Step 1: Prepare the Data

1. Unzip the training data:

   unzip bpp_dataset.zip
   

   This will create a folder named bpp_dataset.

Step 2: Set Up the Environment

1. Change directory to the bpp folder:

   cd bpp
   

   This folder contains the training code.

2. Set up the Python environment:

   • Use the environment.yml file to create a conda environment that includes all necessary packages:

     conda env create -f environment.yml
     

   • Activate the environment:

     conda activate bpp
     

Step 3: Train the Model

1. Run the training script with the specified parameters:

   python train.py --dim 16 --head 4 --transformer_layers 2 --rnn_layers 1 --seed 1234 --epochs 10000 --batch_size 512 --dataset bpp_dataset --multiply_a 10
   

   This script will create a new folder with a timestamp as its name, containing all run information.

Step 4: Test the Model

1. Copy test.ipynb to the newly created timestamped folder.
2. Open and run the notebook cells in sequence to test the model's prediction accuracy.

Step 5: Serialize the Model

1. Copy to_torchjit.ipynb to the timestamped folder.
2. Open and execute the notebook cells to serialize the model using torch.jit, generating a .pt file.

Integrating ML Model with Neural Column Generation

Step 6: Integrate the ML Model with NCG

1. Move the .pt file generated in Step 5 to the NCG folder.
2. Modify the GlobalParams.json file in the NCG folder. Specifically, change the "machine_learning_model" field to the name of your .pt file.

Step 7: Run the NCG Algorithm

Necessary Files

• GlobalParams.json: Configuration file for the algorithm. The default setting can be used to reproduce the results without any changes.
• bpp_with_augment.pt: The trained ML model in .pt format.
• DLL files required for running the executable 2L-CVRP-ML.exe:
  • pytorch_jni.dll, torch.dll, torch_cpu.dll, fbjni.dll, torch_global_deps.dll, c10.dll, fbgemm.dll, asmjit.dll, cublastLt64_10.dll, cublas64_10.dll, cufft64_10.dll, cusparse64_10.dll, curand64_10.dll, libiomp5md.dll, libiompstubs5md.dll, caffe2_detectron_ops.dll, caffe2_module_test_dynamic.dll, cplex2010.dll.
  • We provide all the required DLL files except for cplex2010.dll. Due to licensing and copyright constraints, cplex2010.dll needs to be obtained separately by the user.

Running the Executable

• The executable 2L-CVRP-ML.exe is used to run the NCG algorithm.
• Download and place the required DLL files in the same directory as the executable.
• For cplex2010.dll, please visit the official website of IBM CPLEX Optimization Studio to download.

Configuring GlobalParams.json

Here are some important fields in GlobalParams.json to experiment with:

1. "ml_activate": true - Set to true to activate the ML model; set to false to deactivate.
2. "L_Trie": true - true activates the L-Trie data structure; false deactivates it.
3. "branch_bound": false - true activates the branch-and-bound tree; false deactivates it, making it essentially the NCG approach as described in the paper.
4. "InfSetCut": false - true to generate infeasible set cuts; false to not generate them.

Note: It is advised to only modify the parameters mentioned above. Changing other parameters may lead to decreased performance or unexpected behavior of the algorithm.

Citation

You can find the arXiv version of the paper here: https://arxiv.org/abs/2406.12454.

🌟 If you find this resource helpful, please consider starting this repository and cite our research:

@inproceedings{ijcai2024p218,
  title     = {A Neural Column Generation Approach to the Vehicle Routing Problem with Two-Dimensional Loading and Last-In-First-Out Constraints},
  author    = {Xia, Yifan and Zhang, Xiangyi},
  booktitle = {Proceedings of the Thirty-Third International Joint Conference on
               Artificial Intelligence, {IJCAI-24}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  editor    = {Kate Larson},
  pages     = {1970--1978},
  year      = {2024},
  month     = {8},
  note      = {Main Track},
  doi       = {10.24963/ijcai.2024/218},
  url       = {https://doi.org/10.24963/ijcai.2024/218},
}