[an error occurred while processing this directive]

Publications & Technical Reports
|
R71
Empirical Evaluation of Approximation Algorithms for Probabilistic Decoding
Irina Rish (irinar@ics.uci.edu), Kalev Kask (kkask@ics.uci.edu) & Rina Dechter (dechter@ics.uci.edu)

Abstract
It was recently shown that the problem of decoding messages transmitted through a noisy channel can be formulated as a belief updating task over a probabilistic network [14]. Moreover, it was observed that iterative application of the (linear time) belief propagation algorithm designed for polytrees [15] outperformed state of the art decoding algorithms, even though the corresponding networks may have many cycles. This paper demonstrates empirically that an approximation algorithm approx-mpe for solving the most probable explanation (MPE) problem, developed within the recently proposed mini-bucket elimination framework [4], outperforms iterative belief propagation on classes of coding networks that have bounded induced width. Our experiments suggest that approximate MPE decoders can be good competitors to the approximate belief updating decoders.

  [ps] [pdf]