The Summer School takes place from Monday August 23 at 9 AM, until Friday August 27 at 1 PM. PhD students will have the opportunity to present and discuss their ongoing research with their peers, as well as with senior researchers. A distinguished group of professors from the field of inventory management will participate during the School. These professors will provide keynote speeches and be available for in-depth discussions with the PhD students. Each student is expected to present a (working) paper and to act (together with the professors) as discussants of the other papers. Apart from research presentations, there will also be a social program, an industry panel that is focused on recent developments in inventory management in practice, and a publishing panel that will provide valuable insights into the academic publishing process. The detailed program can be found here. Abstracts of keynotes can be found below.

Human-machine interaction in supply chain decisions making

Monday August 23, 09:30-10:30

Prof. Kai Hoberg, Kuehne Logistics University

Decision support systems for supply chain planning have been supporting planners over decades to improve their decision making in many different ways.  As next-generation planning systems are leveraging advanced artificial intelligence (AI) technologies, companies must not only determine what decision support to use, but effectively shape how the supply chain planner (“the human”) and the system (“the machine”) work together. At the same time, AI-supported planning systems will change the job profiles and required skill sets of supply chain planners.

Forecasting and inventory management: tips & tricks

Monday August 23, 13:30-14:30

Prof. Nikolaos Kourentzes, University of Skövde

Forecasting is central to inventory management, yet the two research areas have often remained disconnected. Historically, forecasting has been primarily influenced by the needs and advances in econometrics. In this talk we revisit many of the common tropes in building forecasting models and evaluate them for the purpose of informing inventory management. We critically look at the need of a model in the statistical sense, the aspects and outputs of models that are important for inventory, and by extension what are the appropriate criteria of forecasting performance. We draw on recent innovations to show some underlying commonalities of well performing forecasts for inventory management, drawing on forecast combinations, temporal hierarchies, and shrinkage estimators.

The impact of sensor technology on spare parts inventory management

Tuesday August 24, 9:00-10:00

Prof. Gudrun Kiesmüller, Technical University Munich

Innovative sensor technology enables the measurement of the condition of a technical system or a component in such a system. However, not every failure can be detected by a sensor, and a signal can be caused by several problems. This means that imperfect advance demand information is available for the planning of the spare parts inventories. In this presentation we discuss the spare parts planning for a company providing field services where a service technician visits customers to repair their failed technical systems. For some repair jobs information in the form of failure codes is available, which can be used for the planning of the repair kit. We show how the problem can be modelled, which challenges occur, and we also discuss some solution approaches. With our model we can quantify the added value of the advance demand information and we can also show the added value of more reliable information.

Deep reinforcement learning for inventory control

Thursday August 26, 13:45-15:00

Prof. Willem van Jaarsveld, Eindhoven University of Technology

After being successfully used to train a computer to play Atari games in 2013, Deep Reinforcement Learning (DRL) has seen countless applications across a vast range of domains. We first give a brief tutorial on applying DRL in inventory control. We introduce the main concepts underlying DRL: deep neural networks and Markov Decision Processes. We then discuss how deep neural networks can be used to represent policies and/or value functions of MDPs. Subsequently, the training of neural networks is discussed. That is, we discuss how to determine weights for the neural network parameters such that the neural networks represents close-to-optimal policies. After the tutorial, we briefly reflect on future research directions, i.e. we discuss where the current approaches fall short, and how future research could overcome this.

Data drives us to battle the curse of dimensionality

Friday August 27, 9:00-10:00

Prof. Joachim Arts, University of Luxembourg

Additional data enables more accurate estimation of existing model. Very often, more data sources reveal structure that can be exploited when the model also includes more structural elements. Such model often have a higher dimension than models that are not informed by data. Thus data drives us to study higher dimensional models. Higher dimensional models suffer from the curse of dimensionality. This talk will outline how data drives us to battle the curse of dimensionality and sketch several weapons in the literature including parsimonious parametric models, approximation algorithms, asymptotic optimality, stochastic programming and robust optimization.