Research Highlight: Shared Autonomous Mobility-on-Demand

05 Dec 2020

Research Highlight: Shared Autonomous Mobility-on-Demand

05 Dec 2020

Here’s a recently published paper by Ivana Dusparic, Enable investigator in Trinity College Dublin.

Paper title: Shared Autonomous Mobility-on-Demand: Learning-based approach and its performance in the presence of traffic congestion
Authors: M. Gueriau, F. Cugurullo, R. Acheampong, I. Dusparic
Journal: IEEE Intelligent Transportation Systems Magazine,

What’s this paper all about?
Ride sharing and car sharing systems consisting of shared autonomous vehicles are expected to improve the efficiency of urban transportation through reduced vehicle ownership, reduced number of trips made, and reduced parking demand. This paper develops a novel reinforcement learning-based decentralized algorithm for matching vehicles with passengers, as well as positioning of vehicles to meet daily fluctuations in demand. Further, it is the first paper to evaluate the efficiency of such a shared mobility system in a full mobility simulation,  investigating the impact that shared vehicles have on the overall congestion by varying the numbers and the ratio of shared vehicles and private vehicles in the simulation.

What exactly have you discovered?
Our proposed algorithm improves the performance of a car and ride sharing system in terms of passenger metrics (a shorter waiting time) and the vehicle/driver metric (an increased vehicle occupancy resulting in increased profits), at the expense of number of requests served. Therefore, potential real-world deployments of ride and car sharing algorithms will require a fine-tuned balance between these metrics that could be achieved by using more sophisticated multi-objective optimization strategies.

The second conclusion arising from the results is that the shared fleet exhibits a similar general behaviour in both the naive simulations and the realistic mobility framework using a real road network and genuine congestion.

However, the impact of traffic congestion is clearly visible. For instance, while a fleet of 200 vehicles was sufficient to serve nearly all requests in the original framework, a significant proportion of requests (roughly 10%) are  now missed by the fleet.

So what?
The implications of our results are twofold. Contrary to the standard in the current literature, this article shows that the evaluation of shared mobility needs to take congestion into account to accurately estimate achievable levels of service. In addition, the results highlight the need for vehicle assignment and rebalancing algorithms to be congestion-aware, in order to be able to finely balance the goals of 3 sets of entities involved in ridesharing systems: passengers/customers, individual vehicles, and fleet operators.

For further information, please contact Ivana.Dusparic (at) scss (dot) tcd (dot) ie