A Markovian Traffic Equilibrium Model for Ride-Hailing

Song Gao, Hanyu Cheng, Chiwei Yan, Guocheng Jiang · 2026

Read A Markovian Traffic Equilibrium Model for Ride-Hailing by Song Gao, Hanyu Cheng, Chiwei Yan, Guocheng Jiang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-21359.txt.

Introduction

Song Gao and three colleagues model ride-hailing traffic balance with vehicles, empty or hired, making step-by-step order and link choices to maximize total future return. The model builds in rivalry among empty vehicles for passengers and congestion from road use at the link level. Balance becomes a fixed-point system whose existence they prove, with relaxed fixed-point iteration methods and convergence results for special networks. Ignoring congestion or drivers' forward-looking behavior, ablation studies show, can bias policy evaluation a lot.

Catalog record

CREATORSong Gao, Hanyu Cheng, Chiwei Yan, Guocheng Jiang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE90613
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.21359

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Markovian Traffic Equilibrium Model for Ride-Hailing Song Gao, Hanyu Cheng, Chiwei Yan, Guocheng Jiang Abstract We develop a Markovian traffic equilibrium model for ride-hailing in which vehicles, whether empty or hired, make sequential order-acceptance and link-choice decisions over a traffic network to maximize total discounted return in an infinite-horizon semi-Markov decision process. The model endogenizes both competition among empty vehicles for passenger demand and traffic congestion arising from road usage at the link level. We characterize equilibrium as the solution to a fixed-point system, establish its existence, and develop relaxed fixed-point iteration algorithms for equilibrium computation, with convergence results for specialized network structures. Computational experiments on realistic networks demonstrate the model’s practical value for transportation planning. Ablation analyses reveal that ignoring either traffic congestion or drivers’ forward-looking behavior can lead to potentially substantial biases in policy evaluation. keywords Markovian traffic equilibrium, ride-hailing † † runningauthor: Gao et al. † † runningtitle: Markovian Traffic Equilibrium for Ride-Hailing † † authors: Department of Civil and Environmental Engineering, University of Massachusetts Amherst, [email protected] Department of Industrial Engineering, Tsinghua University, Beijing, [email protected] Department of Industrial Engineering and Operations Research, University of California, Berkeley, [email protected] Department of Civil and Environmental Engineering, University of Massachusetts Amherst, [email protected] † † affiliation: † † affiliation: † † affiliation: † † affiliation: † † affiliation: † † affiliation: † † affiliation: † † affiliation: † † history: First version, 04/2026. 1 Introduction The expansion of ride-hailing services such as Uber and Lyft has continued steadily since their introduction more than a decade ago ( Yan et al. 2020 ) . Although the pandemic caused a temporary decline, the growth momentum has since recovered. Uber’s gross bookings, which measure the total dollar value of transactions on the platform and provide a more accurate indicator of service scale than revenue, reached $162 billion in 2024, representing an 18% increase from the previous year and more than an eightfold increase relative to 2016 ( Curry 2025 ) . Meanwhile, although autonomous vehicle (AV) technology has moved beyond the peak of its early hype, when major automakers claimed that fully autonomous fleets would be deployed by 2030, steady progress has nevertheless been made. Robotaxi services are now operating in dense cities with favorable weather conditions, such as San Francisco, CA and Phoenix, AZ through Waymo, and most recently in Austin, Texas through Tesla. Globally, China is leading the expansion of robotaxi services, with around 50 cities participating in pilot programs ( The Economist 2025 ) . Although ride-hailing trips currently account for only a small share of overall traffic volume, typically in the single-digit percentages overall but higher in major urban cores ( Bliss 2019 ) , it is conceivable that the continued expansion of both human-driven and AI-driven vehicles will eventually have a non-negligible impact on traffic conditions. A network equilibrium model for ride-hailing is therefore indispensable for addressing a wide range of planning, policy, and regulatory questions, and for enabling relevant agencies to proactively adopt effective strategies for the future transportation system. Our equilibrium model is built on a semi-Markov decision process (SMDP) framework in which ride-hailing vehicles make sequential order-acceptance and link-choice decisions in a traffic network to maximize discounted total return over an infinite horizon, subject to perception errors. These sequential decisions are coupled through flow-dependent congestion and competition for passenger demand, and they equilibrate endogenously within the system. This unified framework treats hired and empty vehicles in a consistent manner: they correspond to different states within a single underlying stochastic decision process. This feature distinguishes our work from existing studies of traffic equilibrium models for ride-hailing, or its predecessor, taxi services, which extend conventional deterministic path-based traffic equilibrium models by incorporating endogenous variables such as vehicle and customer waiting times to balance hired and vacant flows, while still treating the decision processes of hired and vacant vehicles separately ( Wong et al. 2001 , Wong et al. 2008 , Ban et al. 2019 , Xu et al. 2021 ) . The remainder of the paper is organized as follows. Section 2 reviews the related literature and summarizes our contributions. Section 3 presents the model, establishes the existence of the Markovian traffic equilibrium for ride-hailing (MTER), and develops a solution algorithm. Section 4 provides stronger theoretical results for directed-cycle networks. Section 5 reports computational experiments on stylized networks as well as the Sioux Falls and Chicago sketch networks. Section 6 briefly discusses an extension with endogenous driver participation. Finally, Section 7 concludes. 2 Literature Review The literature review is organized into three parts. The first reviews traffic equilibrium studies of taxi and ride-hailing systems with flow-dependent travel times. The second and third examine Markovian equilibrium models for traffic congestion and for ride-hailing, respectively, both of which provide a strong foundation for the integration of traffic congestion and vehicle competition developed in this paper. Traffic equilibrium for taxi/ride-hailing. One of the earliest network equilibrium models for for-hire vehicles is developed by Yang and Wong 1998 . Although traffic congestion is not considered, its core framework has been adopted and extended in subsequent work. In their model, a fixed number of taxis circulate throughout the network, and after dropping off passengers, vacant taxis choose repositioning destinations to minimize the sum of travel time to the target zone and expected waiting time upon arrival. A Logit model is used to capture perception errors and other unobserved factors. At equilibrium, the origin-destination flows of vacant taxis must match those of passengers, with zonal waiting times serving as the endogenous variables that clear the system. This model is extended in Yang et al. 2010 to incorporate bilateral customer-taxi search, in which both taxi and customer search times are endogenous, affect the choices of taxis and customers respectively, and are coupled through a Cobb-Douglas meeting function. In a different direction, Wong et al. 2001 incorporate traffic congestion through a bi-level model, where the lower level solves a combined trip distribution and traffic assignment problem given zonal trip totals of hired and vacant vehicles, which need not match, while treating customer and vehicle waiting times as free variables, and the upper level imposes equilibrium conditions that enforce hired and vacant flow balance, relate customer waiting time to vehicle waiting time, and characterize demand as a function of customer waiting time and travel time. Ban et al. 2019 further consider multiple ride-hailing platforms, each of which centrally determines the repositioning destinations of its drivers by solving a profit-maximization problem, in contrast to the decentralized driver decision-making assumed in most other work; in their model, solo driving is an outside option for travelers and also contributes to congestion. Finally, Xu et al. 2021 introduce an interzonal matching function that allows vehicles to be matched with customers from neighboring zones, thereby accounting for the contribution of both cruising and deadheading miles to traffic congestion. Markovian equilibrium for traffic only. Markov decision processes, or stochastic dynamic programming, provide an appealing framework when an agent’s behavior can naturally be modeled as a sequence of decisions, or when they offer computational advantages in problems with combinatorial solution spaces. An early example of the latter is Dial’s LOGIT traffic loading procedure for flow-independent travel times ( Dial 1971 ) , in which

Read the complete text