Adversarial Deep-Unfolding Networks (ADNs) for Symbol Detection in Communication Systems
Symbol detection is a fundamental and challenging problem in modern wireless communication systems, eg multiuser multiple-input multiple-output (MIMO), block-fading channels. Conventional model-based approaches assume known channel state information; thus, they have limited applicability in commonly unknown wireless channels. Recent advances in machine learning, eg deep learning, showed promising results in learning unknown non-linear channels, however, these neural network models require thorough time-consuming training of the networks before applying and are thus not readily suitable for highly dynamic channels in practice. We introduce an online training framework that can swiftly adapt to any changes in the channel. Our proposed framework unifies the recent deep-unfolding approach, which unfolds an iterative model-based algorithm into neural network layers with trainable parameters, with the emerging generative adversarial networks (GANs) to capture any changes in the channel and quickly adjust the deep unfolding networks to maintain the top performance of the model. We demonstrate that our framework significantly outperforms recent neural network models in various scenarios: 1) MIMO channels with interference where we unfold the Iterative Soft Interference Cancellation (SIC) algorithm; and 2) block-fading channels with inter-symbol interference where we unfold the Viterbi symbol detector. Our results show significant improvements over both model-based and deep learning approaches on highly dynamic channels and even surpasses those on the static channel in our experiments.