Learn2Pool: efficient and effective ride assignment in ride sharing systems

Ryan Cheng, Selvakumar Jayaraman, Robert Fitzgerald, Farnoush Banaei Kashani

Abstract: Ride sharing systems enable people with similar itineraries and pick-up times to share their rides on the same vehicle. This has significant societal benefits as it reduces the number of vehicles used and hence reduces energy consumption and emissions to the environment. However, current systems do not fully address the potential of ride sharing as they do not utilize the transportation network effectively. A change in the ride assignment/scheduling algorithm to batch multiple requests has proven to be effective when extended time utilization is considered. Current batching based assignment methods generate all possible combinations of the trips for a given set of requests and vehicles, and then solve the assignment problem using integer linear programming (ILP). This becomes more computationally expensive as the combinatorial search space grows. In this paper, we propose overcoming this scalability problem by learning the reverse nearest neighbors order for a given location using pointer networks to effectively reduce the number of candidate trips considered in ILP. With extensive experimentation using real data, we show that our proposed solution, termed Learn2Pool, offers the most practical solution for ride assignment by allowing striking a balance between efficacy and efficiency, while demonstrating superior efficiency and efficacy across all existing methods.