A Backhaul Adaptation Scheme for IAB Networks Using Deep Reinforcement Learning With Recursive Discrete Choice Model

Abstract

Challenges such as backhaul availability and backhaul scalability have continued to outweigh the progress of integrated access and backhaul (IAB) networks that enable multi-hop backhauling in 5G networks. These challenges, which are predominant in poor wireless channel conditions such as foliage, may lead to high energy consumption and packet losses. It is essential that the IAB topology enables efficient traffic flow by minimizing congestion and increasing robustness to backhaul failure. This article proposes a backhaul adaptation scheme that is controlled by the load on the access side of the network. The routing problem is formulated as a constrained Markov decision process and solved using a dual decomposition approach due to the existence of explicit and implicit constraints. A deep reinforcement learning (DRL) strategy that takes advantage of a recursive discrete choice model (RDCM) is then proposed. The incorporation of the RDCM helps to address the challenges of backhaul availability and scalability with the aim to increase robustness to backhaul failure in IAB networks. The advantage of applying the RDCM for this problem is that it incorporates choice aversion from prospect theory and the reward is not the only factor affecting the learning rates, but also the punishment. The performance of the proposed algorithm was compared to that of conventional DRL, i.e., without RDCM, and generative model-based learning (GMBL) algorithms. The simulation results of the proposed approach reveal risk perception by introducing certain biases on alternative choices. For instance, applying cumulative prospect theory in the context of route choices reveal different node behaviors as the learning rate and noise levels are varied. The simulation results showed that the proposed algorithm provides better throughput and delay performance over the two baselines.