Ongoing Research Projects
Adaptive data-driven inventory control for dynamic supply chains
Traditional inventory models are based on unrealistic assumptions of isolated control and long-run stability. Modern supply chains are highly dynamic, requiring inventory policies that quickly adapt using all available data. In line with EU sustainable growth aims, this prevents obsolescence and waste. Furthermore, companies can reduce costs by stocking what is needed in the future rather than what was demanded in the past.
We develop data-driven integrated forecasting and inventory models for the situation after a structural supply chain change, where classical inventory theory fails. We devise a general framework to select explanatory variables and dynamically learn from incoming information. We study the performance-interpretability trade-off of machine learning methods, also deriving a new approach based on Extreme Learning Machines. We focus on a just-introduced new product and a just-contracted new supplier, inspired by cases of our business contacts.
We theoretically establish convergence and finite-sample performance, but furthermore apply the policies to real data of European businesses. The result is the first data-driven solution to inventory policy learning in dynamic supply chains and a comprehensive overview of transparency and performance of different machine learning approaches.
Integrated learning and optimization for mobility and transportation
For this TUM IGSSE funded project, which is a collaboration between TU Munich in Germany and DTU in Denmark, I am the project leader. The introduction of new business models in transportation is exposed to a large degree of uncertainty at the stage of design and operations due to limited available historical information. Therefore, integral parameters for product and process design need to be learned. Typically, forecasting and optimization are performed sequentially.
In this interdisciplinary project, we follow an integrated approach and build on concepts from data-driven optimization, stochastic programming and machine learning to develop decision support with the application to transportation and mobility, in particular for bike- and car-sharing. Based on bike- and car-sharing data from the cities of Copenhagen and Munich, exploration and exploitation methods and active learning concepts will be developed to support strategic and operational decision making with regard to capacity, inventory positioning and resource rebalancing under uncertainty in large systems with distributed (competitive) decision making. This integrated approach requires knowledge from the disciplines of transportation and management as well as from stochastic optimization and machine learning.
Other ongoing research
- Compound Poisson parameter estimation for inventory control
- Road logistics price forecasting using machine learning
- Lead time prediction using historical data and planner judgments