On the Joint Optimization of Signal Constellations and Bit Mappings

Abstract

This letter presents an enhanced gradient search algorithm for jointly optimizing signal constellations and bit mappings. This approach stands in contrast to the recent trend of utilizing deep learning (DL) to model communication systems as end-to-end (E2E) systems, which often involve a large number of learnable parameters and high computational complexity during training. We enhance the efficiency of the gradient search algorithm by leveraging the symmetrical properties of the I/Q plane, particularly in higher modulation orders. The resulting constellations are evaluated in terms of bit error rate (BER) performance under additive white Gaussian noise and Rayleigh flat fading channels. Our findings indicate that the optimized constellations obtained via the enhanced gradient search algorithm outperform an attention-empowered DL-based E2E system in terms of BER across both channels, notably in higher signal-to-noise ratio regimes, without encountering error floor issues.