Engineer-To-Order (ETO) Graph Neural Scheduling (GNS) Project

A Graph Attention Network (GAT) to schedule jobs in an ETO manufacturing environment, trained with a Multi-Agent version of the Proximal Policy Optimization (MAPPO) algorithm.

Refer to this repository in scientific documents

Neumann, Anas (2024). A hyper-graph neural network trained with multi-agent proximal policy optimization to schedule engineer-to-order projects *GitHub repository: https://github.com/AnasNeumann/gns*.

@misc{HGNS,
  author = {Anas Neumann},
  title = {A hyper-graph neural network trained with multi-Agent proximal policy optimization to schedule engineer-to-order projects},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/AnasNeumann/gns}},
  commit = {main}
}

Locally try the project

1. python -m venv ./gns_env
2. source ./gns_env/bin/activate
3. pip install --upgrade -r requirements.txt
4. CHOOSE EITHER GNS_SOLVER, EXACT_SOLVER, INSTANCE_GENERATOR, or RESULTS_ANALYS (see bellow for the rest)
5. desactivate

Training process: MAPPO

Test the instance generator

bash _env.sh
python instance_generator.py --train=150 --test=50

Test the exact sat solver (Google OR-Tools)

bash _env.sh
python exact_solver.py --size=s --id=151 --path=./ --mode=test --time=1

Test the GNS solver

bash _env.sh

1. Pre-train on several instances (using MAPPO)

python gns_solver.py --train=true --target=false --mode=test --path=./ --number=1 

2. Solve on instance (inference mode only)

python gns_solver.py --train=false --target=true --size=s --id=151 --mode=test --path=./ --number=1

3. Solve all test instances (inference mode only)

python python gns_solver.py --train=false --target=false --mode=test --path=./ --number=1

4. Fine-tune on target instance (using MAPPO)

python gns_solver.py --train=true --target=true --size=s --id=151 --mode=prod --use_pretrain=true --interactive=false --path=./ --number=1
python gns_solver.py --train=true --target=true --size=s --id=151 --mode=prod --use_pretrain=false --interactive=true --path=./ --number=1

Generate exact jobs for DRAC production

• DRAC = Digital Research Alliance of Canada

# job duration, number of CPUs, and memory used are dynamically related to the instance size (no GPU/TPU for exact jobs)
python ./jobs/exact_builder.py --account=x --parent=y --mail=x@mail.com

Generate the GNS training job for DRAC production

python ./jobs/gns_builder.py --account=x --parent=y --mail=x@mail.com --time=32 --memory=16 --cpu=1 --number=1

Execute all jobs in DRAC production (train GNN and solve testing instances with exact solver)

cd jobs/scripts/
bash 0_run_purge.sh
bash 1_run_all.sh train exact_s exact_m exact_l exact_xl exact_xxl exact_xxxl

Locally analyze the final results with (no hard resources required):

The main results we analyze are:

• Solution Quality: deviation between GNS and EPSIII per instance size
• Scalability: computing time and memory per instance size and solution type (GNS versus EPSIII)
• Convergence: value loss over iterations as well as policy loss and entroy per agent and over time [GNS only]

python results_analysis.py --path=./ --last=9