A Data-Driven Forecasting and Solution Approach for the Dial-A-Ride Problem with Time Windows

The Dial-A-Ride Problem (DARP) consists of designing pick-up and delivery routes for a set of customers with special needs. Particularly, it arises in door-to-door transportation services provided to elderly and impaired people. DARP's main objective is to accommodate as many customers' constraints as possible with minimum operation costs. DARP involves realistic precedence and transit time constraints on the pairing of vehicles and customers. This paper proposes a neural network forecasting approach for DARP with time windows (DARPTW). It develops and compares the results of two-layer and a three-layer artificial neural networks (ANN) which forecast demands, service and travel times based on real-life data provided by a transportation company. Experimental results show that three-layer ANN with hyperbolic tangent (tanh) and sigmoid linear unit (selu) activation functions, coupled with a stochastic gradient descent (SGD) optimizer provide the best forecasting results. This paper also develops a data-driven hybrid adaptive large neighborhood search (DD-HALNS). DD-HALNS selects the local search operators according to their updated success' rates, which are, in turn, guided by a learning mechanism from previous successful moves and cost savings. It applies four hybridization features: simulated annealing, tabu lists, genetic crossovers, and restarts. Experimental results on DARPTW benchmark instances highlight DD-HALNS' ability to improve best known routing solutions, while its application on real life instances, from the Canadian city/region of Vancouver, confirms its implementability.