Neural Network Based Large Neighborhood Search Algorithm for Ride Hailing Services

Abstract

Ride Hailing (RH) services have become common in many cities. An important aspect of such services is the optimal matching between vehicles and customer requests, which is very close to the classical Vehicle Routing Problem with Time Windows (VRPTW).

With the emergence of new Machine Learning (ML) techniques, many researches have tried to use them for discreet optimization problems. Recently, Pointer Networks (Ptr-Net) have been applied to simpler Vehicle Routing Problem (VRP) with limited applicability [14]. We add fixed slots to their approach to make it applicable to RH scenario. The number of slots can vary without retraining the network. Furthermore, contrary to reinforcement learning in [14], we use supervise learning for training.

We show that the presented architecture has the potential to build good vehicle routes for RH services. Furthermore, looking at the effectiveness of Large Neighbourhood Search(LNS) for VRPTW, we combine the approach with LNS by using the trained network as an insertion operator. We generate examples from New York Taxi data and use the solutions generated from LNS for training. The approach consistently produces good solutions for problems of sizes similar to the ones used during training, and scales well to unseen problems of relatively bigger sizes.