The choice of the variable to flip in the Walksat family procedures is always random in that it is selected from a randomly chosen unsatisfied clause c. This choice in Novelty or R-Novelty heuristics also contains some determinism in that the variable to flip is always limited to the two best variables in c. In this paper, we first propose a diversification parameter for Novelty (or R-Novelty) heuristic to break the determinism in Novelty and show its performance compared with the random walk parameter in Novelty+. Then we exploit promising decreasing paths in a deterministic fashion in local search using a gradient-based approach. In other words, when promising decreasing paths exist, the variable to flip is no longer selected from a randomly chosen unsatisfied clause but in a deterministic fashion to surely decrease the number of unsatisfied clauses. Experimental results show that the proposed diversification and the determinism allow to significantly improve Novelty (and Walksat).
展开▼