Séminaire ETIS : David Declercq
Titre du séminaire
Beyond known Limits with Sparse Codes and Iterative Decoding.
Date et lieu
mardi 8 novembre 2011, 11h00.
ENSEA, salle du conseil (salle à confirmer).
In a first part, Finite Alphabet Iterative Decoders (FAID) will be presented. FAID can outperform usual iterative decoders such as the Min-Sum or the Belief-Propagation decoders, which are assumed to be the best known iterative decoders. Here, we use the fact that - although it is not possible to beat the BP decoder 'on average' over all possible graph realizations - it is possible to define a better decoder for a specific graph, for which the detailed topology is known. The issue of FAID decoder selection for universal or a fixed graph will be presented, and the concept of FAID decoder diversity will be introduced. We will show that FAID decoder diversity can address the very general problems of guaranteed error correction and MLD-approaching performance, only with iterative message passing decoders.
In a second part, we will discuss about the possibility to define non-binary cycle LDPC (NB-LDPC) codes that have linear growing minimum distance. Cycle NB-LDPC codes are especially interesting for practical purposes since the encoding/decoding/design is simplified compared to more general families of NB-LDPC codes, but suffer from the well known fact that their minimum distance is growing only logarithmically with the codeword length (like Turbo-Codes). A family of generalized non-binary cycle LDPC, called Cluster-NB-LDPC codes will be presented, with the property that the minimum distance is growing linearly with the code length, while still having the ultra-sparsity induced by the cycle-code structure. Both the asymptotic linear growth and the design of good finite length Cluster-NB-LDPC codes will be presented.