Calibration of the SLEUTH urban simulation model using NOMAD and Genetic Algorithms

Computer simulations often entail an optimization problem, corresponding to the calibration of model parameters to ensure a precise representation of a given scenario. Many complex phenomena such as urban growth have characteristics that make optimization harder; this can be exemplified by the lack of an analytical formulation, presence of nonlinearity, discontinuities, and nondeterminism. SLEUTH is a long-established urban simulator, used to compute forecasts of city evolution. The tool is controlled by five parameters that span a search space of the order of 10 billion combinations, with a calibration procedure that is CPU-intensive and not compatible with gradient-descent methods. In this work we compare the efficiency of a genetic-algorithm version of the simulator with the use of the optimization library NOMAD. Different alternatives for the integration of the library are suggested. The experiments are analyzed using data profiles, a technique designed to handle cases with a limited number of function evaluations. The results confirm interest in NOMAD, and reveal more information than traditional comparisons of number of iterations or final optimization results. The methodology of the study can be applied to similar situations and is not restricted to simulators implementing urban models.