| Abstract | Trailer repositioning in drayage operations is a crucial element in the efficient transportation of goods between global commerce and local communities. Prior research has made two main assumptions to simplify the problem which are the availability of empty trailers at each yard and predetermined movements for empty trailers. In contrast, this paper proposes a demand-dependent trailer repositioning framework that considers delay penalties and limited available trailers while also addressing ad-hoc changes in orders. To address this problem, we formulate a novel drayage operation pickup and delivery framework (Dray-Q) that optimizes the just-in-time movement of empty trailers in response to the prevailing demand. The main objective is to avoid scenarios where there is either a surplus of empty trailers or a deficit, ensuring that the provisioning of trailers in each yard is both timely and adequate. In order to train the agent, our framework utilizes the advanced Reinforcement Learning algorithm, Rainbow-DQN, to learn efficient the real-time trailer repositioning policy. We introduce a multi-objective reward function that balances empty trailer supply and demand, minimizes delays, and considers customer priorities in the dispatching of empty trailers. Dray-Q is flexible and adaptable to changes in customer order settings and can be scaled to accommodate different combinations of order, trailer, and yard sizes. Experimental results demonstrate that Dray-Q outperforms well-known baseline methods in the literature and can be implemented at a production level with exceptional performance and generalizability. |
|---|
