Application of Partial Restarts in a Satisfiability Algorithm

In this paper I describe experiments in the application of dynamic restarts used in heuristic satisfiability algorithms. By modifying a state-of-the-art SAT solver, zchaff, I test three proposed modifications to typical restart policies: exponentially increasing restart thresholds, restricting restarts to occur only at certain critical points, and the addition of transient randomness after a restart. By comparing runtimes on several benchmarks, I evaluate many of the tradeoffs involved in choosing an appropriate restart policy and also attempt to tackle the question of whether or not to use restarts at all. Section 1 introduces Chaff and the basic ideas of restarts, section 2 gives background on research on dynamic restarts, sections 3-5 described our proposed experiment, sections 6-8 describe the procedure, section 9 tabulates the results, and sections 10-11 give my conclusions.