Mühlebach Michael 2880 X 880

Michael Mühlebach

Dr. Michael Mühlebach aims at exploiting analogies between optimization and dynamic systems for gaining new insights and constructing new optimization algorithms that inherently deal with constraints and whose complexity scales mildly in the number of decision variables.



Academic Career

  • Research group leader, Max Planck Institute for Intelligent Systems, Tübingen, Germany, 2021-present
  • Post-doc, University of California, Berkeley, USA, 2018-2021
  • Post-doc, ETH Zurich, Institute for Dynamic Systems and Control, Switzerland, 2018
  • PhD, ETH Zurich, Institute for Dynamic Systems and Control, Switzerland, 2014-2017
  • MSc, Robotics, Dynamic Systems, and Control, ETH Zurich, Switzerland, 2011-2014

Major Awards

  • Hilti prize for innovative research 2019
  • ETH Medal for outstanding doctoral thesis 2018
  • Willi-Studer Award 2014
  • Outstanding D-MAVT Bachelor Award 2011
  • Schweizer Jugend Forscht Award 2008

In the News
  • Prosieben Galileo (German television), robotics presentation, 2015


Branco Weiss Fellow Since

Research Category
Computer Science, Mathematics

Research Location
Max Planck Institute for Intelligent Systems, Tübingen, Germany

Optimization plays a central role in engineering, technology, and society, as it is the basis for rational decision making. It lies at the heart of many new technologies, for example (the list is by no means exhaustive): Computer aided diagnoses assist doctors in interpreting medical images. Due to the availability of large datasets (e.g. mammography images) machine learning algorithms are trained to discriminate malignant abnormalities from benign ones, and thus predict cancer with high certainty and reduce the rate of false-positives. Modern air traffic control algorithms optimize flight schedules for improving efficiency (in terms of reducing total flight delays, fuel consumption, delays per passenger, etc.) and minimizing the propagation of delays, while maintaining safety and security. Self-driving cars use an optimization-based approach to plan trajectories that are in accordance with traffic regulations, avoid obstacles, and guarantee safety, while maximizing passenger comfort at the same time. Speech recognition software recognize natural voice and interpret user commands, thereby providing a convenient user interface to mobile devices. Internet platforms such as Amazon, Spotify, Netflix, YouTube, Facebook analyze the activities of individuals and use optimization algorithms to provide personalized recommendations about products, movies, songs, news, etc. The underlying tool enabling all these applications is mathematical optimization. As a matter of fact, the advances in machine learning and automation that we have witnessed in the last decade are largely due to new optimization algorithms and a better understanding thereof. Due to increases in computational power, emerging low-cost sensing capabilities, and advances in data acquisition in general, possible fundamental research findings in optimization have a high impact and are likely to play a dominant role also in the future.
Details of Research
Dr. Michael Mühlebach’s research aims at exploiting analogies between dynamic systems and optimization. The underlying idea is the following: Consider the minimization of a two-dimensional function. Such a function can be visualized as a landscape of hills, and thus, in order to find a local minimum, one can simply place a marble in the landscape and let the marble roll downhill, accelerated by the gravitational pull. Provided that some friction is added, the marble will ultimately stop at a local minimum. Constraints are naturally implemented as barriers against which the marble can bounce off, resulting in non-smooth dynamics. As highlighted with the following example, such a perspective rooted in non-smooth dynamics, leads to a different treatment of constraints compared to standard optimization strategies. It might also contribute to a better understanding of optimization algorithms and has the potential to enable new algorithms that efficiently deal with large data sets.