Integrating Multi-Head Attention into MADRL for Robust Longitudinal Control of CAVs in Mixed Traffic with Unreliable Communication

Cai, Shuncheng, Garg, Mohit and Bouroche, Melanie (2026) Integrating Multi-Head Attention into MADRL for Robust Longitudinal Control of CAVs in Mixed Traffic with Unreliable Communication. In: IEEE Intelligent Vehicles Symposium, Proceedings. IEEE, Plymouth, MI, USA, pp. 1488-1495. ISBN 979-8-3315-4794-3

Full text not available from this repository.

Abstract

Ensuring reliable longitudinal control of Connected and Autonomous Vehicles (CAVs) in mixed traffic with unreliable communication presents a critical challenge. Existing learning-based approaches often rely on idealized assumptions and suffer from cold start issues, where agents act conservatively or inconsistently in the early driving phases due to a lack of prior experience, which limits their effectiveness in dynamic and uncertain conditions. This paper proposes an advanced Multi-Agent Deep Reinforcement Learning (MADRL) framework designed to improve the longitudinal control of CAVs operating in mixed traffic environments under unreliable vehicle-to-vehicle (V2V) communication. Building on a previously proposed MADRL algorithm, we introduce a multi-head attention (MHA) mechanism designed to improve state representation and enhance overall control performance. The integration of this mechanism presents two significant benefits: (1) dynamic input adaptability, allowing the model to adjust to varying number of CAVs efficiently, and (2) global dependency modelling, which aids in capturing essential inter-vehicle relationships vital for strong and stable multi-agent interactions. Results demonstrate that the MHA-enhanced MADRL strategy improves control stability and responsiveness over value-based MAPPO, distribution-based MAPPO, and MQA-enhanced MAPPO across different range of CAV penetration rates and packet-loss levels.

Item Type:	Book Section
Uncontrolled Keywords:	Longitudinal Control; Mixed Traffic Environments; Multi-Agent Deep Reinforcement Learning (MADRL); Multi-Head Attention (MHA) Mechanism; Unreliable V2V Communication
Subjects:	T Technology > TL Motor vehicles. Aeronautics. Astronautics Q Science > QA Mathematics > Algebra > Algorithms > Computer algorithms Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Staff Research and Publications
Depositing User:	Tamara Malone
Date Deposited:	26 Aug 2026 16:01
Last Modified:	26 Aug 2026 16:01
URI:	https://norma.ncirl.ie/id/eprint/9675

Actions (login required)

View Item