Deep reinforcement learning (RL) algorithms are a powerful class of methods for optimizing sequential decision making and control problems, and are the driver behind many real-world applications. We hypothesize that deep RL applications can considerably benefit from a focus on hyperparameter tuning. Specifically, we aim to develop automated techniques for setting the hyperparameters in a sample-efficient manner for small datasets, thus reducing uncertainty. We will create appropriate benchmarks and then investigate the performance of currently existing hyperparameter optimization (HPO) methods, to then develop automated approaches. This will include a dynamic meta-level control with the use of a robust hyperparameter transfer for online HPO for offline RL and neural architecture search (NAS).
(supervised by PI Awad)
Firstly, the creation of cost-effective benchmarks using small datasets. Secondly, the investigation and development of automated techniques for determining hyperparameters in deep reinforcement learning algorithms, with a specific emphasis on enhancing performance in small dataset scenarios. Additionally, the development of new, efficient HPO and NAS methods that will enhance accuracy and decrease uncertainty for offline deep RL applications.
(supervised by PI Bödecker)
Analysis of importance and the dynamics of hyperparameters of different RL algorithms, robust, hyperparameter transfer methods, and an evaluation in applications from the biomedical domain.