Data Generation

We use a methodology similar to Xu, Hoos, and LeytonBrown (2012) to evaluate the performance of our architecture. First, we used the cnfgen Python library (Lauria et al. 2017) to generate uniform random 3-SAT instances at the phase transition. Following Crawford and Auton (1996), we used m = 4.258n+58.26n 2/3 to estimate the location of the phase transition. We generated 5, 000 SAT instances and 5, 000 UNSAT instances between 100 and 600 variables in steps of 50 variables. In total, we obtained 11, 000 instances in 11 sets. To obtain the prediction targets (satisfiability and satisfying assignments) for each instance, we ran used the lookahead solver march hi (Heule and van Maaren 2009). Although our generation script was limited to a runtime of 24 hours for all instances in a particular set, we repeatedly executed the script until all instances were obtained. For our experiments, we used 61 features from Nudelman et al. (2004). These include 7 problem size features, 29 graphbased features, 13 features related to balance of positive and negative literals, 6 features related to distance to horn formula, and 6 features related to LP relaxation. All instances were generated on a shared compute cluster, on one of 24 nodes equipped with 1 cores, two 2.1 Ghz Intel E52683 v4 Broadwell processors per core, and 16 GB of RAM per core. A memory budget of 15 GB was set for each execution of the generation script.