Back to search

University of Minnesota

Autonomous Mobility-on-Demand Systems: Numerical Optimal Control and Reinforcement Learning Approaches

Abstract

dc:description.abstract

This thesis addresses the mathematical modeling, planning, and operational control of Autonomous Mobility-on-Demand (AMoD) and Electric AMoD (E-AMoD) systems through a combination of model-based and learning-free approaches. First, we develop an optimal control approach to optimize vehicle scheduling and rebalancing (dispatching available empty vehicles to areas undersupplied) in an AMoD system. We propose a linear, discrete-time model of an AMoD system, followed by deriving the desired number of rebalancing vehicles for the AMoD model through an optimization problem. To leverage the proposed model, we implemented the model predictive control (MPC) framework to find the optimal rebalancing and scheduling policy. We demonstrate the MPC's effectiveness and show how it can be implemented in real-time through a real-world case study. Building on this foundation, we extend the framework to E-AMoD systems that model vehicle flows across time, space, and energy states. The framework is built on a fluidic, graph-theoretic model that enforces customer queue, charging queue, node, and link conservation to capture vehicle state transitions and demand dynamics. We derive analytical lower bounds on the minimum fleet size required to satisfy steady-state demand and on the minimum number of chargers required at each station. The control strategy is MPC to manage customer dispatching, rebalancing, and charging in real time. Simulations on the University of Minnesota–Twin Cities campus network demonstrate that the proposed model and strategy maintain queue stability and achieve efficient service with moderate fleet sizes. To complement model-based control, we propose a model-free, real-time, data-efficient Q-learning-based algorithm to solve the H∞ control of linear discrete-time (DT) systems, which reduces computational complexity from 𝒪(q³)to 𝒪(q²), where q depends quadratically on the sum of state variables, control inputs, and disturbances. The algorithm updates the parameters of the actor and critic networks using only a single data point per iteration, and probing noise is only needed in the first iteration, with no impact on the algorithm’s performance. Without requiring an initial stabilizing policy, the algorithm converges to the closed-form solution of the Riccati equation. A simulation study demonstrates the effectiveness of the proposed algorithm in real-time control of an AMoD system. The thesis emphasizes the critical role of infrastructure planning, particularly fleet sizing and charger placement, in shaping system performance. Moreover, the interplay between long-term planning and real-time operational policies such as dispatching, rebalancing, and energy scheduling is highlighted.

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Aalipour, Ali

Subjects

dc:subject × 6

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Citation
Aalipour, Ali. (2025). Autonomous Mobility-on-Demand Systems: Numerical Optimal Control and Reinforcement Learning Approaches. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/282365.

Chain of custody

source
Harvested from
University of Minnesota
Base URL
conservancy.umn.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Aalipour, Ali. Autonomous Mobility-on-Demand Systems: Numerical Optimal Control and Reinforcement Learning Approaches. 2025. https://hdl.handle.net/11299/282365