Client-Centric Radio Access Technology Selection in Heterogeneous Networks

Abstract

Heterogeneity of modern wireless network radio access technologies (RATs) (e.g., 3G/4G/LTE, Wi-Fi, Bluetooth, 5G) is an intrinsic part of providing seamless connectivity and network access in next-generation wireless networks. With modern mobile edge devices increasingly equipped with a variety of wireless interfaces to access these heterogeneous networks (HetNets), these edge clients are capable of dynamically switching between different access networks to optimize their performance.

Such freedom comes at a cost: specifically, the added requirement and complexity of determining which network a client should select at any time. Traditionally, solutions for RAT selection in HetNets focus on the network-centric approach, with a centralized controller solving a global optimization, but this approach is not only non-scalable due to required signaling, but also the practical problem of non-cooperative network operators. However, with the increasing processing power of client edge devices (e.g., smartphones, tablets, etc), it is worth considering if and how RAT selection should be done at the client instead of in the network.

In this dissertation we study RAT selection in HetNets from a client-centric perspective under varying degrees of network-provided information, where the client wishes to maximize throughput on its selected RAT: (1) perfect network knowledge (clients have perfect information on other client-RAT association configurations), (2) partial network knowledge (clients have time-averaged statistics for their own client-RAT channels), and (3) no network knowledge (clients have no statistics provided by the RAT).

In (1), we model the problem as a non-cooperative game between different players under perfect information, and design the client-centric distributed RAT Selection Algorithm using binary exponential backoff and hysteresis (local memory of previous RATs) to guarantee convergence to a bounded Nash Equilibrium.

In (2), we model the problem as a multi-armed bandit with switching costs where the clients only know each channels’ spectral (eigenvalue) gap of its transition matrix, and develop the mHS algorithm that obtains optimal O(ln(t)) total regret.

In (3), we solve the multi-armed bandit problem in the absence of network-provided information by applying reinforcement learning to map locally-obtainable measurements to empirically-obtained past throughputs in our WIFFN algorithm to achieve O(ln(t)) regret.