From Comparison to Selection: Enhancing Cross-Scale Generalization of Pointer Networks Trained with Reinforcement Learning for Capacitated Drone Delivery Routing

TL;DR

The paper improves pointer networks for drone delivery routing by addressing cross-scale generalization issues in reinforcement learning. It introduces a permutation-invariant graph-context aggregation method and an approach selector that adapts model choice based on problem size, significantly reducing routing costs across varying scales.

Why It Matters

Efficient, scalable routing is critical for real-world drone delivery systems. This work enables a single trained model to perform well across diverse client counts without retraining, improving deployment flexibility and operational efficiency.

Key Findings

• Node density shifts, attention dispersion, and diluted graph context degrade pointer network performance when scaling problem sizes.
• Entropy-based attention scaling improves performance when transferring from small to large problem sizes, while graph-context aggregation enhances transfer from large to small.
• An approach selector that dynamically chooses the best model variant achieves up to 3.2% cost reduction across seen and unseen problem sizes.