Improving random number generators on micro-computers

Some random number generators in micro-computer software are found to have serial correlations at low-order lags. Experiments using composite generator methods are performed in an attempt to remove these serial correlations. Success is achieved when proper input parameter settings are used to direct the composite generator methods. A procedure is presented whereby these methods are applied to the global (entire sample) and the local (subsamples) sequences of the random number stream. Results for a wide range of statistical tests of the random numbers are presented.