[ICLR 2025] Graph Assisted Offline-Online Deep Reinforcement Learning for Dynamic Workflow Scheduling

This repository contains the implementation of Graph Assisted Offline-Online Deep Reinforcement Learning (GOODRL) for Dynamic Workflow Scheduling (DWS), which addresses the dynamic workflow scheduling problem. The paper introduces three key innovations in dynamic graph representations, decoupled Actor-Critic encoders for RL stability, and training methods for unpredictable changes.

For reproducible evaluation of Dynamic Workflow Scheduling methods, see DWS-Testbed. It provides a cleaned test environment with EST, PEFT, HEFT, GPHH, ERL-DWS, and GOODRL evaluators, bundled pretrained artifacts, deterministic workflow-instance generation, seed-42 reference results, and a portable SLURM submission example. For a deadline-constrained DWS environment variant that uses total cost as its objective, see Simulator-DCDWSC.

If you find GOODRL helpful for your research or applied projects:

@InProceedings{yang2025graph,
    title={Graph Assisted Offline-Online Deep Reinforcement Learning for Dynamic Workflow Scheduling},
    author={Yang, Yifan and Chen, Gang and Ma, Hui and Zhang, Cong and Cao, Zhiguang and Zhang, Mengjie},
    booktitle={International Conference on Learning Representations},
    year={2025}
}

• Python >= 3.11.5
• PyTorch >= 2.4.1
• PyTorch Geometric >= 2.5.3

Install dependencies:

pip install --upgrade pip
pip install rl-zoo3 deap torch_geometric gym joblib openpyxl

Note: The current implementation is CPU-based. GPU adaptations may require modifications. Multi-CPU parallel execution is supported by adjusting the following:

Non-parallel version:

meanFlowTimes = []
for t in range(configs.valid_num):
    meanFlowTime = validation_H(t, configs)
    meanFlowTimes.append(meanFlowTime)

Parallel version:

meanFlowTimes = Parallel(n_jobs=-1)(delayed(validation_H)(t, configs) for t in range(configs.valid_num))

Run the following command for offline imitation learning:

python step1.py --vm_types 6 --each_vm_type_num 4 --arr_rate 5.4 --lr_a 0.0001 --log_interval 1 --max_updates 10

• Execute multiple independent runs and select the best-performing actor for the next stage: offline PPO.
• The trained actors from this stage can be found in ./validation_data/step1.

Run the following command for offline PPO training:

python step2.py --vm_types 6 --each_vm_type_num 4 --arr_rate 5.4 --lr_a 0.0003 --lr_c 0.001 --warmup_critic 200

• Execute multiple independent runs and select the best-performing actor for offline testing and as the initialization for the online PPO phase.
• The trained actor and critic models from this stage are saved in ./validation_data/step2.

Run the following command for online PPO training:

python step3.py --vm_types 6 --each_vm_type_num 4 --arr_rate 9 --online_start_ac 5_5_5.4 --wf_num 10000 --max_updates 500 --warmup_steps 50000 --lr_a 0.00005 --lr_c 0.0001 --n_epochs 5

• In each run, the final result represents the online performance after processing wf_num workflows.

Additionally:

• mainGP.py is the training script for GPHH.
• mainESRL.py is the training script for ESRL.

The code and documentation will be continuously updated, including multi-objective versions and applications in FJSS environments. For any inquiries, feel free to contact us via 💌.

The implementation references the following works:

• L2D: Learning to Dispatch for Job Shop Scheduling via Deep Reinforcement Learning
• FJSP: Flexible Job Shop Scheduling via Graph Neural Network and Deep Reinforcement Learning
• L2S: Deep Reinforcement Learning Guided Improvement Heuristic for Job Shop Scheduling