Activity detection for next generation multiple access in wireless communications : from optimization to model-based learning

Abstract

To support sixth-generation (6G) mobile services with stringent requirements, deep learning has emerged as a powerful tool for future wireless system design. Unlike conventional optimization-based approaches, deep learning does not rely on high-precision mathematical models, and once trained, it can infer results in real-time. Although being the mainstream approach in deep learning, black-box-based design has obvious disadvantages in wireless communication research. Due to the lack of problem-specific architecture, it requires a large amount of training data, often encounters difficulty in convergence, and performs poorly in the trained network. To overcome these drawbacks, model-driven deep unfolding has emerged as a viable alternative. By embedding domain knowledge into the network design via mimicking a suitable mathematical algorithm, it is competitive with the original optimization algorithm while leveraging the advances in deep learning. But model-driven deep unfolding depends on well-defined mathematical algorithms. This highlights the indispensable role of optimization algorithms even in the era of deep learning. To shed light on how to incorporate model-driven deep unfolding method for future wireless system design, this thesis focuses on the activity detection task in grant-free random access, which is a crucial component to empower future ubiquitous connectivity. For basic single-cell activity detection, common approaches include compressed sensing-based algorithms and covariance-based algorithms. Beyond this basic setting, the activity detection task becomes more challenging. Specifically, this thesis tackles the following three challenges. 1) For joint activity and data detection, it involves an extra discontinuous sparsity constraint and hence lacks an efficient detection algorithm with performance guarantee; 2) For intelligent reflecting surface (IRS)-aided system, the cascaded channels pose challenges in obtaining the distribution of the equivalent channel, thus complicating the detection algorithm design. 3) For the cell-free systems, capacity-limited fronthauls require signal compression and quantization at the access points before forwarding to the central processing unit, making the end-to-end formulation intractable. To address the above challenges, the first part of this thesis derives a new difference-of-norms optimization framework for the joint activity and data detection. Based on the resulting iterative algorithm, the deep unfolding network is designed to further enhance the detection performance. Next, for inaccurate formulation of activity detection in IRS-aided system, the second part of this thesis formulates an approximated optimization problem. Closed-form update algorithm is derived based on the projected gradient approach for the approximated problem and the algorithm is unfolded for deep learning with additional learning parameters to compensate the mismatch issue. Finally, for the intractable formulation of activity detection in cell-free system, the third part of this thesis proposes an augmented deep unfolding learning framework where unfolding technique is leveraged for the detection module, while compression/quantization module is constructed by judiciously designed neural networks. The three cases studied in this thesis exemplify the embedding of optimization process into deep unfolding when 1) effective algorithms are lacking, 2) precise problem modeling is absent, and 3) problem modeling is intractable. These examples not only tackle current challenges in activity detection but also establish the model-based learning paradigm and guidelines for other future wireless system designs.