A deep reinforcement learning approach for the asynchronous dynamic vehicle dispatching proble

Francisco Edyvalberty A. Cordeiro, Anselmo R. Pitombeira-Neto,
A deep reinforcement learning approach for the asynchronous dynamic vehicle dispatching problem,
Artificial Intelligence for Transportation,
Volumes 3–4,
2025,
100038,
ISSN 3050-8606,
https://doi.org/10.1016/j.ait.2025.100038.
(https://www.sciencedirect.com/science/article/pii/S3050860625000389)
Abstract: This paper addresses the asynchronous dynamic vehicle dispatching problem (DVDP), in which vehicle assignments must be made in real time as requests arise or vehicles become available. Unlike traditional synchronous dispatching, the asynchronous DVDP poses unique modeling and computational challenges due to its event-driven nature and stochastic dynamics. We formulate the problem as a semi-Markov decision process and propose a novel solution approach that combines discrete-event simulation with deep reinforcement learning (DRL). Two specialized agents, trained via double deep Q-learning, handle decisions triggered by new requests and freed vehicles, respectively. The proposed approach is validated through a taxi fleet management case study using real-world data from New York City. Computational experiments show that the learned policy outperforms standard asynchronous heuristics, achieving up to a 50.6 % reduction in passenger delay and an 18.4 % reduction in cancellation rates compared with batch policies in resource-constrained scenarios. This study highlights the potential of DRL for optimizing asynchronous dispatching in dynamic and uncertain environments.
Keywords: Dynamic vehicle dispatching; Asynchronous decision-making; Semi-Markov decision process; Deep reinforcement learning; Discrete-event simulation