This study analyzes the potential benefits and drawbacks of taxi sharing using agent-based modeling. New York City (NYC) taxis are examined as a case study to evaluate the advantages and disadvantages of ride sharing using both traditional taxis (with shifts) and shared autonomous taxis. Compared to existing studies analyzing ride sharing using NYC taxi data, reserarchers from the Purdue University proposed a model that incorporates individual heterogeneous preferences; compared traditional taxis to autonomous taxis; and examined the spatial change of service coverage due to ride sharing.
The results show that switching from traditional taxis to shared autonomous taxis can potentially reduce the fleet size by 59% while maintaining the service level and without significant increase in wait time for the riders. The benefit of ride sharing is significant with increased occupancy rate (from 1.2 to 3), decreased total travel distance (up to 55%), and reduced carbon emissions (up to 866 metric tonnes per day). Dynamic ride sharing, wich allows shared trips to be formed among many groups of riders, up to the taxi capacity, increases system flexibility. Constraining the sharing to be only between two groups limits the sharing participation to be at the 50–75% level. However, the reduced fleet from ride sharing and autonomous driving may cause taxis to focus on areas of higher demands and lower the service levels in the suburban regions of the city.
Agent-based modeling was used to study ride sharing because of its capability to incorporate individual agent’s different needs and preferences. Additionally, researchers can collect statistics of each rider entering the system and follow every taxi as it moves through the city, which enables in analyzing the data at an individual level. In this study, there are two types of agents: taxis and rider groups. A rider group refers to one or more passengers that are traveling together as a group (organized before the ride sharing, e.g., a family). Each taxi and every rider group has their own parameters in this model. This model was built using the AnyLogic simulation software.