of the duration of the peak hour, fleet size, and different demand scenarios. The results show that the demand responsive system improves the conventional system up to 10%.
  
  
• Fabio Saitori Vieira (KULeuven - Pieter Vansteenwegen, Kenneth Sòrensen)
  Title:
  Dynamic feeder lines in suburban areas
  
  Abstract:
  This project aims to increase the ƒflexibility of bus services by allowing route deviations to feeder bus lines of suburban areas. Based on a system with a predefined route and timetable, the dynamic service achieves an optimal operation each period. ‘The objective is to reduce users’ travel time. The comparison of daily operations simulations of several instances allows a performance measurement. In low demand periods, it is possible to achieve faster routes to the destination with the optimized service.
  
  
• Bryan Galarza Montenegro (UAntwerpen - Kenneth Sörensen, Pieter Vansteenwegen)
  Title:
  A demand-responsive feeder service for smart cities
  
  Abstract:
  Feeder services are public transit services that transport people from a low demand, typically suburban, area to a high demand area, such as a transportation hub or a city. Here, passengers continue their journey using traditional forms of public transport. These feeder services are essential in geographically secluded communities, like suburbs and mountainous villages, since they connect these communities with the rest of the world. On one hand, on-demand-only feeder services have been a topic of discussion in a number of recent studies, since these services can serve the demand efficiently. On the other hand, traditional feeder services provide predictability and costs are easier to control. In this talk, a demand-responsive feeder service is considered, which has positive characteristics of both traditional services as well as on-demand-only services. To optimize the performance of the feeder service, a large neighborhood search heuristic is developed. Experimental results on 14 benchmark instances illustrate that the heuristic obtains solutions with an average gap of only 1 % compared to the optimal solution within 1 s of run-time. The results also show that under certain circumstances the demand responsive-feeder services outperforms the traditional services by 33 % when the same weighted average for service quality is taken as the objective.

Collaborative logistics - Christof Defryn

• Thomas Hacardiaux (UCLouvain)
  Joint work with Lotte Verdonck, Jean-Sébastien Tancrez and Christof Defryn
  Title:
  Partner selection accounting for product characteristics in horizontal collaboration
  
  Abstract:
  Horizontal collaboration is a promising avenue to improve the efficiency of logistical operations, reducing costs and improving customer service. However, the success of achieving collaborative benefits strongly depends on the degree of fit between the collaborating partners. We analyze the impact of the partners’ product characteristics on the benefits of the collaboration. Depending on the type of product they distribute, companies might have different requirements and expectations on supply chain efficiency and responsiveness. To assess the benefits of collaboration, we use a location-inventory model that accounts for the partners’ individual interests, as well as the costs revealing the efficiency-responsiveness spectrum (i.e. opening of distribution centers, transportation, cycle inventory, as well as safety stocks and stock-outs). The model offers a set of Pareto-optimal solutions that support the decision and negotiation process. Finally, through numerical experiments for companies with functional and innovative products, we reveal valuable managerial insights on the effect of dissimilarities in demand volumes for products with similar or different levels of innovativeness and we support the partner selection decision in the context of horizontal collaboration.
  
  
• Nathalie Vanvuchelen (KULeuven)
  Joint work with Joren Gijsbrechts and Robert Boute
  Title:
  Machine learning for collaborative logistics
  
  Abstract:
  Deep reinforcement learning has been coined as a promising research avenue to solve sequential decision making problems, especially if few is known about the optimal policy structure. We apply the proximal policy optimization algorithm to the intractable joint replenishment problem. We demonstrate how the algorithm approaches the optimal policy structure and outperforms two other heuristics. Its deployment in supply chain control towers can orchestrate and facilitate collaborative shipping in the Physical Internet.
  Keywords: Collaborative Shipping, Physical Internet, Joint Replenishment Problem, Machine Learning, Deep Reinforcement Learning, Proximal Policy Optimization
  
  
• Aymen Aloui (Université de Picardie, France)
  Joint work with Nadia Hamani, Ridha Derrouiche and Laurent Delahoche
  Title:
  Horizontal collaboration and sustainability in freight distribution networks: A Literature Review
  
  Abstract
  In the last few years, competitiveness, problems of globalization and concerns about sustainability require new approaches and models for the planning of transport networks. Horizontal logistics cooperation has been considered an emerging and innovative approach in the design and management of sustainable supply chains. This approach is based on the sharing of resources between actors at the same level in different supply chains. This article provides a Systematic Literature Review (SLR) about sustainability and collaboration in the freight transport sector. It aims to analyse the existing literature in order to reveal the studies already conducted and to identify gaps and opportunities for future research. A total of 89 articles have been published between 2010 and 2020 which have been examined. The results show that the integration of these three dimensions of sustainable development in the field of collaborative network optimization, especially the social considerations have been little studied. In addition, the analysis shows that most of the authors have focused their research on transport optimization at the operational level, with few works on the problem of designing and managing the integrated supply chain.
  
  
• Lotte Verdonck (UHasselt)
  Joint work with Florian Diehlmann, Markus Lüttenberg, Marcus Wiens, Alexander Zienau and Frank Schultmann
  Title:
  Collaboration in Emergency Logistics: A Framework based on Logistical and Game-Theoretical Concepts
  
  Abstract:
  Collaboration in emergency logistics can be beneficial for governmental actors when supply chains need to be set up immediately. In comparison to research on humanitarian-business partnerships, the body of literature on so-called Public-Private Emergency Collaborations (PPEC) remains scarce. Private companies are only rarely considered within research on emergency collaborations, although they could contribute to a more efficient supply of goods given their resources and existing communication networks. Based on this research gap, we develop a logistical and game-theoretical modeling framework for public-private emergency collaborations. We characterize both public and private actors' possible roles in emergency logistics based on literature research and real cases. Furthermore, we provide an overview on existing PPECs and the challenges they are confronted with. To address the challenge of evaluating different objectives in a collaboration, we add a game-theoretical approach to highlight the incentive structure of both parties in such a collaboration.

Data-driven decisions in OR - Victor Bucarey Lopez (VUB) and Hatice Çalik (KULeuven)

• Keynote speaker: Dolores Romero
  Title:
  Making Data Driven Decisions with Tree Models
  
  Abstract:
  Classification and regression trees, as well as their variants, are off-the-shelf methods in Machine Learning. In this paper, we review recent contributions within the Continuous Optimization and the Mixed Integer Linear Optimization paradigms to develop novel formulations in this research area. We compare those in terms of the nature of the decision variables and the constraints required, as well as the optimization algorithms proposed. We illustrate how these powerful formulations enhance the flexibility of tree models, being better suited to incorporate desirable properties such as cost-sensitivity, explainability and fairness, and to deal with complex data, such as functional data.
  
  PhD Talks:
  
  
• Jorik Jooken, Department of Computer Science, KU Leuven, Kulak, Kortrijk, Belgium
  (joint work with Pieter Leyman, Patrick De Causmaecker and Tony Wauters)
  Title:
  Monte Carlo tree search for combinatorial optimization
  
  Abstract:
  Monte Carlo tree search is a popular algorithm in the field of game playing that learns to play good moves by running Monte Carlo simulations. In this presentation we will present our recent work which shows that an adaptation of Monte Carlo tree search can also be successfully used for a very different purpose, namely to solve combinatorial optimization problems. In such problems, the search space tends to be much larger than in a typical game. We propose various enhancements that rely on exploiting the combinatorial structure of the problem to tackle this challenge. The developed algorithm was evaluated on two different combinatorial optimization problems: the 0-1 knapsack problem and the quay crane scheduling problem with non-crossing constraints. The computational results reveal that the algorithm is able to compete with the state-of-the-art methods for both problems.
  
  
• Ziyi Chena, Department of Computer Science, KU Leuven, Kulak, Kortrijk, Belgium
  (joint work with Patrick De Causmaecker)
  Title:
  Neural Network Assisted Branch and Bound Method for the Nurse Rostering Problem
  
  Abstract:
  The Nurse rostering problem (NRP) needs to fully consider the preferences of nurses, legal constraints and other constraints, to formulate a systematic and scientific nurse scheduling plan. The problem aims to optimize the allocation of human resources, effectively reduce the nurses’ work pressure and improve their work efficiency and quality. Due to the need to consider various constraints during the scheduling process, the NRP is complicated and is known to be NP-hard. How to quickly get an efficient schedule has become an urgently needed and challenging problem. In this paper, we propose a new method combining Deep Neural Networks (DNN) and Branch & Bounding (DNNB&B) to make the complex and costly design of heuristics for the problem automatic. It can select branches by analyzing existing (near-)optimal solutions. By treating schedules as matrices, the neural networks can help to decide which branch to explore next and which branch to prune.
  
  
• Maxime Mulamba, Data Analytics Laboratory, Free University of Brussels-VUB, Belgium
  (joint work with Jayanta Mandi, Victor Bucarey, Michelangelo Diligenti and Michele Lombardi)
  Title:
  End-to-end decision-focussed learning over combinatorial problems
  
  Abstract:
  Many decision-making processes involve solving a combinatorial optimization problem with uncertain input that can be estimated from historic data. In this talk we show the main algorithmic challenges involved in the process of learning a decision-task oriented cost vector for combinatorial optimization. In particular we show its relationship with bilevel optimization and what are the different techniques proposed in the literature to find these vectors via an end-to-end learning procedure.
  
  
• Silia Mertens, Universiteit Hasselt
  (joint work with Ahmed K.A. Abdullah, Prof.dr. Yves Molenbruch, Prof.dr. Kris Braekers, Prof.dr. An Caris, Prof.dr. Tias Guns)
  Title:
  Improving vehicle routing by learning demand uncertainties from historic data: a case study
  
  Abstract:
  Vehicle routing problems (VRPs) have already been widely studied in literature. However, the inclusion of uncertainty in VRPs is still less studied. Stochastic problems assume that only a probability distribution is known on the uncertain parameters. This study focuses on the VRP with time windows and stochastic demands, which is based on a real-life problem faced by a logistics service provider. One of the problems the company is facing is that there are often deviations between the specified quantities by the customers and the actual quantities when arriving at the customer, which makes it difficult to make a reliable planning. Stochastic demands can result in inefficient use of vehicle capacity or capacity shortages, leading to costly corrective actions when executing planned collection routes. This study quantifies the importance of this problem. A two-stage stochastic programming with recourse method is applied to model the problem, where the aim is to minimize total expected costs. A detour to depot recourse policy is used to deal with route failures. When a failure occurs, the vehicle returns to the depot to (un)load. Afterwards, it resumes its route as planned, restarting at the customer where the route failure occurred. This study extends the state of the art by including the effect of expected violations of time windows, maximum route duration and other time-related constraints into the recourse cost function. An iterated local search algorithm is presented to solve the problem. Moreover, different learning techniques are used and compared to estimate the probabilities of failure at every customer in a route from real-life historic data. Experimental results demonstrate that in general, total costs can be reduced when taking the expected cost of corrective actions into account when planning the routes.
  
  

ORBEL Award