Implementation of the IEEE TII paper Flexible Job Shop Scheduling via Graph Neural Network and Deep Reinforcement Learning. IEEE Transactions on Industrial Informatics, 2022.

@ARTICLE{9826438,  
   author={Song, Wen and Chen, Xinyang and Li, Qiqiang and Cao, Zhiguang},  
   journal={IEEE Transactions on Industrial Informatics},   
   title={Flexible Job Shop Scheduling via Graph Neural Network and Deep Reinforcement Learning},   
   year={2023},  
   volume={19},  
   number={2},  
   pages={1600-1610},  
   doi={10.1109/TII.2022.3189725}
 }

• python $\ge$ 3.6.13
• pytorch $\ge$ 1.8.1
• gym $\ge$ 0.18.0
• numpy $\ge$ 1.19.5
• pandas $\ge$ 1.1.5
• visdom $\ge$ 0.1.8.9
• matplotlib $\ge$ 3.5.3

Note that pynvml is used in test.py to avoid excessive memory usage of GPU. The code has been modified so that it doesn't call those pynvml relevant functions when using CPU. To allow our program to automatically set up the visdom server and open the browser for you, remember to grant administrator or Internet access privilege to the python interpreter you use. You may use another console and run python -m visdom.server using the virtual environment you created to start the server manually, and hopefully later the Python interpreter will have the privilege to visit the Internet.

For compatibility concerns, we locked the version of the above packages with a suit of capable packages (some are of higher versions than the least requirement) in requirements.yml using conda. You can install them by running:

conda env create -f requirements.yml

This is useful especially as the gym package is keeping updating and the latest version may not be compatible with our code.

After installing the packages, you can activate the environment by running:

conda activate fjsp-drl

• data_dev and data_test are the validation sets and test sets, respectively.
• data saves the instance files generated by ./utils/create_ins.py
• env contains code for the DRL environment
• graph is part of the code related to the graph neural network
• model saves the model for testing
• results saves the trained models
• save is the folder where the experimental results are saved
• utils contains some helper functions
• config.json is the configuration file
• mlp.py is the MLP code (referenced from L2D)
• PPO_model.py contains the implementation of the algorithms in this article, including HGNN and PPO algorithms
• test.py for testing
• train.py for training
• validate.py is used for validation without manual calls

There are various experiments in this article, which are difficult to be covered in a single run. Therefore, please change config.json before running.

Note that disabling the validate_gantt() function in schedule() can improve the efficiency of the program, which is used to check whether the solution is feasible.

python train.py

Note that there should be a validation set of the corresponding size in ./data_dev.

python test.py

Note that there should be model files (*.pt) in ./model.

• https://github.com/zcaicaros/L2D
• https://github.com/yd-kwon/MatNet
• https://github.com/dmlc/dgl/tree/master/examples/pytorch/han