Score-Based Generative Modeling for MIMO Detection Without Knowledge of Noise Statistics
Toluwaleke Olutayo, Benoı̂t Champagne · 2023
Motivated by recent advances in deep generative probabilistic modelling, we propose a robust multiple-input multiple-output (MIMO) symbol detector that aims to perform maximum likelihood (ML) detection without knowledge of the noise statistics. While the optimal MIMO detector (under uniform priors) is the ML detector, its implementation requires knowledge of the noise distribution. Furthermore, for some types of additive noise such as impulsive noise, the probability density function (PDF) of the noise does not admit a closed form expression thus making ML detection intractable. To overcome these limitations, our proposed approach learns a score function of the noise distribution directly from data. Subsequently, the learned score function is used to transform the noise distribution to a known (and tractable) prior distribution through the use of a stochastic differential equation. Via numerical simulations, the proposed detector is shown to outperform recent benchmark approaches for various types of additive noise, and to achieve near optimal ML performance where applicable.