Learning Short-Term Weights for GSAT

We investigate an improvement to GSAT which associates a weight with each clause. We change the objective function so that GSAT moves to assignments maximizing the weight of satis ed clauses, and each clause's weight is changed when GSAT moves to an assignment in which this clause is unsatis ed. We present results showing that this version of GSAT has good performance when clause weights are reduced geometrically throughout the course of a single try. We conclude that clause weights are best interpreted as short-term, context sensitive indicators of how hard di erent clauses are to satisfy.