Multi-Decoder Attention Model with Embedding Glimpse for Solving Vehicle Routing Problems
Authors: Liang Xin, Wen Song, Zhiguang Cao, Jie Zhang12042-12049
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on six different routing problems show that our method significantly outperforms the state-of-the-art deep learning based models. |
| Researcher Affiliation | Academia | 1Nanyang Technological University, Singapore 2Shandong University, China 3National University of Singapore, Singapore |
| Pseudocode | No | The paper includes a figure showing the model architecture (Figure 1) and mathematical equations for model components, but no explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code has been released 1. (Footnote 1: https://github.com/liangxinedu/MDAM) |
| Open Datasets | No | For TSP and CVRP, instances with 20, 50 and 100 nodes (cities) are generated by sampling coordinates from the uniform distribution. For other problems, the settings follow existing work (Kool, van Hoof, and Welling 2019a). The paper describes how data instances are generated but does not provide a link or specific access information for the generated datasets themselves, or state that they are publicly available for download. |
| Dataset Splits | No | The paper mentions training for 100 epochs but does not specify a separate validation dataset or split used for hyperparameter tuning or early stopping. |
| Hardware Specification | Yes | We evaluate our model using a single RTX-2080Ti GPU. We run Concorde and LKH in parallel for 32 instances on a 32 virtual CPU system (2Xeon E5-2620). |
| Software Dependencies | No | The paper mentions general software components and algorithms like "Transformer Model", "Adam Optimizer", and "REINFORCE algorithm" but does not specify their version numbers or the versions of any underlying programming languages or libraries (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | We embed the nodes with element-wise projection to 128-dimensional vectors. The Transformer encoder has 3 layers with 128 dimension features and 8 attention heads... hidden dimension... is 512. We choose the number of decoders in MDAM to be 5... We set the number of steps between re-embeddings to be 2, 4, 8 for TSP20, 50, 100 and 2, 6, 8 for CVRP20, 50, 100... We train the model with 2,500 iterations per epoch and batch size 512 (except 256 for CVRP100)... for 100 epochs. We use Adam Optimizer... with learning rate 10 4... The coefficient of KL loss k KL needs to be large enough... We set k KL to 0.01 based on experiments on TSP20. |