Stochastic Bandits on the Rise: Lever-aging Rising Bandits for AutoML.

This talk explores the use of multi-armed bandit (MAB) algorithms for efficient neural architecture search (NAS) and hyperparameter optimization in machine learning. Selecting suitable architectures and hyperparameters is a computationally expensive problem, typically requiring the evaluation of many candidate configurations. We investigate how bandit-based decision strategies can guide this search more efficiently by balancing exploration of new configurations with exploitation of promising ones. Different bandit algorithms are compared in terms of their ability to allocate computational resources adaptively during the search process. The framework treats candidate architectures as arms, whose performance is estimated through partial training signals. Initial experiments show promising results, indicating that bandit algorithms allocate resources effectively when the search setting is highly constrained.

Till Aust studied Robotics and Autonomous Systems at the University of Lübeck, where he earned both his Bachelor's and Master's degrees. He then joined the Cyber-Physical Systems Group led by Prof. Heiko Hamann as a PhD student, focusing on the analysis of physiological signals in bio-hybrid systems. His research contributes to the EU-funded projects WatchPlant and ChronoPilot, where he explores methods to interpret physiological data for controlling bio-hybrid systems and enabling feedback mechanisms.