Downloaded from https://www.everyspec.com AMCP 701-137 the s-expectation of g(x), where X is a are involved. Special simulation lan-ges random variable, have been developed. Check with your com- A generalization of the law of large num- puter installation to find out what simulation bers comes into play during the repeated facilities are available, and what programming Monte Carlo trials: assistance that installation can offer. 114 MEASURES OF UNCERTAINTY im P )f(x)dx N-o Several methods are available for estab- N lishing s-confidence intervals and estimating 1 - (11-) uncertainties in the results of a simulation. , g(x) > 1 They are essentially the same as in any sam- I pling technique.
Chapter 4 reviews some of where the statistical concepts and gives references e = any positive number for further reading. The procedures are all (x) = pdfofx quite standard and well-known (to mathema- g(x) = any function of x; usually, the ticians). one being simulated N = sample size The required sample size for a given mini. - = sample value of X mum uncertainty is a handy number to have. x m uIt is useful for getting an idea of how much Eq. 11-1 shows that the chance of depar- computer time is likely to be involved.
For ture from the true value of g(x), weighted simulations of equipments, the programming according to the frequencies of the x's, be- and analytic effort to get ready to simulate comes less as N increases, will far outweigh the cost of actually running tr- simulations. Table 11-1 shows typical The reasoning can be extended to a func- sample sizes for various s-confidence levels tion of many variables, and goodness-of-fit (to the Cdf). 11-3 THE SIMULATION METHOD I The simulation method is a way to deter- TABLE 11-1 mine the diatribution of a function of one or MINIMUM SAMPLE SIZE REOUIRED more variables from the distributions of the FOR MONTE CARLO SIMULATION 6 individual variables.
The method involves ran- dom sampling from the distributions of all 6 = 0.90 = 0.96 3:.-0.9 variables and inserting the values so obtained 0.01 6800 9600 16500I in the equation for the function of interest. 0.02 1700 2400 4125 Suppose the function whose distribution is to 0.03 750 1066 1833 be estimated is g(x, , x 2 , ... I x ) and that the 0.05 272 384 660 X, X2, .. , are s-independent random -. m variables whose distributions are presumed to y- maximum deviatlon of sample Cdf from true Cdf beknown.
The procedure is to pick a set of - sco ifidn lve s rnomly froethe distiktio se of This tabla is derived from the Kolmogorov-Sm;rnov test of x's randomly from the distributions of the goodness-of-fit. It does not depend on the form of the X's, calculate g for that set, and store that distribution. value of g. The procedure is repeated many times until enough values of g are obtained. From this sample of g values, its distribution and parameters can be estimated. Very often, one settles for estimating the mean and stan- Since theory shows that the Monte Carlo dard deviation of g. technique gives a true random sample of the Simulation is a well developed art/science.
population (function) to be estimated, there * It is virtually always done on a computer be. is no need to go into special discussions about cause a tremendous number of calculations the statistical theory. 11-2