An Unsupervised LLR Estimation with unknown Noise Distribution

Many decoding schemes rely on the log-likelihood ratio (LLR) whose derivation depends on the knowledge of the noise distribution. In dense and heterogeneous network settings, this knowledge can be difficult to obtain from channel outputs. Besides, when interference exhibits an impulsive behavior, the LLR becomes highly non-linear and, consequently, computationally prohibitive. In this paper, we directly estimate the LLR, without relying on the interference plus noise knowledge. We propose to select the LLR in a parametric family of functions, flexible enough to be able to represent many different communication contexts. It allows limiting the number of parameters to be estimated. Furthermore, we propose an unsupervised estimation approach, avoiding the need of a training sequence. Our estimation method is shown to be efficient in large variety of noises and the receiver exhibits a near-optimal performance.