Deep Learning-assisted Multi-Dimensional Modulation and Resource Mapping for Advanced OFDM Systems

Multi-dimensional modulation (MDM) designed to exploit degrees of freedom across multiple component blocks in communication systems can offer a significant performance improvement compared to conventional modulation schemes. However, the optimal constellation and the corresponding bit to-symbol mapping still remain unsolved. In this paper, an efficient solution to optimize MDM, utilizing deep learning, is proposed. A newly designed neural network structure and an enhanced cost function are proposed for the joint optimization of the MDM constellation and the corresponding bit-to-symbol mapping, which performs in fast and stable manner. Symbol-to resource mapping and link adaptation procedure applicable to MDM for practical orthogonal frequency division multiplexing (OFDM) transceivers are also presented.