RL-GNN-for-RCPSP

A program implemented by Pytorch for solving RCPSP and RCPSP with resource disruptions, based on graph neural network and reinforcement learning. Paper is https://doi.org/10.1016/j.rcim.2023.102628

Requirements:

Python >=3.6
Pytorch >= 1.8.0
Be sure that your CUDA is available!

To reproduce experiment results:

check and run test_rcpsp.py to test on RCPSP
check and run test_srcpsp.py to test on RCPSP-RD

To train a new model:

Check Params.py to set hyperparameters
run ppo.py to train a model on RCPSP
run ppo_for_variant.py to train a model on RCPSP-RD

Use Tensorboard to monitor training process

Bottleneck & suggestions

As discussed in the paper, I think the bottleneck of using these GNN-related method to slove scheduling problems lies in the feature extraction capability and receptive field of GNN. In order to improve the performance, I suggest changing the structure of the GNN, using a structure that can effectively extract information from very distant nodes (such as using Mixhop layers?), and then using supervised learning to quickly verify the GNN learning ability.

Specially thanks!

Part of the GNN code comes from L2D https://github.com/zcaicaros/L2D, the method of this paper mainly imitates them.
The code of PPO is modified from CleanRL https://github.com/vwxyzjn/cleanrl
Specially thanks to them!