Towards Massively Parallel Local Search for SAT

Abstract

Parallel portfolio-based algorithms have become a standard methodology for building parallel algorithms for SAT. In this methodology, different algorithms (or the same one with different random seeds) compete to solve a given problem instance. Moreover, the portfolio is usually equipped with cooperation, this way algorithms exchange important knowledge acquired during the search to solve a given problem instance. Portfolio algorithms based on complete solvers exchange learned clauses which are incorporated within each search engine (e.g. ManySAT [1] and plingeling), while those based on incomplete solvers [2] exchange the best assignment for the variables found so far in order to properly craft a new assignment for the variables to restart from. These strategies range from a voting mechanism where each algorithm in the portfolio suggests a value for each variable to probabilistic constructions.