Optimizing Parameters in Soft-hard BPGD for Lossy Source Coding
Masoumeh Alinia, David G. M. Mitchell · 2023
In this paper, we investigate lossy source coding based on a soft-decision based belief propagation guided decimation (BPGD) algorithm for low-density generator matrix (LDGM) codes, called soft-hard BPGD. For this algorithm, the optimal performance can only be achieved by carefully selecting the softness parameters which are conventionally found by exhaustive empirical search. To address this issue, we have introduced a framework that leverages the cavity method to predict the values of softness parameters at which phase transition occurs. This framework can significantly reduce the need for time-consuming empirical searches to determine the best parameter values for soft-hard BPGD. Our approach has been found to deliver superior rate-distortion performance compared to the hard-decimated BPGD algorithm and faster implementation than the soft-decimated algorithm.