MILP-StuDio: MILP Instance Generation via Block Structure Decomposition
Authors: Haoyang Liu, Jie Wang, Wanbo Zhang, Zijie Geng, Yufei Kuang, Xijun Li, Bin Li, Yongdong Zhang, Feng Wu
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on commonly-used benchmarks demonstrate that with instances generated by MILP-Stu Dio, the learning-based solvers are able to significantly reduce over 10% of the solving time. Experiments demonstrate that MILP-Stu Dio has the following advanced features. |
| Researcher Affiliation | Collaboration | 1Mo E Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China 2 Shanghai Jiao Tong University 3 Noah s Ark Lab, Huawei Technologies |
| Pseudocode | Yes | Algorithm 1: Classification algorithm for constraints and variables in CCMs; Algorithm 2: Partition algorithm for variables in CCMs |
| Open Source Code | No | We implement the model with the code available at https://github.com/sribdcn/Predict-and-Search MILP method to implement PS. (This refers to a baseline, not their own method's code). We will release the code if the paper is accepted. (This is a future promise). |
| Open Datasets | Yes | We consider four MILP problem benchmarks: combinatorial auctions (CA) [16], capacitated facility location (FA) [17], item placement (IP) [32] and workload appointment (WA) [32]. |
| Dataset Splits | Yes | The numbers of training, validation, and testing instances are 100, 20, and 50. |
| Hardware Specification | Yes | We conducted all the experiments on a single machine with NVidia Ge Force GTX 3090 GPUs and Intel(R) Xeon(R) E5-2667 V4CPUs 3.20GHz. |
| Software Dependencies | Yes | This code leverages the state-of-the-art open-source solver SCIP 8.0.3 [6] as the backend solver. |
| Experiment Setup | Yes | In the training process of MILP-Stu Dio, we set the initial learning rate to be 0.001 and the training epoch to be 1000 with early stopping. We list these two parameters used in our experiments in Table 20. |