Reproduce “Inverse Optimization for Routing Problems” 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
Pedro Zattoni Scroccaro, Piet van Beek, Peyman Mohajerin Esfahani, Bilge Atasoy
Researchers at Delft University of Technology developed a tailored Inverse Optimization framework to learn complex human driver routing preferences from historical data, achieving a 2nd place ranking in the Amazon Last Mile Routing Research Challenge with a score of 0.0302. This framework adapts the Augmented Suboptimality Loss for binary decision variables and uses a stochastic first-order algorithm for efficient learning across various routing problem types.
Key insights
- Reformulating the Augmented Suboptimality Loss for binary decision variables yields a convex inner minimization problem, which is critical for practical computability in combinatorial routing.
- Combining the learned cost function with affine penalty terms that capture observed geographical routing patterns (e.g., zone sequencing) is essential for high fidelity in replicating human expert behavior.
- Utilizing a stochastic first-order algorithm with reshuffled sampling proves more effective than uniform sampling or cutting plane methods for learning routing preferences from large datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.