Multi-Player Multi-Armed Bandit Based Resource Allocation for D2D Communications

TL;DR

This paper models D2D resource allocation in 5G networks as a multi-armed bandit problem, proposing algorithms that operate with partial CSI to optimize power and interference management.

Contribution

It introduces novel multi-armed bandit algorithms for D2D resource allocation under partial CSI, extending to multi-user scenarios with fairness considerations.

Findings

Algorithms achieve fair resource distribution.

Proposed methods outperform baseline in simulations.

Effective with partial channel information.

Abstract

Device-to-device (D2D) communications is expected to play a significant role in increasing the system capacity of the fifth generation (5G) wireless networks. To accomplish this, efficient power and resource allocation algorithms need to be devised for the D2D users. Since the D2D users are treated as secondary users, their interference to the cellular users (CUs) should not hamper the CU communications. Most of the prior works on D2D resource allocation assume full channel state information (CSI) at the base station (BS). However, the required channel gains for the D2D pairs may not be known. To acquire these in a fast fading channel requires extra power and control overhead. In this paper, we assume partial CSI and formulate the D2D power and resource allocation problem as a multi-armed bandit problem. We propose a power allocation scheme for the D2D users in which the BS allocates power to the D2D users if a certain signal-to-interference-plus-noise ratio (SINR) is maintained for the CUs. In a single player environment a D2D user selects a CU in every time slot by employing UCB1 algorithm. Since this resource allocation problem can also be considered as an adversarial bandit problem we have applied the exponential-weight algorithm for exploration and exploitation (Exp3) to solve it. In a multiple player environment, we extend UCB1 and Exp3 to multiple D2D users. We also propose two algorithms that are based on distributed learning algorithm with fairness (DLF) and kth-UCB1 algorithms in which the D2D users are ranked. Our simulation results show that our proposed algorithms are fair and achieve good performance.