NeuroCore
We train a simplified NeuroSAT architecture to directly predict the unsatisfiable cores of real problems, and modify MiniSat to periodically replace its variable activity scores with NeuroSAT's prediction of how likely they are to appear in an unsatisfiable core. Our modified MiniSat solves 10% more problems on SAT-COMP 2018 than the original does. Although MiniSat is no longer considered a state-of-the-art solver, our results nonetheless demonstrate the potential for NeuroSAT (and in particular, NeuroCore) to provide useful guidance to CDCL solvers on real problems.
More information can be found in the paper https://arxiv.org/abs/1903.04671.
Overview of repository
solver/
: pybind11 module wrapping Z3, for generating training data.python/gen_data.py
: ray script to generate training data.python/neurosat.py
: simplified reimplementation of NeuroSAT.python/train.py
: distributed tensorflow script to train NeuroCore.hybrids/minisat/
: pybind11 module for a MiniSat extension that periodically queries NeuroCore.python/evaluate_cdcl.py
: ray script for evaluating solvers on benchmarks.weights/
: the trained weights used for the experiments in the paper.
Dependencies
There are several dependencies. For now, we list the ones that cannot be easily installed with apt-get
or pip3
:
- z3 (recent commit)
- drat-trim
- eigen
- pybind11
- minisat
The code also uses a MariaDB database and an Azure storage account. Replace the fake credentials in auth.json
.
Team
- Daniel Selsam, Microsoft Research, Stanford University
- Nikolaj Bjørner, Microsoft Research