Data-Mining-Driven Neighborhood Search

Metaheuristic approaches based on neighborhood search escape local optimality by applying predefined rules and constraints, such as tabu restrictions (in tabu search), acceptance criteria (in simulated annealing), and shaking (in VNS). We propose a general approach that attempts to learn (offline) the guiding constraints that, when applied online, will result in effective escape directions from local optima. Given a class of problems, the learning process is performed offline and the results are applied to constrained neighborhood searches to guide the solution process out of local optimality. Computational results on the Constrained Task Allocation Problem (CTAP) show that adding these guiding constraints to a simple tabu search improves the quality of the solutions found, making the overall method competitive with state-of-the-art methods for this class of problems. We also present a second set of tests on the matrix bandwidth minimization problem.