Data-driven fleet load balancing strategies for shared Mobility-on-Demand systems
MetadataShow full item record
Mobility on Demand (MoD) systems utilize shared vehicles to supplement or replace mass transit and private vehicles. Such systems include traditional taxis as well as Transportation Network Companies (TNCs) that offer bike and ride sharing. MoD systems face myriad operational challenges, but this dissertation focuses on the data-driven load balancing problem of redistributing vehicles among service regions. This is a difficult resource reallocation problem because customer demands follow a stochastic process subject to dynamic temporal-spatial patterns. The first half of this dissertation considers the load balancing problem for a bike sharing system in which bikes are redistributed among stations via trucks. The objective is to avoid situations in which a user wishes to rent (return) a bike to a station but cannot because the station is empty (full). First, a station and interval-specific inventory level is defined as a function of station capacity and interval demand rates as observed from analyzed data. Second, using a graph network framework, a receding horizon controller is proposed to determine the optimal paths -- over a short period of time -- for the fleet of trucks to take. When calculating the optimal paths the controller considers the current and projected inventory subject to the dynamically changing rent and return rates for every station in the network. The second half of this dissertation tackles the redistribution of an autonomous taxi fleet in which the vehicles themselves are capable of performing load balancing operations across service regions. The objective is to minimize the fraction of customers whose demands are dropped due to vehicle unavailability as well as the fraction of time the vehicles spend on load balancing operations (i.e driving empty). The system is represented by a queuing model and, as such, dynamic programming can find the optimal solution; however, the state-space of the model grows quickly rendering all but a minuscule system impossible to solve. To this end a parametric control is proposed that uses thresholds to dictate redistribution actions and well performing parameters are found via concurrent estimation methods of simulation.