A Deep Reinforcement Learning and Graph Convolution Approach to On-Street Parking Search Navigation

Area of Science:

• Urban Planning and Traffic Management
• Artificial Intelligence and Machine Learning
• Operations Research

Background:

• Efficient parking distribution is vital for urban traffic management but faces challenges from variable demand and spatial disparities.
• Current research often focuses on local optimization, neglecting real-time allocation complexities in metropolitan areas.
• Key issues include dynamic supply-demand imbalance and the need for spatial resource optimization to enhance system performance and user satisfaction.

Purpose of the Study:

• To develop a novel framework for dynamic parking allocation that addresses real-time demand fluctuations and spatial inefficiencies.
• To improve overall parking system performance and user satisfaction in complex urban environments.
• To overcome the limitations of static allocation solutions in managing variable parking demand.

Main Methods:

• A Multi-Agent Reinforcement Learning (MARL) framework integrating adaptive optimization and intelligent collaboration.
• A reinforcement learning-driven temporal decision mechanism for real-time parking assignment adjustments.
• A Graph Neural Network (GNN)-based spatial model to analyze and optimize inter-parking relationships.

Main Results:

• The MARL framework significantly outperforms conventional methods like FIFO and SIRO in managing demand variability.
• Demonstrated substantial improvements in optimizing resource distribution for parking.
• Validated the framework's strength and flexibility across various urban contexts using real-world data.

Conclusions:

• The proposed MARL framework offers an effective solution for dynamic parking allocation in urban settings.
• The integration of temporal and spatial optimization models enhances parking management efficiency.
• This approach provides a robust and adaptable method for improving urban traffic flow and user experience.