Dual-Interaction-Aware Cooperative Control Strategy for Alleviating Mixed Traffic Congestion
New AI control system improves traffic flow by distinguishing cooperative CAV interactions from unpredictable human drivers.
A research team led by Zhengxuan Liu has published a paper proposing a novel AI-driven strategy to tackle one of the most complex challenges in autonomous transportation: traffic congestion in mixed environments with both Connected and Automated Vehicles (CAVs) and unpredictable Human-Driven Vehicles (HDVs). The core innovation is the Dual-Interaction-Aware Cooperative Control (DIACC) strategy, which operates within a Multi-Agent Reinforcement Learning (MARL) framework. It directly addresses the major hurdle of HDV uncertainty by teaching CAVs to intelligently distinguish between cooperative interactions with other CAVs and purely observational interactions with human drivers, enabling more effective and safe cooperation in bottleneck areas like highway merges.
The DIACC architecture introduces three key technical components. First, a Decentralized Interaction-Adaptive Decision-Making (D-IADM) module enhances each CAV's local perception by categorizing interactions. Second, a Centralized Interaction-Enhanced Critic (C-IEC) provides a global traffic understanding for better policy guidance. Third, a specialized reward function uses softmin aggregation with temperature annealing to prioritize learning in complex, interaction-heavy scenarios. The paper reports that DIACC outperforms existing rule-based and standard MARL benchmarks, demonstrating marked improvements in overall traffic flow and system adaptability. This research, shared on arXiv, represents a significant step toward practical, large-scale deployment of CAVs by providing a robust AI control system capable of navigating the messy reality of shared roads.
- Uses a Multi-Agent Reinforcement Learning (MARL) framework to enable CAVs to cooperate and alleviate bottlenecks.
- Introduces a novel module to distinguish CAV-CAV cooperative interactions from CAV-HDV observational interactions, handling human driver uncertainty.
- Experimental results show significant improvements in traffic efficiency and adaptability over existing rule-based and MARL benchmark models.
Why It Matters
Paves the way for smoother, safer integration of self-driving cars into existing traffic by making them smarter collaborators with human drivers.