An Attention-Based Approach to the Dynamic Pickup and Delivery Problem

The Dynamic Pickup and Delivery Problem (DPDP) seeks to create dispatch plans for a fleet of vehicles in the face of dynamically revealed orders without complete prior knowledge. In addition to temporal uncertainty, the problem involves operational constraints such as vehicle capacity, precedence between pickup and delivery, unloading rules, and accumulated delays, which makes it difficult to directly use exact methods in continuous decision scenarios. This work proposes an approach based on a Model of Attention trained via Reinforcement Learning and combined with local search. The model was adapted to the DPDP through temporal and operational attributes, feasibility masks and local refinement of the generated routes.