Abstract
The coordination of connected autonomous vehicles (CAVs) at intersections presents a critical challenge for intelligent transportation systems, demanding safe and efficient traffic flow without centralized control. Multi-agent reinforcement learning (MARL) has emerged as a promising paradigm, yet scalability and inter-agent communication remain open issues. This paper proposes a novel MARL framework, termed CommNet-S, which integrates a learnable communication protocol with a centralized training and decentralized execution paradigm to enable scalable coordination among CAVs at unsignalized intersections. The framework employs a deep Q-network architecture augmented with attention-based message passing, allowing agents to selectively share state information. We evaluate CommNet-S in a simulated four-way intersection environment with varying traffic densities, comparing against independent Q-learning, and a state-of-the-art centralized controller. Results demonstrate that CommNet-S achieves up to 34% higher throughput and 28% lower average delay compared to baselines, while maintaining collision-free operation. Communication overhead is analyzed, showing that the selective attention mechanism reduces bandwidth usage by 60% relative to full broadcast. Ablation studies further highlight the importance of message content and recipient selection. The findings underscore the viability of communication-based MARL for real-world intersection management, offering a scalable solution that balances performance and resource efficiency.