Transformer Based Multi-Vehicle Collaborative Decision-Making Under Multi-Dimensional Information Fusion

This paper addresses the critical challenge of enabling effective collaborative decision-making among autonomous vehicles in complex dynamic traffic scenarios by developing a novel multi-agent reinforcement learning framework that integrates multi-dimensional scene information. Our approach uniquely combines vehicle motion characteristics, position information, and driving intentions through a comprehensive information fusion strategy, then employs a Transformer-enhanced QMIX algorithm that leverages selfattention mechanisms to extract features from fused information while maintaining collaborative decision-making capabilities. The framework establishes a local-global observation paradigm under intelligent networking technology that enables coordinated actions through distributed information processing, advancing beyond existing methods that either ignore inter-vehicle interactions or fail to effectively process multi-dimensional scene information in real-time collaborative scenarios. By integrating Transformer's feature extraction capabilities with QMIX's collaborative behavior generation, our method addresses the limitations of rule-based approaches and current MARL methods in complex road structures with dynamically changing scenes. Experimental results demonstrate that our proposed method significantly outperforms baseline approaches including VDN, QMIX, and their Transformer variants, showing improved autonomous vehicle scene comprehension and enhanced collaborative decision-making performance in challenging multi-vehicle coordination tasks.