TL;DR
This paper introduces a novel "multi-level speed guidance cooperative approach" for traffic signal control in intelligent and connected vehicle (ICV) environments. By combining dynamic speed guidance with bidirectional green wave coordination, the method aims to minimize traffic delays and maximize throughput, showing significant improvements over traditional methods.
What Problem Does This Paper Address?
Traditional traffic signal control, even with green wave coordination, often struggles to fully utilize available green light time and adapt to dynamic traffic conditions. This leads to increased vehicle delays, stops, and reduced overall road network efficiency. Existing methods often rely on fixed signal timings and ideal green wave bandwidths, failing to account for the complexities of real-world traffic, such as varying vehicle speeds, multi-lane approaches, and the potential for bidirectional traffic flow. The authors aim to address these limitations by developing a more adaptive and collaborative control strategy for ICVs.
Methods
The core of this paper is a bi-level combinatorial optimization framework called DQNGA (Deep Q Network-Genetic Algorithm).
- Lower Layer (Speed Guidance): This layer focuses on individual vehicle speed control. It uses a Deep Q-Network (DQN), a reinforcement learning algorithm, to determine optimal speed guidance for vehicles entering the control area. The DQN learns to adapt speed based on real-time traffic states, signal phases, and predicted future conditions, aiming to ensure vehicles pass through intersections smoothly and within green light windows. The paper details both two-stage and four-stage speed guidance methods to achieve this.
- Upper Layer (Green Wave Coordination): This layer optimizes the phase offsets between intersections to maximize the bidirectional green wave bandwidth. It employs a Genetic Algorithm (GA), which searches for the optimal combination of phase offsets that best supports the speed guidance from the lower layer.
- Integration: The bi-level approach allows for a synergistic optimization, where the lower layer's speed guidance informs the upper layer's phase offset optimization, and vice-versa, leading to a globally optimal solution.
Datasets
The study utilizes real-world traffic data collected from:
- Microwave radar sensors and onboard diagnostics (OBD) devices in Yangzhou, China.
- On-board GPS data from 68,342 connected vehicles.
- Roadside cameras and LiDAR measurements for vehicle counts.
- VISSIM COM interface for simulating vehicle trajectories.
The data covers three periods: morning peak, flat peak, and evening peak.
Key Insights and Ideas
- Synergy of Speed and Signal Control: The paper highlights that optimizing speed guidance and signal timing simultaneously is crucial for maximizing traffic flow, rather than treating them as independent problems.
- Bidirectional Green Waves: Explicitly considering and optimizing for bidirectional green waves (forward and reverse traffic) is key to improving overall intersection efficiency, especially on arterial roads.
- Adaptive Control is Essential: Static or rule-based traffic control is insufficient for modern ICV environments. Adaptive methods like DQNGA are necessary to dynamically adjust to changing traffic conditions.
- DQNGA's Power: The combination of DQN for adaptive speed guidance and GA for phase offset optimization proves effective in tackling the complex combinatorial problem of traffic control.
Key Results
The proposed DQNGA method demonstrated significant improvements over the conventional MAXBAND approach:
- Reduced Average Delay: 20.76% reduction (as stated in the abstract, and further detailed in tables showing specific reductions like 24.32% and 28.85% in flat peak and morning peak periods, respectively).
- Reduced Number of Stops: 44.49% reduction (as stated in the abstract, and further detailed in tables showing specific reductions like 53.56% and 50.00% in morning peak periods).
- Increased Vehicle Throughput: The number of continuously passing vehicles in the comprehensive green wave bandwidth increased significantly (e.g., ~31.58% in the forward direction and ~35.71% in the reverse direction during the morning peak for Jiangyang Middle Road).
- Enhanced Computational Efficiency: The GA component showed improved convergence rates and reduced computational time compared to other algorithms like BSA and FA.
- Superior Performance: DQNGA outperformed other algorithms (CDRL, MADQN, QT-CDQN, DTA) in terms of reward and computational efficiency, and achieved lower delays and stops.
Limitations and Future Work
- Vehicle Connectivity Assumption: The model assumes full connectivity of vehicles. Integrating traditional vehicles (mixed traffic flow) presents a challenge due to the lack of data on their behavior.
- Transition Time Sensitivity: Future work will focus on analyzing the sensitivity of the model to vehicle acceleration and deceleration transition times.
- Dynamic Weight Adjustment: The authors suggest exploring dynamic weight adjustments in the objective function to prioritize bandwidth maximization during peak hours and energy efficiency during off-peak hours, enhancing scenario-specific adaptability.
- Data Availability: The study relies on specific data from Yangzhou. Generalizability to other urban environments may require further validation.
Summary updated