Comparing Algorithms for Scenario Discovery

While attractive in many ways, traditional scenarios have lacked an appropriate analytic foundation for inclusion in quantitative decision analyses. In previous work, we have proposed to remedy this situation with a systematic, analytic process we call "scenario discovery" that has already proved useful in a wide variety of applications. This study aims to evaluate alternative algorithms needed to implement this novel scenario discovery task, in which users identify concise descriptions of the combinations of input parameters to a simulation model that are strongly predictive of specified policy-relevant results. This study offers three measures of merit -coverage, density, and interpretability and uses them to evaluate the capabilities of PRIM, a bump-hunting algorithm, and CART, a classification algorithm. The algorithms are first applied to datasets containing clusters of known and easily visualized shapes, then to datasets with unknown shapes generated by a simulation model used previously in a decision analytic application. We find both algorithms can perform the required task, but often imperfectly. The study proposes statistical tests to help evaluate the algorithms' scenarios and suggests simple modifications to the algorithms and their implementing software that might improve their ability to support decision analysis with this scenario discovery task.