Impairment-Aware Machine Learning Approaches to Link Adaptation

• Applicant: Dr.-Ing. Sander Wahls
• Subject Area: Electronic Semiconductors, Components and Circuits, Integrated Systems, Sensor Technology, Theoretical Electrical Engineering
• Term: from 2011 to 2015
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 214177965

Wireless channels are highly volatile. Therefore, the transmission parameters of a wireless communications system should be continuously adapted to the state of the channel. This process is commonly referred to as link adaptation. Except a few, current link adaptation algorithms use theoretical models in order to predict the performance of each admissible choice of transmission parameters. The choice with the best predicted performance is then used. While this approach often works well in simulations, it fails in real-world situations where a whole bunch of impairments such as hardware non-linearities, synchronization and estimation errors, interferers, or feedback delay hamper the validity of the theoretical model. The usual resort is to tune the link adaptation using performance feedback (e.g., packet error rates). However, the tuning itself is an intricate task and commonly solutions are quite simplistic. Recently, the use of more sophisticated algorithms from machine learning has been proposed as a way out of this situation. Link adaptation algorithms based on machine learning treat the wireless system as a black box. They do not rely on simplified theoretical models, but instead learn by relating chosen transmission parameters, observed link quality metrics, and resulting performance measurements. An important point here is the choice of link quality metrics. Current algorithms only observe signal-to-noise ratios. However, many impairments depend on observable parameters like the RF state or the Doppler spread. With those, anticipation of impairment errors becomes possible. The development of machine learning based link adaptation algorithms that exploit such parameters is the first goal of the fellowship. On the other hand, there are impairments that cannot be anticipated in this way but that lead to a sudden change in performance. The second goal is a link adaptation that learns how to react quickly to such sudden changes without having to retrain.

• DFG Programme: Research Fellowships
• International Connection: USA
• Host: Professor H. Vincent Poor, Ph.D.