Multi-Agent Reinforcement Learning for Novel Radio Access Network Protocols

• Applicant: Professor Dr.-Ing. Peter Rost
• Subject Area: Communication Technology and Networks, High-Frequency Technology and Photonic Systems, Signal Processing and Machine Learning for Information Technology
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 556070992

This project develops a new category of radio access network protocols that are, on the one hand, specifically optimized for the use of machine learning and, on the other hand, are also “designed” with the help of machine learning. This type of protocols has the potential to partially replace conventional model-based methods in order to enable a more dynamic development of the radio access network architecture, especially with regard to the use of machine learning on the physical layer. The protocols thus enable more flexibility, adaptability and shorten the development cycles for new radio access network protocols, among others with regard to the necessary standardization. The focus of the project is the investigation of “Multi-Agent Reinforcement Learning” (MARL), where “agents” independently learn protocols in order to collaboratively complete tasks or jointly achieve goals. The project investigates how MARL can be applied in radio access networks, i.e., how the individual agents must be dimensioned, how necessary messages can be trained, and how the interaction between these "agents" and the radio access network must be organised. The protocols developed must be adaptable and robust against various, typical influences in mobile networks such as temporal and spatial fluctuations of the signal quality, the available data, and the quality of the data. This adaptability is crucial for the successful integration of novel methods utilizing machine learning. The project also investigates the integration into a mobile network architecture, e.g., how the data needed can be made available and how this data has to be preprocessed. It also investigates how the distribution and the continuous learning process of the individual “agents” must be organized to consistently meet the performance requirements. The algorithms are evaluated and compared with regard to their convergence behavior, scalability, robustness, and computational efficiency. To enable a fair comparison, two exemplary applications are used that are highly relevant for radio access networks. One is the mobility process and the other one is the allocation of radio resources. During the mobility process, a mobile terminal must be handed over from one cell to another based on the available channel state information and without causing radio link or handover failures. The aim of radio resource allocation is to assign the available radio resources to mobile terminals such that the quality of service criteria are met and the signaling overhead is minimized. In this project, the MARL-based methods will be compared to conventional protocols that are currently in use.

• DFG Programme: Research Grants