A machine-learning framework for a novel 3-step approach for real-time taxi dispatching

Abstract

In the status quo, taxi dispatching is not fully optimized. Low taxi capacity utilization rates along with high passenger wait times suggest inefficiency with dispatching. Computationally, the taxi dispatch problem (TDP) faces key constraints: the problem is very dynamic with information about trips unknown beforehand and must be computed in real-time. These constraints force quick, simple, intuitive, but inefficient solutions like local greedy approaches to be applied. This research study presents a novel solution for TDP. Through TaxiNet, future taxi demand is predicted in four components: the number of passengers picked up, the number of passengers dropped off, the distribution of passengers picked up and the distribution of passengers dropped off for a 15-minute time-step. The predicted demand is inputted into a proposed Monte Carlo algorithm which can link the pickup demand with the drop-off demand to generate a series of predicted trips that will occur shortly. Not only do these predictions allow for a clearer idea of where passengers and taxis will be in the future, but it also extends the window of computation time provided to calculate and find optimal dispatching solutions. A proposed ACO algorithm inputs in the predicted passenger and taxi locations and outputs an optimal dispatching solution. Simulations were run to compare the performance of a taxi fleet operating under existing systems versus the developed algorithm. The results show that the algorithm increased fleet profitability and lowered passenger wait times.