Finding maximum a posteriori solution for 2-D ISI channels
is intractable for the large sizes, therefore one must resort
to approximate inference methods such as message passing
algorithms. For clarity, the detection method is explained
through an example. This example will be used through the
remainder of the paper. Let us consider a 4 ⇥ 4 cell square
and assume that the read-head response spreads over 3 ⇥ 3
cells. The factor graph of our example is depicted in Fig. 4(a).
Variable nodes, denoted by Vi,j , are shown as circles and factor
nodes, denoted by fCi,j , as cubes with an edge connecting a
variable node Vi,j to a factor node fCi,j if and only if Vi,j is
an argument of fCi,j . As it is shown in Fig. 4(a), there exists
many cycles in 2-D ISI channel factor graph in contrast to
the sparse graphs of LDPC codes. As a result, the tree-like
assumption used in BP does not hold and BP approximation
is poor. GBP algorithm can be used to resolve this issue.
Finding maximum a posteriori solution for 2-D ISI channelsis intractable for the large sizes, therefore one must resortto approximate inference methods such as message passingalgorithms. For clarity, the detection method is explainedthrough an example. This example will be used through theremainder of the paper. Let us consider a 4 ⇥ 4 cell squareand assume that the read-head response spreads over 3 ⇥ 3cells. The factor graph of our example is depicted in Fig. 4(a).Variable nodes, denoted by Vi,j , are shown as circles and factornodes, denoted by fCi,j , as cubes with an edge connecting avariable node Vi,j to a factor node fCi,j if and only if Vi,j isan argument of fCi,j . As it is shown in Fig. 4(a), there existsmany cycles in 2-D ISI channel factor graph in contrast tothe sparse graphs of LDPC codes. As a result, the tree-likeassumption used in BP does not hold and BP approximationis poor. GBP algorithm can be used to resolve this issue.
การแปล กรุณารอสักครู่..
