Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
A Learning-based Iterative Method for Solving Vehicle Routing Problems
Authors: Hao Lu, Xingwen Zhang, Shuang Yang
ICLR 2020 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present our experiment results. First we introduce a detailed setup of CVRP and hyper-parameters used. Then we compare our performance with prior neural network approaches, i.e., Nazari et al. (2018), Kool et al. (2019), Chen & Tian (2019), and a classic state-of-the-art heuristic algorithm. |
| Researcher Affiliation | Collaboration | Hao Lu Princeton University, Princeton, NJ 08540 {haolu}@princeton.edu Xingwen Zhang & Shuang Yang Ant Financial Services Group, San Mateo, CA 94402 EMAIL |
| Pseudocode | No | The paper describes the proposed method in text and with diagrams (e.g., Figure 2, Figure 6) but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | Code will be available at github.com/rlopt/l2i upon approval of the company. |
| Open Datasets | Yes | Following the same protocol of data generation as in Uchoa et al. (2017) (the authors of the paper were the creators and owners of the CVRPLib website4), we tested our method for four additional scenarios, namely, central depot positioning, eccentric depot positioning, clustered customer positioning, and random-clustered customer positioning. (...) 4http://vrp.galgos.inf.puc-rio.br/index.php/en/ |
| Dataset Splits | No | The paper describes how problem instances are generated and used for training and testing, but it does not specify a separate validation set or explicit train/validation/test dataset splits for reproducibility. |
| Hardware Specification | Yes | Our method was implemented in Python, and the experiments were run computer nodes, each with a single Nvidia Tesla T4 GPU. |
| Software Dependencies | No | The paper states that the method was implemented in Python and uses ADAM for training, but it does not provide specific version numbers for Python, ADAM's underlying library (e.g., PyTorch/TensorFlow), or any other software dependencies. |
| Experiment Setup | Yes | Setup and hyper-parameters. We follow the same settings as previous works (Nazari et al., 2018; Kool et al., 2019; Chen & Tian, 2019) for CVRP. We consider three sub-problems with number of customers N = 20, 50, 100, respectively. (...) The capacity of a vehicle is 20, 30, 40 for N = 20, 50, 100, respectively. After L = 6 consecutive step of no improvement, we perturb the solution. To train the policy network, we use ADAM with a learning rate of 0.001. Unless otherwise stated, for a problem instance and a given policy we randomly initiate a feasible solution, and then iteratively update the solution T = 40000 times following the policy. |