An artificial neural network metamodel to solve semi-expensive simulation optimization problems: A comparative study

Although a considerable number of problems whose analysis depends on a set of complex mathematical relations exist in the literature due to recent developments in decision-making, still very simplified and unrealistic assumptions are involved in many. Simulation is one of the most powerful tools for dealing with this kind of problem, and it enjoys being free of any restricting assumptions that may generally be considered in a stochastic system. Besides, simulation optimization techniques are categorized into two broad classes: model-based and meta-model-based methods. In the first class, simulation, and optimization components interact, causing an increase in simulation times and costs. To cope with this problem, a third component, defined as a metamodel that estimates the relationships between the inputs and outputs of the system being simulated, comes to the picture in the second-class problems. Furthermore, optimizing semi-expensive simulation optimization problems requires numerous simulation runs in model-based methods. However, as the validation cost increases rapidly in each iteration of the metamodel-based techniques, a new method consisting of two phases has been introduced in the literature to solve semi-expensive simulation optimization problems in less computational time. In the first phase, similar to a model-based algorithm, the simulation output is used directly in the optimization stage. The simulation model is changed to a validated metamodel in the second phase. In this paper, an artificial neural network is employed as the metamodel to compare its performance to the ones of the original algorithm that uses a Kriging metamodel in five widespread test problems as well as an ( s , S ) inventory problem.