This thesis addresses the problem of inference in factor graphs, especially the LDPC codes, almost solved by message-passing algorithms. In particular, the Belief Propagation algorithm (BP) is investigated as a particular message-passing algorithm whose suboptimality is discussed in the case where the factor graph has a loop-like topology. From the equivalence between the BP and the Bethe approximation in statistical physics that is generalized to the region-based approximation, is detailed the Generalized Belief Propagation algorithm (GBP), a message-passing algorithm between clusters of the factor graph. It is experimentally shown to surpass the BP in the cases where the clustering deals with the harmful topological structures that prevents the BP from rightly decoding any LDPC code, namely the trapping sets. We do not only confront the BP and the GBP algorithms according to their performance from the point of view of the channel coding with the error-rate, but also according to their dynamical behaviors for non-trivial error-events for which both algorithms can exhibit chaotic beahviors. By means of classical and original dynamical quantifiers, it is shown that the GBP algorithm can overcome the BP algorithm.

Dans cette thèse, nous étudions le problème de l'inférence bayésienne dans les graphes factoriels, en particulier les codes LDPC, quasiment résolus par les algorithmes de message-passing. Nous réalisons en particulier une étude approfondie du Belief Propagation (BP) dont la question de la sous-optimalité est soulevée dans le cas où le graphe factoriel présente des boucles. A partir de l'équivalence entre le BP et l'approximation de Bethe en physique statistique qui se généralise à l'approximation basée régions, nous détaillons le Generalized Belief Propagation (GBP), un algorithme de message-passing entre des clusters du graphe factoriel. Nous montrons par des expériences que le GBP surpasse le BP dans les cas où le clustering est réalisé selon les structures topologiques néfastes qui empêchent le BP de bien décoder, à savoir les trapping sets. Au-delà de l'étude des performances en termes de taux d'erreur, nous confrontons les deux algorithmes par rapport à leurs dynamiques face à des événements d'erreur non triviaux, en particulier lorsqu'ils présentent des comportements chaotiques. Par des estimateurs classiques et originaux, nous montrons que l'algorithme du GBP peut dominer l'algorithme du BP. / This thesis addresses the problem of inference in factor graphs, especially the LDPC codes, almost solved by message-passing algorithms. In particular, the Belief Propagation algorithm (BP) is investigated as a particular message-passing algorithm whose suboptimality is discussed in the case where the factor graph has a loop-like topology. From the equivalence between the BP and the Bethe approximation in statistical physics that is generalized to the region-based approximation, is detailed the Generalized Belief Propagation algorithm (GBP), a message-passing algorithm between clusters of the factor graph. It is experimentally shown to surpass the BP in the cases where the clustering deals with the harmful topological structures that prevents the BP from rightly decoding any LDPC code, namely the trapping sets. We do not only confront the BP and the GBP algorithms according to their performance from the point of view of the channel coding with the error-rate, but also according to their dynamical behaviors for non-trivial error-events for which both algorithms can exhibit chaotic beahviors. By means of classical and original dynamical quantifiers, it is shown that the GBP algorithm can overcome the BP algorithm.