Learning to Design Constellation for AWGN Channel Using Auto-Encoders

This paper proposes a novel constellation design in AWGN channel through learning based auto-encoder (AE). Additionally, this paper illustrates the reason why learning based constellation has better performance than the classical square-shaped QAM design by analyzing the Euclidean distance distribution and the bound of symbol error rate between learning designed symbols and other constellations. Moreover, the performance of learning based constellation will be compared to constellation based on convex optimization design. To solve the bit mapping problem of the learning based constellation, $Q-$ary LDPC encoding is applied to these specifically designed QAM modulation systems, where the soft decoding of $Q-$ary LDPC codes can be carried out with the symbol-level soft outputs of demodulation.