Finding Good Subtrees for Constraint Optimization Problems Using Frequent Pattern Mining

Abstract

Making good decisions at the top of a search tree is important for finding good solutions early in constraint optimization. In this paper, we propose a method employing frequent pattern mining (FPM), a classic datamining technique, to find good subtrees for solving constraint optimization problems. We demonstrate that applying FPM in a small number of random high-quality feasible solutions enables us to identify subtrees containing optimal solutions in more than 55% of problem instances for four real world benchmark problems. The method works as a plugin that can be combined with any search strategy for branch-and-bound search. Exploring the identified subtrees first, the method brings substantial improvements for four efficient search strategies in both total runtime and runtime of finding optimal solutions.

Finding Good Partial Assignments during Restart-Based Branch and Bound Search

“…Frequent Pattern Mining-based Search (FPMS) (Li et al 2020) finds good subtrees containing high-quality solutions for loosely constrained optimization problems. It employs a random search procedure to generate sample solutions to represent the feature of the solution space and uses frequent pattern mining to find a good partial assignment.…”

“…different restart runs (Boussemart et al 2004;Michel and Van Hentenryck 2012;Li, Yin, and Li 2021). Recently, Frequent Pattern Mining-based Search (FPMS) (Li et al 2020) is proposed to find good subtrees for solving COPs. It makes BBS directly zoom into a promising sub-search-space by running a pre-processing phase using frequent pattern mining on high quality sampled solutions.…”