Learning to Handle Complex Constraints for Vehicle Routing Problems

Authors: Jieyi Bi, Yining Ma, Jianan Zhou, Wen Song, Zhiguang Cao, Yaoxin Wu, Jie Zhang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To verify our PIP designs, we conduct extensive experiments on the highly challenging Traveling Salesman Problem with Time Window (TSPTW), and TSP with Draft Limit (TSPDL) variants under different constraint hardness levels.
Researcher Affiliation Academia 1Nanyang Technological University 2Shandong University 3Singapore Management University 4Eindhoven University of Technology
Pseudocode No No explicitly labeled 'Pseudocode' or 'Algorithm' blocks were found in the paper.
Open Source Code Yes Our implementation in PyTorch are publicly available at https://github.com/jieyibi/PIP-constraint.
Open Datasets Yes For TSPTW [7, 9, 30, 67], we generate three types of instances: Easy, Medium and Hard, by adjusting the width and overlap of the time window. For TSPDL [68 70], we consider two levels of hardness: Medium and Hard.
Dataset Splits No While the paper describes instance generation and evaluation on test instances, it does not explicitly provide details on training/test/validation dataset splits with percentages or counts for a fixed dataset. The training is conducted using reinforcement learning on generated instances, and evaluation is performed on separate generated test instances.
Hardware Specification Yes All the experiments are conducted on servers with NVIDIA Ge Force RTX 3090 GPUs and Intel(R) Xeon(R) Gold 6326 CPU at 2.90GHz.
Software Dependencies No The paper mentions 'Our implementation in PyTorch', but it does not specify a version number for PyTorch or any other key software dependencies.
Experiment Setup Yes Hyper-parameters for training follow the original settings of the backbone models except for the ones related to the added PIP decoder. Pertaining to the PIP-D model (e.g., AM*+PIP-D), we employ an auxiliary decoder (i.e., PIP decoder) that is trained with the ground-truth PIP labels in a supervised manner. To balance the trade-off between training efficiency and empirical performance, we update it periodically. Specifically, within E total training epochs, the PIP decoder is first trained with Einit epochs, and then periodically updated Eu epochs per Ep epochs. To boost the performance, we switch Eu to El for the final El epochs. The detailed settings are presented in Table 9. ... We set the Lagrangian multiplier λ to 1 in the main experiments