A review on reinforcement learning methods for mobility on demand systems

TL;DR

The paper addresses the challenge of operating Mobility-on-Demand systems using reinforcement learning by surveying RL-based strategies for vehicle dispatch, rebalancing, and joint control. It introduces a unified sequential-decision-making lens and categorizes approaches into policy-function, cost-function, value-function, and direct-lookahead approximations, emphasizing deep RL, graph neural networks, and actor-critic architectures. Key contributions include a structured literature synthesis across MoD tasks, critical appraisal of evaluation methodologies, and identification of research gaps such as benchmarking consistency and transferability. The findings highlight that rebalancing is the most studied sub-task, urban settings dominate, and there is a push toward open platforms and transferable policies to enable real-world deployment and cross-city adaptation. Overall, the work charts a path for more rigorous, comparable, and scalable RL-based MoD control in increasingly complex and heterogeneous mobility environments.

Abstract

Mobility on Demand (MoD) refers to mobility systems that operate on the basis of immediate travel demand. Typically, such a system consists of a fleet of vehicles that can be booked by customers when needed. The operation of these services consists of two main tasks: deciding how vehicles are assigned to requests (vehicle assignment); and deciding where vehicles move (including charging stations) when they are not serving a request (rebalancing). A field of research is emerging around the design of operation strategies for MoD services, and an increasingly popular trend is the use of learning based (most often Reinforcement Learning) approaches. We review, in this work, the literature on algorithms for operation strategies of MoD systems that use approaches based on Reinforcement Learning with a focus on the types of algorithms being used. The novelty of our review stands in three aspects: First, the algorithmic details are discussed and the approaches classified in a unified framework for sequential decision-making. Second, the use cases on which approaches are tested and their features are taken into account. Finally, validation methods that can be found across the literature are discussed. The review aims at advancing the state of the art by identifying similarities and differences between approaches and highlighting current research directions.