SIMULATION METAMODELLING VIA LOCAL LINEAR INTERPOLATION

Simulation metamodels are constructed by fitting a model to a dataset wherein each sample in the dataset typically consists of a set of inputs (metamodel `design points'), and estimates of one or more simulation outputs that form the dependent variable(s) for the metamodel. In this paper, we present a novel approach for constructing simulation metamodels that involves working directly with the simulation replications generated for each metamodel design point. Simulation output estimates corresponding to a test point - a set of simulation inputs not part of the design space - are generated by computing a convex combination of replicate outcome estimates from a set of nearest `neighbours' of the test point in the design space. This approach eliminates the curve-fitting step in the metamodel construction process, and because multiple realisations of outcome estimates for a single test point are generated, possesses the potential of preserving the stochasticity associated with simulation outcomes.