Invented by JohnMiller of CarnegieMellonUniversity in the early 1990s, a little-known but useful way of stress-testing models. Not software models, and not UML models. Actual simulations and predictive models of real-world systems that involve algorithmic or mathematical formalism. For instance, the 400 variables that HighlyPaidConsultants think affect the future profitability of a business interact in certain equations (usually a fancy spreadsheet). That whole thing, the set of hypotheses it embodies, is a model. ActiveNonlinearTesting examines whether any of those equations, and the implied empirical measurements and assumptions and approximations that go into setting the constants that are in them, are unduly affected by error or uncertainty. This is different from, and better than, SensitivityAnalysis, since SensitivityAnalysis involves looking at the ''individual'' contribution of uncertainty in every variable in a model on the variation in the predicted outcome. ActiveNonlinearTesting allows many variations at once, all within "reasonable" limits. Probably the biggest reason that it isn't well accepted yet in the statistical community is that it uses a GeneticAlgorithm to do it. See JohnMiller's paper on the subject at *** http://www.santafe.edu/research/publications/wpabstract/199603011 ( BrokenLink ) *** http://www.santafe.edu/research/publications/workingpapers/96-03-011.ps Find more papers with the search: *** http://www.google.com/u/sfi?q=active+nonlinear+testing&domains=santafe.edu&sitesearch=santafe.edu ( BrokenLink ) ---- CategoryTesting