Rethinking Light Decoder-based Solvers for Vehicle Routing Problems
Authors: Ziwei Huang, Jianan Zhou, Zhiguang Cao, Yixin XU
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on cross-size and cross-problem benchmarks, and demonstrate the effectiveness of our method in significantly improving both in-distribution and OOD generalization performance of light decoder-based solvers. Notably, our work narrows the performance gap between light and heavy decoder paradigms, reaffirming the potential of light decoder-based solvers when properly adjusted. |
| Researcher Affiliation | Academia | 1School of Computing and Information Systems, Singapore Management University, Singapore 2College of Computing and Data Science, Nanyang Technological University, Singapore |
| Pseudocode | No | The paper describes the methodology using mathematical equations and textual descriptions, but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at: https://github.com/ziweileonhuang/reld-nco. |
| Open Datasets | Yes | We first evaluate performance on the CVRP using the test dataset introduced in (Luo et al., 2023), which consists of 10,000 instances with 100 nodes, and 128 instances for each problem size of 200, 500, and 1,000 nodes. We also evaluate our Re LD on the well known benchmark dataset CVRPLib Set-X (Uchoa et al., 2017) and Set-XXL. |
| Dataset Splits | Yes | We first evaluate performance on the CVRP using the test dataset introduced in (Luo et al., 2023), which consists of 10,000 instances with 100 nodes, and 128 instances for each problem size of 200, 500, and 1,000 nodes. Re LD employs the same training algorithm as POMO (Kwon et al., 2020), but is trained over 90 epochs, with each epoch comprising 600,000 training instances and a batch size of 120. |
| Hardware Specification | Yes | All experiments were conducted using a single Nvidia A800 GPU with 80GB of memory. |
| Software Dependencies | No | The paper mentions using the Adam optimizer, but does not specify versions for any programming languages, libraries, or other software components. |
| Experiment Setup | Yes | Re LD employs the same training algorithm as POMO (Kwon et al., 2020), but is trained over 90 epochs, with each epoch comprising 600,000 training instances and a batch size of 120. We use the Adam optimizer with an initial learning rate of 1e-4, which is decayed by a factor of 0.1 at the 70th and 80th epochs. Weight decay is set to zero. |