This is the repository to the corresponding paper "Learning to Control Local Search for Combinatorial Optimization"
- ECML-PKDD 2022: https://2022.ecmlpkdd.org/wp-content/uploads/2022/09/sub_689.pdf
- arXiv (with appendix): https://arxiv.org/abs/2206.13181
We implement a Neural Local Search solver combining a GNN-based learned meta-controller (meta-heuristic) with local search for routing problems (via VRPH) and job shop scheduling problems (custom solver implementation)
In this repo we provide our GNN models and the final checkpoints used in the experiments. Furthermore, the open source code of our custom JSSP solver and the extensions and python bindings for the VRPH C++ solver.
Install the requirements as conda environment
conda env create -f requirements.ymlTo run NeuroLS on JSSP Taillard benchmark instances of size 15x15 for 100 steps
python run_benchmark.py -r run_nls_jssp.py -d data/JSSP/benchmark/TA/ -g 15x15 -p jssp -m nls -e eval_jssp --args "env=jssp15x15_unf" -n 100
To run PDR "FIFO" on JSSP Taillard benchmark instances of size 15x15
python run_benchmark.py -r run_pdr.py -d data/JSSP/benchmark/TA/ -g 15x15 -p jssp -m pdr --args "env=jssp15x15_unf" --policy fifo
To run MetaHeuristic "SA" on JSSP Taillard benchmark instances of size 15x15 for 100 steps
python run_benchmark.py -r run_meta_jssp.py -d data/JSSP/benchmark/TA/ -g 15x15 -p jssp -m meta --args "env=jssp15x15_unf" --policy sa -n 100
To run NeuroLS on CVRP Uchoa benchmark instances group n100 for 200 steps
python run_benchmark.py -r run_nls_rp.py -d data/CVRP/benchmark/uchoa/ -g n100 -p cvrp -m nls -e eval_cvrp --args "env=cvrp100_unf" -n 200
@inproceedings{falkner2022learning,
title={Learning to Control Local Search for Combinatorial Optimization},
author={Falkner, {Jonas K.} and Thyssens, Daniela and Bdeir, Ahmad and Schmidt-Thieme, Lars},
booktitle={Proceedings of the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD)},
year={2022}
}
example rendered episodes on JSSP15x15 and CVRP100
