Arti®cial neural networks are often proposed as an alternative approach for formalizing various quantitative and qualitative aspects of complex systems. This paper examines the robustness of using neural networks as a simulation metamodel to estimate manufacturing system performances. Simulation models of a job shop system are developed for various con®gurations to train neural network metamodels. Extensive computational tests are carried out with the proposed models at various factor levels (study horizon, system load, initial system status, stochasticity, system size and error assessment methods) to see the metamodel accuracy. The results indicate that simulation metamodels with neural networks can be e ectively used to estimate the system performances.
IntroductionSimulation has been widely accepted by the scienti®c community and practitioners as a¯exible tool in modelling and analysis of complex systems. It reduces the cost, time and risks associated with the implementations of new designs. However, due to its lengthy computational requirements and the trail-and-error nature of the development process, simulation may not be always the ®rst preferred tool in solving real-life problems. This issue is particularly important for problems where the solution space is very large and when the time available for decisionmaking is too limited for extensive analysis to be performed (i.e. on-line applications). Harmonosky and Robohn (1995) stated that CPU requirements appeared to be a major obstacle for the on-line applications of simulation. Time considerations are still an important issue, even with the o -line use of simulation due to the increasing complexity of the modern production systems. In this paper, it is argued that the use of simulation metamodels may help alleviate these problems.A simulation metamodel is a simpler model of the real system. The simulation model is an abstraction of the real system in which a selected subset of inputs is considered. The e ect of the excluded inputs is represented in the model in the form of the randomness to which the system is subject. As illustrated in ®gure 1, a metamodel is a further abstraction of the simulation model. Simulation is used to generate data sets, which in turn are used to build the metamodel. A simulation metamodel with neural networks is a neural network whose training is provided by a simulation model. In general, a metamodel takes a fewer number of inputs