Gradient and parameter sensitivity estimation for systems evaluated using Monte Carlo analysis

The performance evaluation of many practical systems can be handled only through computationally intensive Monte Carlo simulation. Although a number of specialist techniques have been proposed, in general, estimation of the sensitivity of the outcome to changes in parameters involves duplicate simulations and finite differences for each parameter of interest. An approximate technique for gradient sensitivity estimation was outlined previously. It is appropriate when the performance function is uni-modal and relatively smooth in the region of interest. It generates all gradients simultaneously by converting Monte Carlo simulation run outcomes to an approximate analytic problem defined by a simplified response surface. The gradients then follow immediately. No extra simulation runs are required. Herein that approach is extended to non-Normal random variables and to the estimation of parameter sensitivities for random variable means and standard deviations. Some illustrative examples are given with comparisons to sensitivities computed by conventional Monte Carlo. The influence of constraint function(s) defining the admissible solution region is also considered.