TL;DR
This paper introduces a cooperative decision-making framework for Connected Automated Vehicles (CAVs) navigating complex multi-lane merging zones. By leveraging coalitional game theory and Model Predictive Control (MPC), the system allows vehicles to dynamically form cooperative groups (coalitions) based on their personalized driving styles (aggressive, moderate, conservative). This approach uniquely balances the overall traffic efficiency of the system with the individual preferences and safety requirements of each vehicle.
What Problem Does This Paper Address?
Highway lane merging is a classic bottleneck that causes traffic congestion, reduces efficiency, and increases the risk of collisions. While Connected Automated Vehicles (CAVs) offer a solution through vehicle-to-vehicle (V2V) communication, existing control strategies usually fall into two extremes.
Centralized controllers optimize the entire traffic flow but force all vehicles to surrender control to a master system. This ignores passenger preferences—for example, a family wanting a smooth, conservative ride versus a commuter prioritizing aggressive efficiency. Conversely, single-vehicle decision-making prioritizes the individual but degrades overall traffic efficiency and can lead to selfish, unsafe merging behaviors. This paper addresses this gap by proposing a decentralized, cooperative approach for multi-lane merging that respects individual, human-like driving characteristics while still optimizing for system-wide safety and efficiency.
Methods
The authors model the multi-lane merging scenario as a coalitional game, a type of cooperative game theory where players (vehicles) can form alliances to reduce their individual costs.
The framework consists of several key components:
- Motion Prediction: The system uses a simplified single-track bicycle model to predict the future states (position, velocity, yaw) of surrounding vehicles.
- Cost Function via MPC: Formulated using Model Predictive Control, the decision-making cost function balances three metrics: safety (longitudinal/lateral distance and lane-keeping), comfort (minimizing jerk), and efficiency (maintaining desired velocity).
- Personalized Driving Styles: Vehicles are assigned one of three profiles—aggressive, moderate, or conservative. These profiles dictate the weighting coefficients in the cost function (e.g., an aggressive profile heavily weights efficiency over comfort).
- Coalition Formation: Vehicles dynamically evaluate four types of coalitions: single-player, multi-player, a "grand coalition" (all vehicles cooperating), or a grand coalition with sub-coalitions. Crucially, the system enforces individual rationality: a vehicle will only join a coalition if its allocated cost within the group is less than or equal to its cost of acting alone.
Datasets
Not discussed in the paper. The authors evaluate their framework using custom simulated driving scenarios built in the MATLAB/Simulink platform rather than relying on pre-existing static datasets.
Key Insights and Ideas
- The "Grand Coalition" is optimal, but not always rational: Mathematically, a grand coalition where all vehicles cooperate minimizes the total cost for the system. However, if vehicles have mismatched driving styles (e.g., one aggressive driver among conservative ones), the aggressive vehicle might suffer a higher individual cost by cooperating. The game-theoretic model elegantly allows that vehicle to break away, mimicking real-world human behavior.
- Dynamic Alliances: Vehicles with similar goals and driving styles can form "sub-coalitions" (e.g., two moderate vehicles merging together) to simplify the game and lower their collective cost, effectively teaming up to navigate the merge.
- Personalization is Key: By embedding driving characteristics directly into the cost function weights, the framework proves that future autonomous systems don't have to act like rigid, uniform robots. They can exhibit personalized, human-like behaviors while remaining safe and predictable to other connected vehicles.
Key Results
- Case 1 (3-Vehicle Scenario): When all vehicles had "moderate" driving styles, they successfully formed a grand coalition, reducing the total system cost from 166,679 (if acting alone) to 165,265. However, when one vehicle was switched to "aggressive," it rejected the grand coalition because its individual cost would have increased (from 71,527 to 72,903). The system successfully adapted by allowing it to act as a single-player coalition while the others cooperated.
- Case 2 (5-Vehicle Scenario): The algorithm successfully managed complex, dynamic splitting and merging of coalitions based on mixed driving styles. For instance, aggressive vehicles successfully broke away to form their own sub-coalitions when conservative vehicles refused to yield.
- Computational Efficiency: The algorithm demonstrated strong real-time potential, with an average computational time of just 0.06 to 0.07 seconds per step in the MATLAB/Simulink environment.
Limitations and Future Work
The authors acknowledge a few limitations in the current study. First, the framework assumes a 100% CAV environment where all vehicles can communicate and participate in the game, which does not reflect near-term transitional traffic. Second, the game currently only includes direct participants and immediately adjacent vehicles, ignoring the broader impact on vehicles further upstream or downstream.
For future work, the authors plan to expand the model to handle mixed traffic environments—where CAVs must interact with unpredictable, non-connected human-driven vehicles—and to incorporate interactions between multiple, spatially separated coalitions.
Summary updated