Optimization of discrete variable stochastic systems by computer simulation

A heuristic procedure is developed for determining the optimum values of the decision variables of discrete variable systems whose performances are evaluated by computer simulation. The objective function and some of the constraints of this optimization are various responses of the simulated model. The constrained simplex search method is the basis of this development. However, due to the stochastic nature of the simulation responses, the vertices of the simplex are compared statistically. The algorithm uses a variable simulation run length to minimize the required computer time. The data on the simulation output at each decision point are monitored continuously and, as soon as a statistically reliable comparison among the alternatives can be made, the simulation run at that point is terminated. The whole procedure is developed into an algorithm that can be interfaced with the simulation model built by the analyst. In this paper, the significant aspects of the algorithm and its application to a practical problem as well as the results of the comparison of its performance with respect to two other optimization search methods are presented.