Deep Reinforcement Learning for Backhaul Link Selection for Network Slices in IAB Networks

TL;DR

This paper introduces a DRL-based method for dynamic backhaul link selection in IAB networks to efficiently manage capacity and congestion in 5G mobile networks.

Contribution

It proposes a novel DRL approach using DDQN for backhaul link selection in IAB networks, addressing the challenge of dynamic traffic and multiple slice configurations.

Findings

DDQN agent successfully performs backhaul selection without failures.

The proposed method converges after approximately 20 training episodes.

A simple neural network architecture suffices for effective decision-making.

Abstract

Integrated Access and Backhaul (IAB) has been recently proposed by 3GPP to enable network operators to deploy fifth generation (5G) mobile networks with reduced costs. In this paper, we propose to use IAB to build a dynamic wireless backhaul network capable to provide additional capacity to those Base Stations (BS) experiencing congestion momentarily. As the mobile traffic demand varies across time and space, and the number of slice combinations deployed in a BS can be prohibitively high, we propose to use Deep Reinforcement Learning (DRL) to select, from a set of candidate BSs, the one that can provide backhaul capacity for each of the slices deployed in a congested BS. Our results show that a Double Deep Q-Network (DDQN) agent using a fully connected neural network and the Rectified Linear Unit (ReLU) activation function with only one hidden layer is capable to perform the BS selection task successfully, without any failure during the test phase, after being trained for around 20 episodes.