Analysis of variable neighborhood descent as a local search operator for total weighted tardiness problem on unrelated parallel machines

Abstract

Variable Neighborhood Descent (VND) is a metaheuristic commonly used as a local search operator of other metaheuristics. This work analyzes the hypothesis in which the substitution of the local search operator by VND may increase the performance of three metaheuristics (Iterated Greedy Search - IGS, Artificial Bee Colony - ABC, and Genetic Algorithm - GA), proposed in the literature for the solution of minimization problems of the total weighted tardiness in Unrelated Parallel Machines environments. For the validation of this hypothesis, six neighborhood structures are proposed, considering the characteristics of the problem for reducing the search space exploitation. The analysis is carried out considering three VND variations, two neighborhood structures exploitation order, as well as exploitation by the First Improvement and Best Improvement methods. The Taguchi Robust Parameter method is used to design a specific configuration of the VND for each metaheuristic. Additionally, some experiments to analyze the contribution of each neighborhood structure for the convergence of the metaheuristics are performed. The results show that three neighborhood structures, act together, and domain the convergence influence of the local search. The results also show that the use of VND as a local search operator in place of those original local search increases the performance of all metaheuristics evaluated. Moreover, the results achieved by the metaheuristics integrated to the VND in the most evaluated scenarios have become equivalent or better, on average, than state-of-the-art approaches.