System Identification in Wireless Relay Networks via Gaussian Process

TL;DR

This paper introduces a Gaussian Process-based stochastic model for identifying relay functionalities in wireless networks, providing efficient algorithms that balance accuracy and computational complexity, with promising simulation results.

Contribution

It develops a novel Gaussian Process framework for relay system identification and proposes three algorithms with different complexity-accuracy trade-offs.

Findings

At worst 9% total error compared to the lower bound at low SNR

Relay functional estimation error impacts BER by less than 2dB

Algorithms perform well across various relay functionalities

Abstract

We present a flexible stochastic model for a class of cooperative wireless relay networks, in which the relay processing functionality is not known at the destination. In addressing this problem we develop efficient algorithms to perform relay identification in a wireless relay network. We first construct a statistical model based on a representation of the system using Gaussian Processes in a non-standard manner due to the way we treat the imperfect channel state information. We then formulate the estimation problem to perform system identification, taking into account complexity and computational efficiency. Next we develop a set of three algorithms to solve the identification problem each of decreasing complexity, trading-off the estimation bias for computational efficiency. The joint optimisation problem is tackled via a Bayesian framework using the Iterated Conditioning on the Modes methodology. We develop a lower bound and several sub-optimal computationally efficient solutions to the identification problem, for comparison. We illustrate the estimation performance of our methodology for a range of widely used relay functionalities. The relative total error attained by our algorithm when compared to the lower bound is found to be at worst 9% for low SNR values under all functions considered. The effect of the relay functional estimation error is also studied via BER simulations and is shown to be less than 2dB worse than the lower bound.