Neural Solver Selection for Combinatorial Optimization
Authors: Chengrui Gao, Haopu Shang, Ke Xue, Chao Qian
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that our framework can effectively distribute instances and the resulting composite solver can achieve significantly better performance (e.g., reduce the optimality gap by 0.88% on TSPLIB and 0.71% on CVRPLIB) than the best individual neural solver with little extra time cost. |
| Researcher Affiliation | Academia | 1National Key Laboratory for Novel Software Technology, Nanjing University, China 2School of Artificial Intelligence, Nanjing University, China. Correspondence to: Chao Qian <EMAIL>. |
| Pseudocode | No | The paper describes methods and architectural components (e.g., 'Hierarchical graph encoder' in Figure 2), but does not contain explicitly labeled 'Pseudocode' or 'Algorithm' blocks with structured steps. |
| Open Source Code | Yes | Our code is available at https://github.com/lamda-bbo/neuralsolver-selection. |
| Open Datasets | Yes | To evaluate the out-of-distribution performance, we utilize two well-known benchmarks with more complex problem distributions and larger problem scales (up to N = 1002): TSPLIB (Reinelt, 1991) and CVRPLIB Set-X (Uchoa et al., 2017). |
| Dataset Splits | Yes | For training, we generate 10,000 TSP and CVRP instances and apply 8-fold instance augmentation (Kwon et al., 2020). For test, a smaller synthetic datasets with 1,000 instances is used. |
| Hardware Specification | No | The paper discusses 'average time' and 'inference time' for various methods but does not provide specific details about the hardware (e.g., CPU, GPU models, or memory specifications) used for running the experiments. |
| Software Dependencies | No | The Adam optimizer (Kingma & Ba, 2015) is employed for training... Re Zero normalization (Bachlechner et al., 2021) for stable convergence of training. These references are to techniques or algorithms, not specific software libraries or platforms with version numbers. |
| Experiment Setup | Yes | The Adam optimizer (Kingma & Ba, 2015) is employed for training, where we set the learning rate to 1e-4 and the weight decay to 1e-6. The number of epochs is set to 50. ... For the top-k strategy, we set k = 2. For the rejection-based strategy, we reject the 20% of instances with the lowest confidence levels ... For the top-p strategy, we set p = 0.5 for TSP and p = 0.8 for CVRP. |