Machine Learning for Reconfigurable Intelligent Surface Assisted Communication (ML4RIS)

• Applicant: Dr.-Ing. Bile Peng
• 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 566937681

The reconfigurable intelligent surface (RIS) is a promising technology for next generation communication systems. It passively modifies the characteristics of wireless channels to meet requirements of advanced signal processing techniques (such as multiple access techniques). With a properly configured RIS, throughput and robustness of wireless communications can be significantly improved. In order to fully exploit the potential of RIS and make it a reality, we propose a RIS optimization approach with more than 1000 RIS elements, without assuming known channel state information (CSI) and considering user mobility. The proposed approach is validated and fine-tuned with a RIS prototype. Due to the inherent uncertainty of channel state and user movement, randomness is an unavoidable part in RIS configuration. The robustness of RIS-assisted communication is optimized by explicitly considering the aforementioned randomness. In order to perform the above optimizations, which are not amenable to analytical methods due to their high complexity, our project pioneers a novel approach called problem-specific unsupervised machine learning (ML). In contrast to supervised methods, our approach operates autonomously without given labels, empowering the neural network to search for optimal solutions autonomously, since the optimal RIS configuration is often unknown. Moreover, we depart from generic machine learning methodologies, combine expertise in communications technology and ML. Central to our methodology is the dedicated neural network architecture RISnet, which is designed according to the RIS property and can be seamlessly integrated with well-established analytical techniques such as multiple access precoding. Taking into account the inherent uncertainty described above, we propose a groundbreaking framework termed randomness-aware machine learning (RAML), which addresses the difficulty that typical neural network cannot handle randomness well because both input and output of the neural network are deterministic numbers, not distributions. RAML represents a paradigm shift in optimization strategies, aiming to enhance expected performance or outage probability under varying stochastic conditions, including complicated non-parametric conditional distributions, such as possible future channel conditions due to user movement given past and present channel conditions. The outcome of the project not only enables robust RIS configuration with high performance in real time, but also inspires optimization of other complex technical systems with a hybrid approach of ML and domain-knowledge based analytical methods.

• DFG Programme: Research Grants