Learning Collaborative Policies to Solve NP-hard Routing Problems
Authors: Minsu Kim, Jinkyoo Park, joungho kim
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments This section reports the experimental results of the LCP scheme on TSP, PCTSP, and CVRP (N = 20, 50, 100, 500, N: number of nodes). Also, we report several ablation studies in section 5.3 and Appendix B-F. |
| Researcher Affiliation | Academia | Minsu Kim Jinkyoo Park Joungho Kim Korea Advanced Institute of Science and Technology (KAIST) School of Electrical Engineering, Dept. Industrial & Systems Engineering {min-su, jinkyoo.park, joungho}@kaist.ac.kr |
| Pseudocode | Yes | See pseudo-code in Appendix A.4 for a detailed technical explanation. |
| Open Source Code | Yes | 2See source code in https://github.com/alstn12088/LCP |
| Open Datasets | Yes | Extensive experiments demonstrate that the proposed two-policies collaboration scheme improves over single-policy DRL framework on various NP-hard routing problems, including TSP, prize collecting TSP (PCTSP), and capacitated vehicle routing problem (CVRP). ... The detailed descriptions for these problems are in Appendix A.1. ... We evaluate performance on real-world TSPs in the TSPLIB [20]. |
| Dataset Splits | No | The paper mentions 'dataset configuration' in Appendix A.5 but does not provide explicit training, validation, or test split percentages or sample counts in the main body. |
| Hardware Specification | No | The paper mentions running experiments 'in our machine' but does not provide specific hardware details such as CPU/GPU models or memory specifications. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies (e.g., Python, PyTorch, TensorFlow, or specific libraries and solvers). |
| Experiment Setup | Yes | Throughout the entire training process of the seeder and reviser, we have exactly the same hyperparameters as Kool et al. [12], except that the training batch size of our seeder is 1024. To train the seeder s policy, we set α = 0.5 (2) and Nw = PN i=1 i for linear weight wt = N t Nw for entropy scheduling. Details in the experimental setting, including hyperparameters, dataset configuration, and run time evaluation, are described in Appendix A.5. |