TY - CHAP
TI - A framework for testing, comparing and visualizing the performance of non-uniform random variate generators
AB - There exist powerful libraries for testing uniform pseudo-random number generators, e.g., Marsaglia's Diehard test suite or L'Ecuyer's TestU01 library. Commonly used generators are extensively tested. For non-uniform random variate generators the situation seems to be quite different. Authors propose new algorithms, create proof-of-concept implementations, and report their experiences. Our personal impression is that code for the performance and validation tests often is especially written for the particular generator and hard to reuse. When running tests on a large range of parameter values for an experimental generator it is often hard to choose suitable sample sizes that result in practicable running times and to avoid to run the test at all on parameter values where the generator is prohibitively slow. Another poor practice in literature is that performance figures like marginal generation times or rejection constants are often reported and compared to competitive algorithms by means of (small) tables. This is sufficient for the standard normal distributions but inadequate for distributions with one or more shape parameters like the GIG or the generalized hyperbolic distribution. We therefore present a general framework for testing the correctness and performance of non-uniform random variate generators. It allows to run tests for many different parameter values and visualizes their results. To avoid problems with unexpected long running times a timeout can be set for each test. The test suite has been implemented in R as this allows interactive programming and provides many routines for further processing of the test results. Moreover generators that are coded in C/C++/Fortran can be easily included. We demonstrate the practical importance of our new framework by analyzing a couple of generators for the GIG distribution as well as generators for the beta distribution.
AF - 10th IMACS Seminar on Monte Carlo Methods (MCM 2015)
PP - Linz
PY - 2015-01-01
AU - Leydold, Josef
ER -