Reproduce “SSPO: Structure-Aware Similarity-Weighted Preference Optimization for Neural Combinatorial Optimization” 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
Yuanyu Li, Jintao Xu, Zijiang Liu, Yongzhi Qi, Ningxuan Kang, Jianshen Zhang, Wei Qi, Chen Xie and 1 more
A structure-aware, similarity-weighted preference optimization (SSPO) framework enhances Neural Combinatorial Optimization (NCO) policy training by effectively leveraging information from co-sampled solutions, leading to improved solution quality across TSP, EFL, and JSP benchmarks. Developed by researchers from JD.com and Tsinghua University, SSPO achieved a 0% relative gap on Entity Facility Location and reduced optimality gaps on TSP and Job-Shop Scheduling problems.
Key insights
- Reusing the NCO encoder's node representations to construct zero-parameter solution embeddings allows for quantifying structural similarity, providing a rich, underutilized signal for variance reduction.
- A structure-aware, similarity-weighted baseline effectively navigates the trade-off between best-anchor methods (which can polarize gradients) and uniform leave-one-out baselines (which can be redundant), consistently improving NCO policy performance.
- The performance gains from structure-aware weighting are directly correlated with the degree of structural diversity among co-sampled solutions, with greater improvements observed in problems or instances with higher diversity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.