Bernard Fleury, "Variational Inference Applied to MIMO-OFDM - A Message Passing View"


Bernard Fleury, Aalborg University, Denmark


Message-passing has proven to be an efficient, tractable tool to solve Bayesian inference problems in telecommunications and in numerous fields in engineering and natural sciences. The efficiency combined with the intrinsic iterative structure of message-passing-based solutions is often reflected in the prefix “turbo” added in naming them: turbo-decoding, turbo-equalization, etc.

In this talk, we discuss a reformulation of the mean-field method for conjugate-exponential models in Bayesian networks: variational-Bayesian message passing (VBMP). We shortly revisit the theory elaborated by Winn and Bishop (2004), and apply it to a Bayesian network representation of the received signal in a generic MIMO-OFDM system. Thus, a message passing algorithm for parameter estimation and detection is derived for a general case in which no particular assumptions on the amount of information already available at the receiver are made.

Using this general algorithm as a starting point, we impose different sets of assumptions and restrictions to the signal model, obtaining as a result a variety of solutions for different particular scenarios. Among others, we present a pilot-based channel estimator for MIMO transmission with overlapped pilot symbols and an iterative receiver structure for MIMO channel estimation, noise estimation and symbol detection. We also briefly address topics as the integration of the channel decoder in the framework, optimal message scheduling, convergence properties of the iterative algorithm and computational complexity considerations.

Finally, we sketch in an outlook a unifying framework embedding variational-Bayesian message passing and belief-propagation message passing, and stress the importance of such a unified view for receiver design.

When Fr, 29. Jan 2010 14:00 – 15:30
Where EI1 lecture hall, TU Wien
Export iCal , vCal