PhD in Efficient Online Optimization

How do you optimize a system you can’t fully model, where the operational objectives vary over time, and where safety is a major concern? Drones flying in cluttered changing environment, buildings that manage their own climate, and supply chains under changing market conditions and global events all face this problem. We are looking for a curious, mathematically minded PhD candidate to help build the theory that lets such systems learn on the fly, safely and efficiently and without a central coordinator. If you like elegant algorithms, mathematical challenges and seeing your theory tested on real drones and a real building, we’d like to hear from you.

Information
Bayesian optimization is a data-efficient way to optimize controllers and systems without explicit models, because it learns surrogate models and comes with built-in uncertainty estimates. Its theory still has major gaps when the system is large-scale, distributed and safety-critical, and when the reward and safety constraints change over time. In this project you will develop a new framework that combines safe Bayesian optimization, online learning and distributed estimation, with a strong focus on provable guarantees. This is primarily theoretical research. Your main contributions will be new algorithms and convergence and safety guarantees. Experimental validation will be done together with our partners in the drone lab of the Control Systems group at TU Eindhoven and at the NEST building in Dübendorf, Switzerland.

You will be embedded in the Control Systems group at TU/e, contributing to its research program on intelligent and responsible control of networked systems.

As a PhD candidate, you will:

• Develop safe Bayesian optimization algorithms for time-varying settings, including adaptive sampling strategies that track moving optima while keeping constraints satisfied at every step.
• Design distributed learning and optimization schemes in which multiple agents, each with only partial information, cooperate without central coordination.
• Study how surrogate models and the optimization process can be broken down in time and space to stay tractable in high dimensions and over long horizons.
• Derive rigorous convergence, regret and safety guarantees for constrained large-scale optimization in non-stationary, safety-critical environments.
• Use tools from online learning, Gaussian processes, distributed estimation and control theory, and connect them in new ways.
• Turn your theory into algorithms that can be validated on UAV reference tracking and obstacle avoidance, and on temperature control in a real residential building in Dübendorf, Switzerland.
• Publish in leading venues in control, optimization and machine learning (e.g., CDC, TAC, NeurIPS, ICML), present your work internationally and work with our academic partners.

This PhD position is in collaboration with Dr. Marta Zagorowska from TU Delft and other collaboration partners in Switzerland.