Distributed Stochastic Optimization for Traffic-Aware Efficient Design of 5G Wireless Networks

Retour à la liste des thèses
Ajouté le: 21/05/2014
Directeur : ASSAAD Mohamad - mohamad.assaad@supelec.fr
Titre : Distributed Stochastic Optimization for Traffic-Aware Efficient Design of 5G Wireless Networks
Thèmes : Automatique, Signal, Télécoms, Systèmes embarqués
Laboratoires : E3S Supélec Sciences des Systèmes EA 4454
Description :

 Advisors: Mohamad Assaad – Mohamad.Assaad@supelec.fr
                   Richard Combes – Richard.Combes@supelec.fr

Context

The proliferation of wireless multimedia applications necessitates the development of more advanced wireless systems that can support the expected high amount of mobile data traffic in the next years. It has been adopted by the 3GPP that the future 5G cellular networks must support the 1000-fold increase in traffic demand. This requires developing new physical layer techniques, e.g. Massive MIMO and adopting a new architecture of the network since the increase of the capacity of macro cells cannot meet the requirement of the high traffic demands. Future 5G networks are expected to be dense, self-organizing, and having a user-centric and distributed architecture. This necessitates the development of decentralized algorithms that require very low information exchange between the nodes. In addition, the use of massive MIMO introduces additional challenges, mainly in terms of complexity, since the number of resources is very high.


Research Proposal

The objective of this thesis is to develop an arsenal of smart distributed algorithms that optimize the network resource utilization in 5G dense networks taking into account the traffic pattern, the advanced physical layer (massive MIMO) and the signaling overhead. The resource allocation problems can be formulated as an optimization problem where each user has a payoff function to maximize. This payoff function must incorporate the key physical layer advances and traffic patterns in such a way to capture the QoS/QoE satisfaction requirements of the users. Two cases will be explored in this thesis.

i) When the reward function has a predefined structure (e.g. convex, Lipschitz, unimodal etc.), the resource allocation problem can be formulated as a multi-armed bandit with the objective to minimize the convergence time (or minimize the regret).
ii) When the reward function does not have a simple mathematical model or when only a numerical value of the reward can be estimated, the problem is also a stochastic optimization problem but it is more generic and based only on numerical observations. Distributed learning policies based on extremum-seeking techniques can be developed in this case.