Adaptive Measurement Strategies for Network Optimization and Control

QC 20250303

Abstract

The fifth generation mobile networks has become the new network standard, with the sixth generation on the horizon, and its new technological capabilities have enabled a far wider variety of services compared to the fourth generation networks. To ensure that these services can co-exist and meet their standardized requirements, the network's resources must be provisioned, managed and reconfigured in a far more complex manner than before. As such, it is no longer sufficient to select a simple, static scheme for gathering the necessary information to take decisions. Instead, it is necessary to adaptively, with regards to network system dynamics, trade-off the cost in terms of power, CPU and bandwidth consumption of the taken measurements to the value their information brings. Network optimization is a wide field, and the way to quantify the value of a given measurement heavily depends on the problem studied. As such, this thesis addresses adaptive measurement schemes for a number of well-defined network optimization problems. The thesis is presented as a compilation, where after an introduction detailing the background, purpose, problem formulation, and contributions of our work, we present a more detailed view of the use cases, the theoretical, technological and methodological background together with an outline of our results. Finally we go into even more detail about our work through the appended conference and journal papers.

First, we study the problem of optimal spectrum access for low priority services. We assume that the network manager has limited opportunities to measure the spectrum before assigning one (if any) resource block to the secondary service for transmission, and this measurement has a known cost attached to it. We study this framework through the lens of multi-armed bandits with multiple arm pulls per decision, a framework we call predictive bandits. We analyze such bandits and show a problem specific lower bound on their regret, as well as design an algorithm which meets this regret asymptotically, studying both the case where measurements are perfect and the case where the measurement has noise of known quantity. Studying a synthetic simulated problem, we find that it performs considerably better compared to a simple benchmark strategy.

Secondly, we study a variation of admission control where the controller must select one of multiple slices to enter a new service into. The agent does not know the resources available in the slices initially, and must instead measure these, subject to noise. Mimicking three commonly used admission control strategies, we study this as a best arm identification problem, where one or multiple arms is "correct" (the arm chosen by the strategy if it had full information). Through this framework, we analyze each strategy and devise sample complexity lower bounds, as well as algorithms that meet these lower bounds. In simulations with synthetic data, we show that our measurement algorithm can vastly reduce the number of required measurements compared to uniform sampling strategies. 

Thirdly, we study a network monitoring system where the controller must detect sudden changes in system behavior such as batch traffic arrivals or handovers, in order to take future action. We study this through the lens of change point detection but argue that the classical framework is insufficient for capturing both physical time aspects such as delay as well as measurement costs independently, and present an alternative framework which decouples these, requiring more sophisticated monitoring agents. We show, both through theory and through simulation with both synthetic data and data from a 5G testbed, that such adaptive schedules qualitatively and quantitatively improve upon classical change point detection schemes in terms of measurement frequency, without losing classical optimality guarantees such as the one on required measurements post change, and can do so even when very little is information a priori about the distribution of measurements post-change.

Finally, we combine the problems of dynamic spectrum access and change monitoring by studying change detection in a finite-state Markov chain model, with the two-state model as an interesting special case. Here, both classical methods as well as our methods from the case of independent measurements are insufficient to capture the complexities of the problem. We show fundamental limits of measurements on any agent in this case and study two heuristic policies, showing their advantages and disadvantages to both each other and a naive measurement strategy. Then, we evaluate their performance and measurement costs through simulation both with synthetic data and data from Wi-Fi spectrum utilization measurements.

urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-360485