Research Interests

  • CP-SAT solver and reinforcement learner for planning and scheduling
  • Business & industrial process analysis and optimization
  • Modeling of digital twins in Industry 4.0
  • Complex and scalable AI solution architectures
  • AI safety and enterprise architecture


Current Projects

Process Optimization with CP-SAT Solvers

Recent progress in CP-SAT solvers makes it possible to apply these AI algorithms to industrial problems of impressive complexity. We investigate methods to model these problems such that CP-SAT solvers scale well, study and research automatic model rewriting and compilation methods. Furthermore, we are interested in better understanding the root causes of model unsatisfiability for easier modeling and spotting modeling errors. Our work also results in creating benchmark sets for competitions and in empirical comparisons of contemporary AI solvers, e.g. see the MiniZinc Challenge 2020 to which we contributed the CTW domain.
Cooperation with IWI DFKI

Reinforcement Learning in Real-world Domains with Many Dead-End States

Real-world scenarios differ significantly from gaming domains where reinforcement learning has been applied with impressive success. In many industrial applications for example, actions often lead to dead-end states from where the goal is no longer reachable. Simply providing a negative reward does not allow a RL learner to understand why a sequence of actions leads to an unrecoverable failure. We are working towards methods that improve the “informedness” of the reward and that will allow RL learners to build a model of failure to avoid dead-end states in learned policies

Constraint Optimization in the Life Sciences

The life sciences are not only the source of many challenging problems that can be addressed using machine learning, but also require for example to find optimal combinations of measurement results or optimized compositions of cells and other components. We are investigating if and how these optimization problems can be modeled using AI-based methods and if AI algorithms such as CP-SAT solver can scale to compute optimal or near-optimal solutions.
Cooperation with Prof. Jörn Walter

Architectural Thinking for AI Systems

Going from an AI prototype to a market-ready AI solution is a challenging endeavor. Many projects fail on this way, because the AI system context is poorly understood or the integration effort underestimated. AI technology requires to rethink functional requirements and system qualities due, for example, due to the inherent stochastic nature of modern AI Algorithms. In addition, operational risks have to be explicitly addressed and managed. Our framework for architectural thinking and operational risk management addresses these challenges.

Management of AI Projects

Successfully managing projects that use AI technology is a challenging task due to the multi-disciplinary teams that must come together, the complexity of the technologies, or the difficulty to maintain the data pipeline. Many projects fail, but not for technological reasons. How to address the needs of feasibility, viability, and desirability in an AI project is often not well understood by practitioners. Furthermore, important tasks that need to be prepared during the various project phases are sometimes overlooked. We work on a concise and proven management method that helps to ask the right questions at the right time in the project and to successfully manage upcoming challenges. The method brings together elements from AI technology, design thinking, agile software development, and psychological models of change and transformation.
Cooperation with University of St. Gallen, Institute of Information Management.

Postgraduate AI Education for Professionals

I regularly teach AI courses for professionals that provide a general introduction into the field of Artificial Intelligence, its history, key assumptions, paradigms, concepts, and fundamental methods. In these courses we also discuss current trends, challenges, and risks related to AI technologies and the field in general. I also share lessons learned from AI application projects where I discuss best practices and derailment factors based on real case studies.
Some of the material is currently prepared for the AI Campus I teach these courses directly for companies or via partner offers, e.g. CAS Digital Business Innovation and CAS Artificial Intelligence.