The impact of using different choice functions when solving CSPs with autonomous search

Abstract

Constraint programming is a powerful technology for the efficient solving of optimization and constraint satisfaction problems (CSPs). A main concern of this technology is that the efficient problem resolution usually relies on the employed solving strategy. Unfortunately, selecting the proper one is known to be complex as the behavior of strategies is commonly unpredictable. Recently, Autonomous Search appeared as a new technique to tackle this concern. The idea is to let the solver adapt its strategy during solving time in order to improve performance. This task is controlled by a choice function which decides, based on performance information, how the strategy must be updated. However, choice functions can be constructed in several manners variating the information used to take decisions. Such variations may certainly conduct to very different resolution processes. In this paper, we study the impact on the solving phase of 16 different carefully constructed choice functions. We employ as test bed a set of well-known benchmarks that collect general features present on most CSPs. Interesting experimental results are obtained in order to provide the best-performing choice functions for solving CSPs.