Multi-level regression analysis as a tool to design and evaluate heuristic algorithms

Abstract

In this paper we propose an experimental setup and statistical methodology to design and evaluate heuristic algorithms. We apply our approach to various heuristics proposed in literature for the classical VRPTW. First, a multi-level regression analysis is used to determine the algorithms’ optimal parameter values and to construct decision rules stating which heuristic elements should be activated for a particular problem instance. Second, the performance of the various algorithms are statistically compared.