Optimization is one fundamental building block of current machine learning algorithms and many other methods in the natural sciences. In contrast to local optimization and convex problems, where principled methods can be derived, global optimization for non-convex problems often times reverts back to heuristic and evolutionary search procedures.

Meta learning methods enable the automatic discovery of high-quality optimization schemes tailored to a class of problems. Such approaches have shown to outperform heuristic approaches in the respective fields: e.g. stochastic gradient based local optimization [1] and data efficient global optimization [2].

This work strives to explore learning global optimizers in a big data regime, where both problem dimension and number of objective evaluations might be large and unknown a-priori. One main challenge therefore is dealing with large and dynamically sized datasets e.g. using set-based methods [3]. Another important issue is the learning procedure itself. Here, a wide range of supervised as well as reinforcement learning methods are conceivable.

A scientific publication of the results is possible and desirable. If desired, there is also the option for a research internship prior to the master thesis.

The project is offered by the Bosch Center for Artificial Intelligence, which we collaborate with. If desired, the project can be in collaboration and exchange with the Intelligent Control Systems Group at MPI Stuttgart.