On Continuous Local BDD-Based Search for Hybrid SAT Solving

Authors: Anastasios Kyrillidis, Moshe Vardi, Zhiwei Zhang3841-3850

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We study the capabilities and limitations of our versatile CLS solver, Grad SAT, by applying it on many benchmark instances. The experimental results indicate that Grad SAT can be a useful addition to the portfolio of existing SAT and Max SAT solvers for solving Boolean satisfiability and optimization problems.
Researcher Affiliation Academia Anastasios Kyrillidis, Moshe Vardi, Zhiwei Zhang Rice University, 6100 Main Street, Houston, TX, USA {anastasios, vardi, zhiwei}@rice.edu
Pseudocode Yes Algorithm 1: Fourier SAT, a CLS-based SAT Solver. ... Algorithm 2: The Probability-Assignment Algorithm ... Algorithm 3: Objective Evaluation (Top-Down) ... Algorithm 4: Objective Evaluation (Bottom-Up) ... Algorithm 5: Efficient Gradient Computation ... Algorithm 6: Grad SAT
Open Source Code Yes We implement Grad SAT in C++ 9. Available at https://github.com/vardigroup/Grad SAT
Open Datasets Yes Benchmark 2: Max SAT problems. We gathered 570 instances from the crafted (198 instances) and random track (377 instances) of Max SAT Competition 2016-2019.
Dataset Splits No No specific details about training, validation, or test dataset splits (e.g., percentages, sample counts, or methodology like cross-validation) were provided for the benchmarks used.
Hardware Specification Yes Each experiment is run on an exclusive node in a Linux cluster with 16-processor cores at 2.63 GHz and 1 GB of RAM per processor.
Software Dependencies Yes Thus practical packages for handling BDDs, e.g., CUDD 2 have been developed. Such packages provide sub-function sharing between individual BDDs so that a set of BDDs can be stored compactly. ... 2CUDD: CU Decision Diagram Package Release 3.0.0
Experiment Setup Yes Time limit is set to be 100s. ... We set the time limit to 60s. ... We tuned r and T by running Grad SAT on random 3-CNF benchmarks with r {1, 1.5, 2, 2.5, 3} and T {1, 2, 4, 8, 16} and choosing (r, T) that solves most cases. We got r = 2, T = 8. ... The constraint weights are initialized to be the length of the constraint. When a local optimum is reached, we multiply the weights of unsatisfied constraints by r and start from the same point once again. Grad SAT only does random restart when T trials have been made for the same starting point, as shown in Alg. 6.