New Optimization Guidance for Dynamic Dial-a-Ride Problems

Abstract

In the dial-a-ride problem, customers have to be transported from different pickup to drop-off locations. Various constraints such as time windows and a maximum ride time per passenger need to be considered. In the dynamic version of the problem, not all customer requests are known in advance, but arrive during the operation time. Therefore, the maximization of the number of served customers is usually set as the optimization goal. Nevertheless, in the vast majority of known heuristics, the total distance is used to guide the optimization. In this paper, we present different metrics that should enable the evaluation of the insertion potential of future customers and lead to a higher acceptance rate through their use in solution procedures. We show that even a single metric can provide better results than the distances and present a Markov decision process-based approach to enable an agent trained by reinforcement learning to perform even more anticipatory decision making by considering multiple metrics simultaneously.