gos-ufsc/sat-gnn

GNNs for Nanosatellite Task Schedulling

This repository implements the experiments described in the paper "Graph Neural Networks for the Offline Nanosatellite Task Scheduling Problem". All experiments were logged at weights & biases. Overall, the interesting bits of code can be found in src/, while notebooks/ contain drafts, experiments and visualizations. The scripts in root are for generating the results we present in the paper.

Steps to reproduce results:

1. Get the data, either by
   • Generating new instances using instance_generation.py and then computing the solutions using compute_solutions.py; or
   • Downloading our publicly available dataset (instances/*.json must go into data/raw/, while solutions/*.npz musto go into data/interim).
2. Maybe save the datasets into hdf5 files using save_datasets.py
   • This is useful if you don't want (or can't) keep the complete dataset in memory
3. Run the feasibility classification experiments using feasibilty_classification_experiments.py
4. Run the optimality prediction experiments
   1. First tune the models using tune_multitarget.py and tune_optimals.py
   2. Then train the best models found using train_best_models.py
5. Evaluate the SatGNN-based heuristics using evaluate.py
   • All of our evaluations are packed in run_evaluation_test.sh and run_heuristic_experiments.sh
6. Generate plots using notebooks/6.0-bmp-paper.plots.ipynb

Other useful scripts: tmp_save_preds.py saves the predictions for a fine-grained analysis; remove_repeated_instances_and_rename.py drops repeated instances when generating on multiple machines (probably outdated); pred_feasibility stores the resuts of the feasibility models; bce_test_set.py computes the test set loss without using the Trainer classes.