Colloquium - Details

Master-Vortrag: Optimized Compression Functions for Reduced Complexity Information Bottleneck Detection and Decoding

Sumedh J. Dongare
Montag, 31. August 2020
14:00 Uhr
virtueller Konferenzraum

The detection and the decoding at the receiver side is of crucial importance and often the optimum signal processing algorithms result in high implementation complexity. Therefore, sub-optimal algorithms with close to optimum performance are needed in practice. The information bottleneck method is a novel method for the detection and decoding with the low complexity that maximizes the mutual information. The main idea of such signal processing method is to design mutual information preserving mappings that replace the traditional signal processing operations to reduce complexity. These mappings are typically implemented as look-up tables. For instance, the literature successfully applies this method for the decoding of binary low-density parity-check codes. The low-density parity check codes are gaining more and more attention since their non-binary generalization has been found out which has better error correction capabilities than their binary equivalents. Due to the advancements in the computational capabilities of the devices, research in this field is a current hot topic. It turns out that the decoding of the non-binary low-density parity-check codes is an application which does not allow the straight forward application of the mutual information maximizing signal processing. The main problems are that the decoding requires systems with many input variables and in addition, the symbols from higher order fields can take more than two values. As a result the look-up table based approach which works well in case of binary codes, here results in look-up pf prohibitive size.

This motivates me to explore and investigate compression functions which can maximize the relevant mutual information but can be characterized using much fewer parameters than look-up tables. Such functions are designed in this thesis with a novel approach which relies on the genetic algorithms. Such algorithms are inspired from the natural evolution of the species. The novel approach allows to construct and analyze systems which cannot be designed with the look-up table based approach. This thesis compares the resulting systems to other state-of-the-art signal processing systems in terms of symbol error rate performance and also in terms of the ability to preserve relevant mutual information.

The reference system which is considered typically in this thesis is the soft symbol demodulator which has to be applied when non-binary low-density parity-check codes shall be used with binary modulation. Such demodulators designed using the novel approach are compared with the look-up table based approach and traditional soft symbol demodulators. The thesis develops a powerful class of parametrizable mappings that can be optimized using the genetic algorithms. Most importantly, the novel approach allows to achieve performance close to that of a soft symbol demodulator in many investigated scenarios.

back