An Incentive Algorithm for a Closed Stochastic Network: Data and Mean-Field Analysis

Bianca Marin Moreno, Christine Fricker, Hanene Mohamed, Amaury Philippe, Martin Trépanier2024 · DOI

Abstract

Free-floating car-sharing systems frequently suffer from vehicle imbalance, where cars migrate from high-demand city centers to peripheral residential areas. This paper investigates a "gift" incentive policy used by Communauto in Montreal to encourage passive rebalancing. By integrating empirical analysis of 1.5 million transactions with a mean-field mathematical model, we demonstrate that targeted user incentives can reduce station vacancy rates from 60% to 20%. However, we also identify a "saturation trap" where excessive rebalancing may lead to parking shortages in high-demand zones.

Introduction

Free-floating car-sharing systems provide flexible urban mobility but face significant logistical challenges. Without intervention, vehicles naturally accumulate in low-demand areas, leaving high-demand "hotspots" empty. Operators typically address this through active rebalancing (manual relocation) or passive rebalancing (user incentives). This study focuses on the latter, specifically a "gift" policy where users receive 30 minutes of free driving for moving stagnant cars from low-demand zones to designated high-demand zones.

Empirical Study and Dataset

The research utilizes a proprietary dataset from Communauto, the primary car-sharing operator in Montreal. The data spans October 2019 to October 2020, encompassing approximately 1.48 million transactions. The service area, covering roughly 100 km², was partitioned into a high-demand "Intensive Zone" (Zone 1) and a larger "Normal Zone" (Zone 2).

Map of the Montreal service area divided into high-demand (Intensive Zone) and normal zones.

The analysis focused on vehicle idle times and user engagement with the incentive program. A key observation was the timing of vehicle availability: the supply of "gift" cars typically peaks at noon, just before the city center experiences its daily minimum vehicle count, suggesting the policy is well-aligned with natural demand cycles.

Modeling Approach: Mean-Field Analysis

To evaluate the system's performance at scale, the authors modeled the car-sharing service as a closed stochastic network. Tracking individual vehicles in a large fleet is computationally prohibitive; therefore, a mean-field limit was applied. This approach approximates the discrete behavior of hundreds of agents using a system of Ordinary Differential Equations (ODEs). This mathematical framework allows for the prediction of the system's "steady state"—the long-term average percentage of empty or full stations—under various fleet sizes and incentive parameters.

Key Findings

The empirical data revealed a highly skewed participation rate in the incentive program, following a Zipf’s law distribution. Over 90% of users never utilized a "gift" trip, while a small group of "power users" performed the vast majority of rebalancing. For instance, the top-ranked user alone accounted for 2.6% of all "used gift" trips.

Daily average number and proportion of gift-eligible and used-gift trips across winter and summer periods.

On average, "gift" trips were 31.3% cheaper for users than standard trips. The effectiveness of the policy was most evident in high-vacancy scenarios: the incentive could reduce the vacancy rate in the intensive zone from 60% to 20%.

Discussion: The Saturation Trap and Limitations

While effective at preventing empty stations, the gift policy introduces the risk of "saturation." If the high-demand zone is too small or the incentive too aggressive, the influx of cars can exceed local parking capacity, preventing other users from ending their trips.

The study also notes several limitations. The current mean-field model assumes exponential trip time distributions, which does not fully capture the "heavy-tailed" nature of real-world car-sharing (where many trips are very long). Additionally, the model treats zones as homogeneous, ignoring micro-level popularity differences between streets.

Conclusion

The combination of empirical data and mean-field modeling provides a robust framework for optimizing car-sharing incentives. While passive rebalancing via "gifts" is a powerful tool for maintaining system equilibrium, operators must carefully calibrate zone sizes and incentive levels to avoid the transition from vehicle scarcity to parking saturation.

Summary updated

BibTeX
@inproceedings{moreno2024incentive,
  author = {Bianca Marin Moreno and Christine Fricker and Hanene Mohamed and Amaury Philippe and Martin Trépanier},
  title = {{An Incentive Algorithm for a Closed Stochastic Network: Data and Mean-Field Analysis}},
  booktitle = {La},
  year = {2024},
  doi = {10.1007/s44007-024-00098-x},
}
Figure