Enhancing latency performance through intelligent bandwidth allocation decisions: a survey and comparative study of machine learning techniques

Converged access networks consolidating 5G and beyond and fixed optical fiber access are expected to support future latency-sensitive human-to-machine applications over the Tactile Internet. Making intelligent bandwidth allocation decisions among end users/machines/robots of the converged network is thus crucial to meeting stringent latency requirements. The recent renewed interest in machine learning (ML) has contributed towards a plethora of undeniable performance improvements in communication networks. Current insights into how ML can be exploited to provide intelligent bandwidth allocation decisions to enhance latency performance, along with guidance on the most suitable ML technique in that regard, remain elusive. This paper provides the first insights, to the best of our knowledge, into the suitability of commonly adopted ML techniques for this purpose by first presenting an in-depth survey focusing on the technical details of these techniques and how each technique is used in existing studies. The benefits, drawbacks, resultant time and space complexity incurred, and prediction accuracy are then evaluated for each ML technique reviewed. Next, a comprehensive comparative study of the ML techniques is presented for the first time, to our best knowledge, to provide guidance on the selection of ML technique that provides intelligent bandwidth allocation decisions towards supporting emerging latency-sensitive applications. The uplink latency performance of a converged network adopting an artificial neural network (ANN) supervised bandwidth allocation scheme is then compared with those arising from using existing bandwidth allocation schemes. Results highlight the ability of the ANN to learn the association among bandwidth demand, network parameters, and the resulting uplink latency such that, in operation, the allocated bandwidth will always be optimized to enhance latency performance.