Seminar AI in Logistics
11:00-12:30 in Ravelijn 1501, University of Twente, Enschede, The Netherlands
Niels Agatz
Real-Time Routing Cost Predictions for Time Slot Management
In the context of online grocery delivery, last-mile operations are challenging and costly. It is common in e-grocery retailing to allow customers to select a delivery window from a menu of available options. The retailer may decide to incentivize those time slots that allow for a more efficient route plan. To make these decisions, the retailer must determine the cost of serving customers in each time slot. Determining this marginal detour cost is a computationally challenging problem because it involves solving a vehicle routing problem with time windows. In practice, the number of customer orders can be large and the available time extremely limited (<1 second). Therefore, it is common to use fast routing heuristics to estimate the "insertion cost". Here, we explore the use of machine learning models to better predict marginal routing costs.
Robert Boute
Predict-and-Optimize Machine Learning for Feature-Based Inventory and Maintenance Optimization
Prediction and optimization have traditionally been considered two related yet separable tasks in operations management literature and practice. For example, before making inventory decisions, demand is typically forecasted independently. Analogously, before making maintenance decisions, the remaining useful lifetime of a machine is forecasted. A growing body of research suggests that it may not always be a good idea. We explore using supervised learning with custom loss functions to integrate prediction and optimization. Such a predict-and-optimize approach ‘predicts the optimal decision.’ We show its application in multi-period inventory and maintenance problems.
Willem van Jaarsveld
Inventory Planning in Capacitated High-Tech Assembly Systems Under Non-Stationary Demand
We study inventory planning in high-tech manufacturing supply chains driven by the semiconductor market, focusing on capacitated assembly systems. The research addresses demand variability through a Markov decision process and a new demand model that captures non-stationary uncertainties. Comparing traditional base-stock policies with deep reinforcement learning (DRL), we find that custom base-stock policies perform well in stable environments but falter under non-stationary demand. In contrast, DRL consistently achieves optimality gaps below 1%, proving superior particularly in dynamic settings, as shown in a case study with industry partner ASML. This demonstrates DRL's effectiveness in managing capacitated systems by strategically allocating downstream inventory to enhance service levels.
PhD defence Fabian Akkerman
14:30 - 16:30 in Waaier 4, University of Twente, Enschede, The Netherlands
Logistics decision-making is becoming more complex due to uncertainty, real-time demands, and disruptions. While traditional operations research (OR) methods provide structured solutions, machine learning (ML) offers complementary, data-driven tools that can adapt to dynamic and high-dimensional environments. This dissertation explores how ML can support decision-making across supply logistics, distribution logistics, and revenue management in distribution logistics.
In supply logistics, ML is used to improve inventory management under uncertainty, addressing demand variability and inventory record inaccuracies. In distribution logistics, ML contributes to dynamic routing and customer selection through predictive models and reinforcement learning. In revenue management, ML helps adjust pricing and delivery offerings (e.g., parcel lockers, time slots) in response to customer demand.
The research distinguishes between data analytics (using historical data to generate insights) and decision analytics (directly optimizing decisions), and proposes a framework for integrating ML into logistics. The goal is to support more adaptive, efficient, and resilient logistics operations.
