Reproduce “Combinatorial Optimization by Graph Pointer Networks and Hierarchical Reinforcement Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qiang Ma, Suwen Ge, Danyang He, Darshan Thaker, Iddo Drori
Columbia University researchers developed Graph Pointer Networks (GPN) and Hierarchical Graph Pointer Networks (HGPN) that integrate graph neural networks and hierarchical reinforcement learning to address combinatorial optimization problems. GPN improves performance on TSP by better capturing spatial relationships, while HGPN effectively handles constrained problems like TSP with Time Windows, achieving a higher percentage of feasible solutions.
Key insights
- Representing city relationships with vector context, instead of absolute coordinates, improves model generalization to larger problem instances.
- Integrating graph neural networks into pointer networks better captures the graph structure of combinatorial optimization problems, leading to improved performance.
- Decomposing complex constrained optimization problems into a hierarchy of simpler sub-problems with layer-specific policies can effectively satisfy constraints and optimize objectives.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.