# A Cooperative Merging Method for Mixed Traffic Based on Enhanced Graph Reinforcement Learning with Vehicle Collaboration Graphs

## TL;DR

This paper introduces a new method for connected and autonomous vehicles to safely and efficiently merge in mixed-traffic environments using advanced graph-based reinforcement learning.

## Contribution

The novel VCG-EGRL algorithm uses a vehicle collaboration graph and enhanced graph reinforcement learning to improve cooperative merging decisions in mixed traffic.

## Key findings

- The proposed VCG-EGRL method outperforms baseline models in merging success rate.
- The method demonstrates improved efficiency and robustness in varying traffic conditions.
- The use of a local–global cooperative graph enhances the representation of vehicle interactions.

## Abstract

Achieving cooperative perception and decision-making among connected and autonomous vehicles (CAVs) in mixed-traffic ramp merge scenarios is crucial for building a swarm intelligence-based traffic control system. However, existing cooperative decision-making methods struggle to adequately model and represent the dynamic collaborative interactions among heterogeneous agents in mixed-traffic environments, which can lead to traffic congestion or even severe accidents in ramp merging areas. Therefore, this paper proposes an Enhanced Graph Reinforcement Learning algorithm based on a Vehicle Collaboration Graph (VCG-EGRL) to enable cooperative merging decisions for CAVs in mixed-traffic ramp merging scenarios. First, a vehicle collaboration intensity (VCI) model is designed to effectively model the intensity of collaborative interactions among vehicles. Then, based on the VCI model, the perception–communication relationships between vehicles and the vehicle-to-infrastructure (V2I) communication relationships are jointly constructed to form a local–global cooperative graph, which represents the dynamic collaborative relationships of the vehicle network from macro and micro perspectives and deeply explores the driving behavior of vehicles. Subsequently, a Graph Convolutional Network enhanced with Kolmogorov–Arnold Networks (KANs), referred to as GKAN, is employed to extract and aggregate the driving features of vehicles from the local–global graph. Finally, a graph mutual information maximization method is used to optimize the iterative process of the Graph Reinforcement Learning strategy, ensuring the generation of accurate lane-changing decisions for CAVs. Experimental results in ramp merging scenarios under varying traffic conditions demonstrate that the proposed method outperforms baseline models in terms of merging success rate, efficiency, and robustness.