Entropic harmony: Graph-based multi-agent reinforcement learning with dynamic decay for mixed traffic in highway

The evolution of intelligent transportation systems is driving a paradigm shift in highway driving from isolated individual intelligence to collaborative swarm intelligence. However, realizing efficient coordination in mixed traffic flows remains impeded by two critical challenges: the non-stationarity of reward signals, which leads to inconsistent cooperative behaviors, and the overwhelming perceptual noise within complex vehicle interaction topologies. To bridge these gaps, this paper proposes a robust Multi-Agent Reinforcement Learning framework incorporating Adaptive Reward Entropy and Dynamic Graph Attention. First, to address the instability in coordination, we devise a Reward Entropy mechanism with a Sliding-Window Coordination Factor. Unlike static historical baselines, this approach dynamically tracks local reward extrema to calibrate entropy regularization, effectively guiding the transition from disordered individual exploration to synchronized group performance. Second, to tackle environmental noise, we introduce a Graph-Structure-Based Multi-Head Attention Network equipped with a Dynamic Decay Adapter. By modeling vehicle interactions as a topological graph, this module precisely extracts critical spatial-temporal dependencies while automatically attenuating irrelevant interference from distant vehicles. Extensive evaluations in the CARLA simulation platform demonstrate that our method exhibits highly competitive coordination efficiency and robust convergence characteristics across varying traffic densities compared to established multi-agent baselines.