Managing Spatiotemporal Imbalance in Mobility on Demand Systems

This direction is merged into another direction: spatial demand imbalance. Its last version stays readable and citable here.

Version 1 of 123 papers23 conceptsfrozen 2026-10-04

In any modern city, transportation is a game of chasing ghosts. At 8:00 AM, thousands of commuters in residential neighborhoods need a ride to the city center. By 6:00 PM, the demand flips. This creates a fundamental "spatiotemporal imbalance": vehicles naturally accumulate where trips end, but new demand emerges where trips begin.

If a system only reacts to current requests, it will always be behind. Drivers will spend half their time "deadheading"—driving empty to reach far-away passengers—while riders wait in frustration. To solve this, researchers are moving beyond simple matching toward proactive rebalancing: the art of moving empty vehicles to where demand is about to be.

The Core Challenge: Predicting the Invisible

The first step in balancing a system is knowing where the imbalance exists. While ride-hailing apps have direct data on user requests, traditional taxi systems often do not. Shili Xiang, Yu Lu, and their colleagues demonstrated in Taxi Queue, Passenger Queue or No Queue? that we can infer passenger demand by looking at taxi state transitions. By analyzing when a taxi switches from "FREE" to "Occupied," the system can map out "hotspots" of unmet demand in real-time without needing a single GPS point from a passenger's phone.

Proactive Rebalancing: Moving Before the Request

Once demand is identified, the system must decide how to move the fleet. Traditional dispatching is reactive, but as Fei Miao and Shan Lin argue in their work on Receding Horizon Control, true efficiency requires looking into the future. By optimizing for both current needs and anticipated future demand, they showed that idle distance can be cut by more than half.

This proactive movement is known as rebalancing. In Efficient Collaborative Multi-Agent Deep Reinforcement Learning, Kaixiang Lin and Renyu Zhao frame this as a massive coordination problem. They use "contextual" reinforcement learning to help thousands of individual agents (cars) understand the global state of the city, allowing them to spread out or cluster together based on the time of day.

Scaling to the City: The Mean-Field Approach

As fleets grow to tens of thousands of vehicles, calculating the best move for every individual car becomes computationally impossible—a "curse of dimensionality."

To solve this, Kenan Zhang, Matej Jusup, and their team proposed a shift in perspective in Ride-Sourcing Vehicle Rebalancing with Service Accessibility Guarantees. Instead of tracking every car, they treat the fleet as a continuous "fluid" or distribution—a Mean-Field Reinforcement Learning approach. This allows the system to maintain service guarantees across an entire city by managing the density of cars rather than the specific path of each one.

Pricing as a Rebalancing Lever

Moving empty cars is expensive. An alternative is to move the people. In dockless bike-sharing systems, where there are no drivers to direct, the system must rely on the users themselves.

Ling Pan, Pingzhong Tang, and others developed Hierarchical Reinforcement Pricing to solve this. By offering small monetary incentives, the system encourages users to park their bikes in "deficit" areas or pick them up from "surplus" areas. This transforms the users into a distributed rebalancing workforce. Similarly, Pingzhong Tang and Song Zuo showed in Optimal Vehicle Dispatching Schemes via Dynamic Pricing that randomized pricing can outperform traditional "surge" pricing by more effectively smoothing out the peaks and valleys of urban demand.

Spatial Intelligence and Portability

A major hurdle in this research is that every city is different. A policy learned in Manhattan might fail in Singapore. Kaidi Yang addressed this in Graph Neural Network Reinforcement Learning for Autonomous Mobility-on-Demand Systems. By representing the city as a graph—where nodes are neighborhoods and edges are travel times—the system learns the underlying logic of urban movement. This "spatial intelligence" allows a model trained on one city's data to be deployed in another with almost no retraining, making the solution as mobile as the fleets it manages.

By combining real-time sensing, predictive modeling, and sophisticated incentive structures, this field is turning the chaotic "chase" of urban transit into a synchronized, self-balancing system.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_managing_spatiotemporal_imbalance_in_mob_82903e1c,
  title = {Managing Spatiotemporal Imbalance in Mobility on Demand Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/managing-spatiotemporal-imbalance-in-mobility-on-demand-systems/rdr_82903e1cbad44929a2cffc6f5c6b9618},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_managing_spatiotemporal_imbalance_in_mob_82903e1c_v1,
  title = {Managing Spatiotemporal Imbalance in Mobility on Demand Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/managing-spatiotemporal-imbalance-in-mobility-on-demand-systems/rdr_82903e1cbad44929a2cffc6f5c6b9618/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}