Section: New Results

Results of axis 2: algorithms

1. Massive random access in LPWAN

   Table 6.

   Principal Investigator:	Jean-Marie Gorce, Claire Goursaud
   Students:	Diane Duchemin, Lélio Chetot
   Funding:	ANR Ephyl, Inria-Nokia common lab
   Partners:	Sequans, Supelec Rennes, ISEP, CEA Leti, Nokia
   Publications:	[30], [31], [37], [47]

   The optimization of IoT access techniques was the objective of the ANR Ephyl collaborative project, where we studied different solutions at the PHY and MAC layers as presented in [47].

   The main question Maracas group addressed in this research is the detection of simultaneous random transmissions from distributed nodes. The underlying mechanism is a coded slotted Aloha allowing to avoid hand-skake mechanisms. Each node can transmit randomly and the receiver tries to detect several packets simultaneously. Our objective is to identify a good code family, and to determine the fundamental trade-off in terms of nodes density versus reliability. During this year, we focused on the detection of a small subset of simultaneous active nodes, exploiting optimal detection. We developed a MAP based iterative detector at a multi-antennas receiver in [30]. We also proposed a low complexity detector in [37].

   This joint coding-decoding optimization problem will be also investigated from extensive simulations and experimental data (see section 3.4), and represents an interesting problem to evaluate deep learning based approaches.

2. Interference management

   Table 7.

   Principal Investigator:	Léonardo Cardoso, Jean-Marie Gorce
   Students:	Hassan Khalam
   Funding:	Fed4PMR (PIA)
   Partners:	Thales
   Publications:	[41]

   Interference management and resource management is a very complex problem in wireless environment (e.g. [55]). The capacity region is known for some specific scenarios and some specific channel conditions. But the optimal performance relies on perfect feedback mechanisms, to get channel state information at the transmitters and to coordinate them. As proposed by Jafar et al, topological interference management (TIM) [56] is a seducing framework to balance performance with feedback complexity. In the context of the Fed4PMR project, we develop new algorithms to allow partial coordination between interfering transmitters [41], relying only on some partial interference information. This approach suits particularly well with the requirements of PMR networks, since their deployments is not optimized. The algorithm relies on an association of degrees of freedom evaluation, graph theory and interference alignment.

   Based on this first study, we will explore the suitability of TIM in other application scenarios (especially for the standard IEE802.11ax under preparation). For short, TIM allows to build optimal graph representations of a wireless networks, with reduced coordination needs. TIM can be seen as an approach to optimally quantize a complex interfering graph and to distribute its knowledge in an optimal fashion.

3. Learning in radio systems

   Table 8.

   Principal Investigator:	Léonardo Cardoso, Malcolm Egan, Jean-Marie Gorce
   Student:	Cyrille Morin, Mathieu Goutay
   Funding:	ADR Analytics, Inria-Nokia common lab
   	AI chair ANR program (applied)
   Partners:	Jakob Hoydis, Nokia Bell Labs
   Publications:	[45]

   Following the artificial intelligence tsunami, the research community in wireless systems (both industry and academia) is engaged in a strong competition to determine how this revolution could change the paradigm of wireless networks. Following the preliminary studies made by Jakob Hoydis [54], we investigate in this research action, the potential of deep learning in radio communications. The central question is to identify which processing could take advantage from neural networks against classical approaches.

   Our joint strategy with Nokia follows: we target the production of a huge set of experimental data with FIT/CorteXlab to facilitate the comparison of different solutions and to train neural networks on real data. We currently investigate three original problems : transmitter identification from its RF signature (Cyrille Morin PhD) [45], self-synchronization procedures based on neural networks (Cyrille Morin PhD) and dirty RF compensation (Mathieu Goutay PhD, patents submitted). Last but not least, we believe that an intelligent radio should be able to learn from its environment and to adapt its behavior. Therefore, in the future, we will explore reinforcement principles associated to neural networks and applied to learning based radio.

   This topic is very hot, and most top ranked conference have special sessions on this topic. We believe that our partnership with Nokia, our data sets from FIT/CorteXlab and our experience in estimation theory let us be highly competitive.