Description
Date depot: 1 janvier 1900
Titre: New Baseband Architectures using ML/DL in Presence of non-linearities and Dynamic Environment
Directeur de thèse:
Jérémie SUBLIME (LISITE)
Encadrant :
Yahia MEDJAHDI (LISITE)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini
Resumé:
The increasing complexity foreseen in B5G and 6G is expected to go over the capabilities of classical mathematical modelling tools. This can be a challenging problem especially in dynamic environments e.g. time variant channels. In the literature, many model-based solutions investigate the detection under changing channel conditions. However, these solutions assume always that the channel is invariant in coherence time . Accordingly, it is critical for these solutions to have accurate knowledge of instantaneous CSI, which entails an important overhead affecting thus the data rate.
In B5G, data-driven solutions are expected to play a major role in the design of future transceiver architectures. Motivated by the recent achievements on deep learning over the past few years, there has been significant interest in the application of machine and deep learning to various communications scenarios , with a focus on the physical layer [Shea17]. Recently, some works have proposed DL-based channel estimation techniques. Note that many of the ML/DL-based solutions consider stationary wireless communication systems. However, these solutions may not work in fast time-varying channels where the outdated CSI has detrimental effects on the performance. Relatively few studies have been carried out to address this issue. The authors in [Far18] have proposed Sliding Bidirectional Recurrent Networks (SBRNN) to learn varying molecular channels. Inspired by the previous work, Sliding Bidirectional Gated Recurrent Units (SBGRU) have been proposed in [Bai20] to dynamically track time varying channels. Compared to traditional algorithms and other Neural Networks schemes, SBGRUs provide better performance and robustness.
Main targets of this PhD proposal :
Our aim is to investigate the capability of ML in the design of new receiver architectures considering time-varying channels and power amplifier nonlinearities. In such varying environment, the quantity of training data is extremely limited due to the shortness of the channel coherence time. Two approaches can be explored: a pure data-driven approach and a combination of data-driven and model-driven. We believe that the hybrid approach can provide a significant improvement compared to the pure data-driven strategy since domain knowledge (e.g. correlation, approximated models …) is exploited in the design of the receiver.
Doctorant.e: Dos Reis Ana Flávia