Project Details
Projekt Print View

Algorithm Control: Efficient Learning to Control Algorithm Parameters

Subject Area Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
Term from 2020 to 2024
Project identifier Deutsche Forschungsgemeinschaft (DFG) - Project number 442750095
 
Final Report Year 2025

Final Report Abstract

It is widely known that the performance of many algorithms depends on their hyperparameters. To avoid tedious and error-prone manual tuning, automated approaches have been developed to efficiently achieve peak algorithm performance. However, these approaches have the drawback of setting hyperparameter values only once, even though many hand-designed internal heuristics of AI algorithms are typically reactive in nature, adjusting parameters dynamically during runtime. In this project, we developed approaches to automatically learn such reactive, dynamic heuristics for hyperparameter control from data and thus further enhance the efficiency of algorithms. To achieve this, we leveraged the latest advances in reinforcement learning by framing the task of algorithm control as a contextual Markov Decision Process (MDP). This allowed us to make algorithms across several domains significantly more efficient, such as in automated planning, evolutionary algorithms, and deep learning. To enable the research community to continue working dynamically in these domains, we also created a freely available benchmark library. Additionally, we developed methods to adapt reinforcement learning for algorithm control by explicitly modeling the temporal component and instance context. This project, therefore, provided the theoretical, practical and empirical foundations and the first benchmark for dynamic algorithm configuration, demonstrated the potential of reinforcement learning for this task for the first time, extended reinforcement learning through the lens of algorithm control, and thus significantly improved algorithms across various domains.

Publications

 
 

Additional Information

Textvergrößerung und Kontrastanpassung