Channel charting (CC) is an emerging application of self-supervised machine learning (ML) to wireless communication which leverages the fact that wireless communications systems continuously collect data about the electromagnetic propagation channel. This data, known as channel state information (CSI), is high-dimensional and acquired at fast rates but typically discarded immediately after use. In contrast, CC recycles acquired CSI data by means of dimensionality reduction (DR) to learn a so-called channel chart. This channel chart is essentially a low-dimensional representation of the CSI with the salient property that users who are close in the channel chart are also close in physical space. Put simply: CC is a method that produces a pseudo-location with no recourse to classical positioning methods, potentially opening up a range of location-based applications to operate with significantly reduced overhead. Real-world experiments in single base-station scenarios have demonstrated the practical viability of CC. Despite these encouraging preliminary results, it is unclear whether CC can generate pseudo-location information at the radio access network (RAN) scale (e.g., for multiple base stations) with sufficient quality to replace true location information for certain location-based services and applications. The objective of this project is to develop methods and algorithms allowing to implement network-wide CC, develop its predictive capabilities when applied to real-world use cases with heterogeneous users and dynamically changing environments in which CSI may be affected by changes in the electromagnetic propagation environment. With the ultimate goals of developing CC into a robust and versatile pseudo-positioning method and of enhancing the expressivity of the channel chart beyond merely replacing location information to serve as a general basis for context-based services, we envision channel charting as a service (CaaS), which assists a number of network functions and user-level applications with a CC-based pseudo-location management architecture accessible through a dedicated application programming interface. To achieve these goals, we will perform fundamental research on the algorithm and implementation levels, validate our solutions by gathering long-term CSI measurements in dynamically changing environments, and optimize CaaS for radio resource management (RRM) in RAN. The project will lift CC from technology readiness level (TRL) 3 to TRL 4.

The expected impacts of the CHASER project lie in providing ML-based predictive RRM to enhance the efficiency, reliability, and latency of wireless communication. In particular, predictive handover and MIMO beamforming management will be performed. Furthermore, the project will consider new context-aware applications around which we will develop a set of representative benchmarks consisting of use cases, reference algorithm implementations, and CSI datasets enabling benchmarking of CC algorithms in a reproducible manner. We will also develop novel evaluation strategies of predictive tasks that rely on self-supervised ML methods in wireless systems. The cross-disciplinary research of CHASER will enhance long-term interaction between the machine learning, wireless communications, real-time, and distributed computing communities. We shall expand classical DR towards more realistic non-stationary observation models, provide solutions to distributed manifold alignment problems, and implement real-time dataset distillation. Thus, beyond its immediate impacts on the field of wireless communications, our research has the potential to significantly advance the versatility of the ML and DR toolbox and will strengthen the links with the real-time computing community.

Project partners

• Aalto University - Finland (coordinator)
• INRIA Lyon - France
• University of Minho - Portugal
• Eidgenössische Technische Hochschule Zürich - Switzerland