Learning local search heuristics for SAT

Code accompanying the NeurIPS 2019 paper Learning local search heuristics for Boolean satisfiability.

Setup

Make sure to have minisat in your path. It is available here: http://minisat.se/MiniSat.html

Run the following sequence of commands to clone the repository and create the required conda environments.

git clone --recursive https://github.com/emreyolcu/sat.git
cd sat
conda env create -f sat.yaml
conda env create -f cnfgen.yaml
(cd code/PyMiniSolvers && make)

Instructions

Data generation

scripts directory includes bash scripts to generate the data used for training. As an example, to generate the coloring formulas, run the following commands.

conda activate cnfgen
scripts/generate_kcolor.sh

When generating the SR formulas, activate the sat environment instead of cnfgen before running the script.

Training

configs/search directory includes training configurations for each set of formulas. Run the following commands to train models on sets of coloring formulas of increasing size.

conda activate sat
python code/train_search.py --config_path configs/search/kcolor/5.yaml
python code/train_search.py --config_path configs/search/kcolor/10.yaml
python code/train_search.py --config_path configs/search/kcolor/15.yaml
python code/train_search.py --config_path configs/search/kcolor/20.yaml

If a configuration file has the line eval_multi: True, the training script will create 25 processes to perform evaluation during training. Remove this line or change the value to False if you prefer to use a single process instead.

When attempting to use multiple processes you may see the error Too many open files. This can be avoided by executing ulimit -n 4096 (or another large value that your system allows) before starting training.

Evaluation

After training is completed, run the following commands to evaluate a learned model and WalkSAT on a set of formulas.

conda activate sat
python code/evaluate.py --dir_path data/kcolor/3-10-0.5 --samples 5 --model_path results/search/kcolor/5/model_best.pth --no_multi
python code/walksat.py --dir_path data/kcolor/3-10-0.5 --samples 5

Arguments used:

• --dir_path: Directory containing the formulas.
• --samples: Number of formulas to run the solver on.
• --model_path: Saved parameters of the graph neural network to evaluate.
• --no_multi: Whether to use a single process for evaluation. Without this argument, the script creates 25 processes.