Collaboration! Towards Robust Neural Methods for Routing Problems

Authors: Jianan Zhou, Yaoxin Wu, Zhiguang Cao, Wen Song, Jie Zhang, Zhiqi Shen

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments verify the effectiveness and versatility of CNF in defending against various attacks across different neural VRP methods.
Researcher Affiliation Academia Jianan Zhou Nanyang Technological University jianan004@e.ntu.edu.sg Yaoxin Wu Eindhoven University of Technology y.wu2@tue.nl Zhiguang Cao Singapore Management University zgcao@smu.edu.sg Wen Song Shandong University wensong@email.sdu.edu.cn Jie Zhang, Zhiqi Shen Nanyang Technological University {zhangj,zqshen}@ntu.edu.sg
Pseudocode Yes We present the pseudocode of CNF in Algorithm 1, and elaborate each part in the following subsections.
Open Source Code Yes The source code is available at https://github.com/Royal Skye/Routing-CNF.
Open Datasets Yes We further evaluate all neural methods on the real-world benchmark datasets, such as TSPLIB [58] and CVRPLIB [67].
Dataset Splits No We report the average optimality (opt.) gap over 1000 test instances.
Hardware Specification Yes All experiments are conducted on a machine with NVIDIA V100S-PCIE cards and Intel Xeon Gold 6226 CPU at 2.70GHz.
Software Dependencies No No specific version numbers for key software components or libraries (e.g., PyTorch version, Python version) are provided for their own method's implementation.
Experiment Setup Yes Adam optimizer is used with the learning rate of 1e 4, the weight decay of 1e 6 and the batch size of B = 64. To achieve full convergence, we pretrain the model on 300M and 100M clean instances for TSP200 and CVRP200, respectively. After obtaining the pretrained model, we use it to initialize M = 3 models, and further adversarially train them on 5M and 2.5M instances for n = 100 and n = 200, respectively. To save the GPU memory, we reduce the batch size to B = 32 for n = 200. The optimizer setting is the same as the one employed in the pretraining stage, except that the learning rate is decayed by 10 for the last 40% training instances. For neural methods, we use the greedy rollout with x8 instance augmentations following [38].