Optimizing QoS in HD Map Updates: Cross-Layer Multi-Agent with Hierarchical and Independent Learning
Jeffrey Redondo, Nauman Aslam, Juan Zhang, Zhenhui Yuan
TL;DR
This work tackles the challenge of delivering HD map updates with low latency in dense VANETs by introducing a cross-layer, multi-agent reinforcement learning framework. It deploys three agents—CWmin, CWmax/IFSn interdependent via hierarchy, and a separate waiting-time learner—within a cross-layer design that directly accesses MAC parameters from the application layer. The approach demonstrates substantial latency reductions and improved HD Map throughput compared to IEEE802.11p EDCA and fixed CW baselines, with the three-agent configuration yielding the largest gains across Voice, Video, HD Map, and Best-effort services. The findings suggest that hierarchical and independent learning, coupled with cross-layer coordination, can effectively manage QoS in multi-service vehicular networks, enabling more reliable HD map dissemination in real-time deployments.
Abstract
The data collected by autonomous vehicle (AV) sensors such as LiDAR and cameras is crucial for creating high-definition (HD) maps to provide higher accuracy and enable a higher level of automation. Nevertheless, offloading this large volume of raw data to edge servers leads to increased latency due to network congestion in highly dense environments such as Vehicular Adhoc networks (VANET). To address this challenge, researchers have focused on the dynamic allocation of minimum contention window (CWmin) value. While this approach could be sufficient for fairness, it might not be adequate for prioritizing different services, as it also involves other parameters such as maximum contention window (CWmax) and infer-frame space number (IFSn). In response to this, we extend the scope of previous solutions to include the control of not only CWmin but also the adjustment of two other parameters in the standard IEEE802.11: CWmax and IFSn, alongside waiting transmission time. To achieve this, we introduced a methodology involving a cross-layer solution between the application and MAC layers. Additionally, we utilised multi-agent techniques, emphasising a hierarchical structure and independent learning (IL) to improve latency to efficiently handle map updates while interacting with multiple services. This approach demonstrated an improvement in latency against the standard IEEE802.11p EDCA by $31\%$, $49\%$, $87.3\%$, and $64\%$ for Voice, Video, HD Map, and Best-effort, respectively.