Long‐term Perspectives
Sławomir Stańczak, Alexander Keller, Renato L. G. Cavalcante, Nikolaus Binder, Soma Velayutham · 2021
Mobile communications resources (spectrum and energy) are scarce. In addition, signals in the real world are distorted by noise and interference while the statistics of wireless connections are highly dynamic due to mobility. Machine learning (ML) has great potential to cope with these challenges and to significantly increase the capacity of wireless networks, but novel ML methods need to be developed to fully exploit this potential. This chapter examines ML techniques for future wireless access networks (RAN) and in particular the physical layer. We argue in favor of hybrid methods that benefit from both model-based approaches and data-driven approaches while incorporating domain knowledge. In particular, we will consider the combination of kernel-based methods and deep neural networks. Key to efficient algorithms is the exploitation of structures intrinsic to wireless channels, transmission signals, and inferred and measured data. While structure like sparsity and compressibility can be designed right into the algorithms, data is not available at a single point but distributed across different locations. Nevertheless, the novel ML methods need to provide robust results based on small datasets and under latency constraints. This requires a massively parallel and distributed implementation of future wireless networks. In fact, the massive parallelism offered by graphics processing unit (GPU) architectures allows for concurrent execution of certain radio access network (RAN) functions to achieve orders-of-magnitude acceleration. Based on open standards, a popular software ecosystem (compute unified device architecture [CUDA]), and speed of deployment on components-off-the-shelf (COTS), such truly software-defined radios (SDRs) are highly attractive, especially in self-hosted campus networks.