Empirical comparison of search algorithms for discrete event simulation

Abstract

Discrete-event simulation is a significant analysis tool for designing complex systems. In the research literature, several deterministic search algorithms have been linked with simulation for industrial applications; but there are few empirical comparisons of the various algorithms. This paper compares the Hooke–Jeeves pattern search, Nelder–Mead simplex, simulated annealing, and genetic algorithm optimization algorithms on variations of four industrial case study simulation problems. The simulation models include combinations of real variables, integer variables, non-numeric variables, deterministic constraints, and stochastic constraints. The genetic algorithm was the most robust, as it found near best solutions for all 25 test problems. However, it required the most replications of all the algorithms. The pattern search algorithm also found near best solutions to small- and medium-sized problems with no non-numeric variables, while requiring fewer replications than the genetic algorithm.