Font Size: a A A

Probabilistic graphical models and variational Bayesian inference in receiver design for MIMO-OFDM systems

Posted on:2014-05-25Degree:Ph.DType:Dissertation
University:Purdue UniversityCandidate:Chen, ChulongFull Text:PDF
GTID:1458390008454447Subject:Engineering
Abstract/Summary:
Variational Bayesian methods performing approximate inference over complex probabilistic graphical models have been celebrated with success in many areas. In this work, a novel receiver design based on variational Bayesian inference for MIMO-OFDM systems is developed. The system architecture and underlying physical processes are modeled by sophisticated Bayesian probabilistic graphical models. Hierarchical probabilistic models are proposed to model the prior structures of signals of interest while retaining the conjugate property of the model: Two-layered hierarchical model is used to exploit the sparse property of multipath channels which lead to robust channel estimation with noisy measurements. And a hierarchical model combined with binary representation is employed to model prior knowledge to discrete transmitted symbols. The proposed approach lifts the need for unnecessary approximations on transmitted symbols in previous work in order to apply variational iterations. Therefore, the proposed scheme, for the first time, incorporate the channel estimation, equalization, detection and decoding into the unified variational Bayesian framework with minimum approximations. The structured mean field approximation for performing inference on the proposed Bayesian network is then developed. By preserving tractable substructures in the network, better estimation to the true posterior is achieved. The proposed framework naturally generates the approximate posterior probability of transmitted symbols which may be used for soft decoding. The proposed algorithms' performance is demonstrated by close-to-capacity bit error rate in a simulated LTE-like transceiver (with minimum simplifications). Low complexity implementations of the proposed algorithms are also developed. And their performance is also studied with simulations.
Keywords/Search Tags:Probabilistic graphical models, Variational bayesian, Inference, Proposed
Related items