Design of Noncoherent Communications: From Statistical Method to Machine Learning
Jianhao Huang, Muhang Lan, Han Zhang, Chuan Huang, Wei Zhang, Shuguang Cui
Abstract
The upcoming Internet of Things and fifth generation communications are expected to support short package transmissions with low complexity and low energy consumption, which motivates applications of noncoherent communications. First, we review the design methods for noncoherent communications based on two statistical schemes, that is, maximum likelihood (ML) decoding and energy-based decoding, which heavily rely on models of channel state information distributions. Then a data-driven machine learning method is proposed to design the noncoherent transceiver for short package transmissions. Neural networks are trained separately or jointly by utilizing finite channel realizations to construct the training samples. With the proposed method, two nondeterministic polynomial-time hard problems, joint transmitters design and ML decoding, are efficiently and approximately solved. Simulations reveal that the proposed machine learning method outperforms the conventional statistical method for cases with imperfect knowledge of the channel state information distributions or multiple transmitters.