Pareto Set Learning for Neural Multi-Objective Combinatorial Optimization
Authors: Xi Lin, Zhiyuan Yang, Qingfu Zhang
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that our proposed method significantly outperforms some other methods on the multiobjective traveling salesman problem, multiobjective vehicle routing problem, and multiobjective knapsack problem in terms of solution quality, speed, and model efficiency. |
| Researcher Affiliation | Academia | Xi Lin, Zhiyuan Yang, Qingfu Zhang Department of Computer Science, City University of Hong Kong xi.lin@my.cityu.edu.hk |
| Pseudocode | Yes | Algorithm 1 Neural MOCO Training |
| Open Source Code | Yes | The source code can be found in https://github.com/Xi-L/PMOCO. |
| Open Datasets | No | We randomly generate 100, 000 problem instances on the fly for each epoch, and train the model for 200 epochs. |
| Dataset Splits | No | The paper describes generating instances 'on the fly' for training and evaluating on '200 random test instances', but does not explicitly detail train/validation/test dataset splits with percentages, counts, or predefined splits for reproducibility. |
| Hardware Specification | Yes | We train our models on a single RTX 2080-Ti GPU, and it costs about 10 minutes for an epoch on MOTSP100. |
| Software Dependencies | No | The paper mentions using 'ADAM' as the optimizer and refers to specific model architectures, but does not provide specific version numbers for software dependencies like programming languages or libraries (e.g., Python, PyTorch versions). |
| Experiment Setup | Yes | For all problems, we train our proposed model for 200 epochs, with 100, 000 problem instances randomly generated on the fly at each epoch. At each iteration step, we need to sample K preferences, B problem instances, and N tours to calculate the policy gradient. We set K B = 64 to make the batch of 64 instances for training a single AM model, and let N equal to the problem size (e.g., the number of nodes) as in Kwon et al. (2020). |