It typically can be difficult to create and solve optimization models for large-scale sequential decision problems, examples of which include applications such as communications networks, inventory problems, and portfolio selection problems. Monte Carlo simulation modeling allows for the creation and evaluation of these large-scale models without requiring a complete analytical specification. Unfortunately, optimization of such simulation models is especially difficult given the large state spaces that often produce a combinatorially explosive number of potential solution policies. In this paper we introduce a new technique, Simulation for Model Generation (SMG), that begins with a simulation model of the system of interest and then automatically builds and solves an underlying stochastic sequential decision model of the system. Since construction and implementation of the created model requires approximation techniques, we also discuss several types of error that are induced into the decision process. Fortunately, the decision policies produced by the SMG approach can be directly evaluated in the original simulation model-thus the results of the SMG model can be compared against any other possible strategies, including any decision policies currently in use.