چكيده به لاتين
Abstract
Device-to-device (D2D) communication is a dramatic departure from the conventional cellular architecture as it allows for user equipment (UE) in a cellular network to act as transmission relays for each other without the involvement of network infrastructures, realizing a co-existing massive ad-hoc network. While a hybrid D2D-cellular architecture can enhance the spectral efficiency, resource allocation in such a two-tier system is faced with unique challenges to ensure minimal impact on the performance of existing cellular users. In this paper, we address the D2D resource allocation problem under unknown channel state information (CSI). CSI-free schemes are important as certain practical limitations (e.g., finite CSI feedback delay) make the knowledge of instantaneous CSI impossible in systems with fading channels. Our proposal for extending the D2D resource optimization to unknown CSI settings is applicable only to a subset of scenarios that can be formulated as an equivalent graph-theoretic weighted bipartite matching problem. While at first sight, this may appear rather restrictive, we argue that still a fairly rich set of scenarios fall within this category, a claim also supported by citing numerous prior works. Without CSI, matching will pose as a combinatorial problem with unknown random edge weights and generally unknown distributions. To compensate for this lack of knowledge, we resort to the formalism of combinatorial multi-armed bandit (CMAB) from machine learning theory. A CMAB-based network controller can converge to the optimal matching configuration by following rules which can strike a proper balance between exploring alternative D2D configurations and exploiting the gradually accumulated knowledge. We formulate and numerically experiment with two problems as “proof of concept”, namely: (i) joint mode selection and relay assignment, and (ii) joint mode selection and channel allocation. We also compare with existing/baseline schemes with various flavors of CSI availability assumptions, including: perfect instantaneous CSI, perfect statistical CSI, erroneous CSI, as well as static CSI.
