Neural Ruin Recreate

This is the repository accompanying our paper

"Too Big, so Fail? -- Enabling Neural Construction Methods to Solve Large-Scale Routing Problems".

A pre-print of our paper can be found on arxiv.

Environment setup

conda env create -f environment.yml

To install verypy, directly install from the repo:

pip install https://github.com/yorak/VeRyPy/zipball/master

for the "learning to delegate" baseline a different environment is needed, please also see this github issue of the original repo:

conda env create -f l2d_environment.yml

Running NRR

The specific parameters are defined in the config files at config/nrr_config. The different configurations can be specified on the command line. E.g. for a standard run on the CVRP dataset of size 500 using the

• SGBS sub-solver,
• knn construction,
• greedy SG selection,
• initial solution via savings,
• random scoring
• in debug mode

python run_nrr.py meta=debug solver=sgbs constr_mthd=knn select_mthd=greedy init_mthd=savings score_mthd=rnd problem=cvrp500_mixed

An overview can be displayed via

python run_nrr.py -h

Learning the neural scoring function (NSF)

1) First some general dataset has to be generated via the create_cvrp_data notebook.

2) Then, we need to run a solver (e.g. SGBS or LKH3) on some subgraphs of the RR procedure. The respective command to run it for the NRR configuration using

• savings initial solutions,
• sweep based SG selection,
• the SGBS sub-solver
• on uniform CVRP data of problem size 500

  python create_scoring_data.py method=score_savings_sweep_all solver=sgbs problem=cvrp500_unf

  Some predefined configurations for the data scoring can be found in config/nrr_config/method. 3) Next, the NSF has to be trained

  python train_model.py meta=train problem=sgbs_merged_500_unf

  The corresponding configuration can be found in config/nsf_config

Running baseline models

We have a unified interface for all baselines (but the "learning to delegate" aka L2D model), which can be invoked via

python run_baseline.py -m savings -d data/CVRP/benchmark/uchoa/n3.dat --n_seeds 3

where

• -m is the baseline model
• -d is the path to the respective dataset
• --n_seeds is the number of random seeds for which to rerun the baseline

Further flags can be found in the run file run_baseline.py and the registry at baselines/methods_registry.py

For the installation of required packages for some baselines (e.g. LKH3 and HGS) please see the respective readmes in the baselines directory