Online configuration of reservable parking spaces: An agent-based deep reinforcement learning approach

Abstract

Unevenly distributed parking demand frequently leads to the overconsumption of popular parking lots, resulting in increased regional travel costs and traffic congestion. Configuring reservable parking spaces in parking lots based on online reservation systems is a prevalent solution to alleviate these issues. However, existing static configuration methods are inadequate for addressing time-varying parking demand, presenting significant challenges in determining the optimal number of reservable parking spaces across different parking lots over time. Thus, to address these challenges and reduce the total travel time in popular reservation-enabled management areas, this paper proposes a dynamic configuration model for reservable parking spaces utilizing agent-based deep reinforcement learning. The model can dynamically schedule the ratio of reservable parking spaces in an environment where reserved users and non-reserved users coexist, thereby influencing parking users’ choice behavior and balancing demand distribution. Experimental results on a real-world simulator show that, compared to baseline methods, the proposed model can effectively configure reservable parking spaces online. It conservatively reduces the total travel time by 21.4% and alleviates parking cruising and waiting in the management area. This approach is prospective for smart parking management.