Learning to Solve Grouped 2D Bin Packing Problems in Manufacturing Industry

This repo contains the source code and datasets for our KDD'23 paper.

Please cite us as:

@inproceedings{Ao2023,
  doi = {10.1145/3580305.3599860},
  url = {https://doi.org/10.1145/3580305.3599860},
  year = {2023},
  month = aug,
  publisher = {{ACM}},
  author = {Wenxuan Ao and Guozhen Zhang and Yong Li and Depeng Jin},
  title = {Learning to Solve Grouped 2D Bin Packing Problems in the Manufacturing Industry},
  booktitle = {Proceedings of the 29th {ACM} {SIGKDD} Conference on Knowledge Discovery and Data Mining}
}

Requirements

python>=3.8
einops==0.6.0
numpy==1.21.2
ray==1.10.0
scipy==1.7.3
setproctitle==1.2.3
torch==1.13.0
tqdm==4.62.3

And install pybp by running

pip install git+https://github.com/John-Ao/pybp

Datasets

The training set and testing set of the synthetic datasets G200 and G100 are in the ./dataset folder. Due to privacy concerns, the Real dataset is not included. Instead, the synthetic dataset G200 can be used, which is generated with statistics similar to Real.

Let's recap the key definitions of the Grouped 2D bin packing problem are as follows:

Bin: we refer to the raw material sheets as bins, which are rectangular and all of the same size. We denote the width as $W$ and height as $H$.
Item: we refer to the parts or panels that need to be cut from bins as items, which are also rectangular but with different sizes.
Order: an customer order contains a list of items. The items in the same order must be manufactured together.
Group: a group is a list of orders. The group is the basic unit of manufacturing. As the items need to be sorted out for the assembly of each order after manufacturing, there is a limit on the maximum number of items a group contains.
Bin packing plan: it is a way to arrange and fit the given list of items on a list of bins such that the edges of the items are parallel to the edges of the bins and the items do not overlap.
Grouping plan: it is a way to partition the given list of orders into several groups.

In the datasets, the dimensions of the bin are $W=2440mm$ and $H=1220mm$. For G200, the maximum number of items in a group is $200$ and for G100 it is $100$.

The datasets are provided as 3-dimensional lists. The first dimension is day, the second is customer order, the third is item (represented as width x height), as shown in the following example:

[
  # Day 0
  [
    # Customer order 0
    [
      [1,1], # item 0, size is 1x1
      [2,2], # item 1, size is 2x2
    ],
    # Customer order 1
    [
      [3,5], # item 2, size is 3x5
      [6,7], # item 3, size is 6x7
      [8,9], # item 4, size is 8x9
    ],
    ...
  ],
  # Day 1
  [
    ...
  ],
  ...
]

Baselines

The baseline methods can be run with run_baseline.py.

For example, to test GGA-H on the first instance of dataset G100, use

python run_baseline.py --upper GGA --lower height --dataset G100 --instance 0

Our method

To train the lower agent on G200 and G100, use

python train_lower.py --dataset G200 --embed-dim 32 --num-heads 4 --pomo 10 --cuda 0
python train_lower.py --dataset G100 --embed-dim 32 --num-heads 4 --pomo 10 --cuda 0

To train the upper agent on G200 and G100, use

python train_upper.py --dataset G200 --lower RL --embed-dim 32 --cuda 0
python train_upper.py --dataset G100 --lower RL --embed-dim 32 --cuda 0

We provide checkpoints for pre-trained models on G200 and G100. To test our method on the first 10 test instances of dataset G100 and G200, use

python run_ours.py --dataset G100 --lower RL --steps 1000 --cuda 0
python run_ours.py --dataset G200 --lower RL --steps 1000 --cuda 0

Demo

To see our method in action, please refer to demo.ipynb, where we demonstrate how to generate the grouping plan and bin packing plans.

tsinghua-fib-lab / G2DBP