Although many problems in the literature involve complex mathematical relationships, many still rely on simplified and unrealistic assumptions. Simulation is one of the most powerful tools for dealing with such problems, as it avoids the restrictive assumptions often required in stochastic systems. Simulation optimization techniques are generally classified into two broad categories: model-based and metamodel-based methods. In the first category, simulation and optimization components interact directly, thereby increasing simulation time and cost. To address this issue, a third component—called a metamodel—is introduced in the second category to estimate the system's relationship between input and output variables. Optimizing semi-expensive simulation problems often requires many simulation runs in model-based methods. However, the cost of validating metamodels also rises rapidly during iterations. A two-phase method has been proposed in the literature to reduce computation time. In the first phase, similar to a model-based algorithm, the simulation output is used directly in the optimization process. In the second phase, a validated metamodel replaces the simulation model. In this paper, an artificial neural network (ANN) is employed as the metamodel, and its performance is compared with that of the original algorithm, which employs a Kriging metamodel, on five well-known test functions and an (s, S) inventory model.