The variable branching heuristics used in the most recent and most effective SAT solvers, including zChaff and BerkMin, can be viewed as consisting of a simple mechanism for rewarding the variables participating in conflicts during the search process. In this paper we propose to extend the simple rewarding mechanism used in zChaff and BerkMin, and develop different rewarding mechanisms based on information provided by the search algorithm, namely the size of learned clauses and the size of the backjumps in the search tree. The results show that very significant gains can be obtained for real-world problem instances of SAT.
Elsa Carvalho, João P. Marques Silva