A New Approach using Machine Learning and Data Fusion Techniques for Solving Hard Combinatorial Optimization Problems

Abstract

We investigate the possibility of using kernel clustering and data fusion techniques for solving hard combinatorial optimization problems. The proposed general paradigm aims at incorporating unsupervised kernel methods into population-based heuristics, which rely on spatial fusion of solutions, in order to learn the solution clusters from the search history. This form of extracted knowledge guides the heuristic to detect automatically promising regions of solutions. The proposed algorithm derived from this paradigm is an extension of the classical scatter search and can automatically learn during the search process by exploiting the history of solutions found. Preliminary results, with an application to the well-known vehicle routing problem (VRP) show the great interest of such paradigm and can effectively find near-optimal solutions for large problem instances.