A survey of meta-heuristics used for computing maximin latin hypercube

Abstract

Finding maximin latin hypercube is a discrete optimization problem believed to be NP-hard. In this paper, we compare different meta-heuristics used to tackle this problem: genetic algorithm, simulated annealing and iterated local search. We also measure the importance of the choice of the mutation operator and the evaluation function. All the experiments are done using a fixed number of evaluations to allow future comparisons. Simulated annealing is the algorithm that performed the best. By using it, we obtained new highscores for a very large number of latin hypercubes.