Value-ordering heuristics: Search performance vs. solution diversity

Abstract

We examine the behavior of dynamic value-ordering heuristics in a CSP under the requirement to generate a large number of diverse solutions as fast as possible. In particular, we analyze the trade-off between the solution search performance and the diversity of the generated solutions, and propose a general probabilistic approach to control and improve this trade-off. Several old/new learning-reuse heuristics are described, extending the survivors-first value-ordering heuristics family. The proposed approach is illustrated on a real-world set of examples from the Automatic Test Generation problem domain, as well as on several sets of random binary CSPs. © 2010 Springer-Verlag.