Distributed optimization techniques for 4G and beyond

Abstract

In today’s and future wireless telecommunication systems, the use of multiple antenna elements on the transmitter side, and also on the receiver side, provides users a better experience in terms of data rate, coverage and energy efficiency. In the fourth generation of such systems, precoding emerged as a relevant problem to be optimally solved so that network capacity can be increased by exploiting the characteristics of the channel. In this case, transmitters are equipped with few antenna elements, say up to eight, which means there are a few tens of precoding matrices, assuming a discrete codebook, to be coordinated per transmitter. As the number of antenna elements increases at communication nodes, conditions to keep in providing good experience for users become challenging. That’s one of the challenges of the fifth generation. Every transmitter must deal with narrow beams when a massive number of antenna elements is adopted. The hard part, regarding the narrowness of beams, is to keep the spatial alignment between transmitter and receiver. Any misalignment makes the received signal quality drop significantly so that users no longer experience good coverage. In particular, to provide initial access to unsynchronized devices, transmitters need to find the best beams to send synchronization signals consuming as little power as possible without any knowledge on unsynchronized devices’ channel states. This thesis thus addresses both precoding and beam finding as parameter coordination problems. The goal is to provide methods to solve them in a distributed manner. For this purpose, two types of iterative algorithms are presented for both. The first and simplest method is the greedy solution in which each communication node in the network acts selfishly. The second method and the focus of this work is based on a message-passing algorithm, namely min-sum algorithm, in factor graphs. The precoding problem is modeled as a discrete optimization one whose discrete variables comprise precoding matrix indexes. As for beam finding, a beam sweep procedure is adopted and the total consumed power over the network is optimized. Numerical results show that the graph-based solution outperforms the baseline/greedy one in terms of the following performance metrics: a) system capacity for the precoding problem, and b) power consumption for the beam finding one. Although message-passing demands more signaling load and more iterations to converge compared to baseline method, it usually provides a near-optimal solution in a more efficient way than the centralized solution.