This paper proposes a system for automatic calibration of two-lane traffic simulation models based on a genetic algorithm. The calibration process requires searching for values of selected model parameters, as to minimize the differences between simulation results and the observed data. The calibration parameters are usually related to driver behavior and because of their large number and the fact that their affects are often highly correlated they are difficult to calibrate for specific applications. In this paper the differences between the simulation results and the observed data are assessed based on performance measures that are related to the application of the model (e.g., average travel speed, percent vehicles in platoons, etc.). The proposed procedure calibrates each model using several different highway sections to find the set of calibrated parameters that can best represent a typical two-lane highway in Brazil. Data for the calibration were collected in five different locations in the state of São Paulo which were considered representative of roads in the region in terms of terrain and traffic mix. Two data sets were collected, one for model calibration and the other for model validation. The results indicate that the proposed approach is highly efficient, albeit computationally intensive; the average difference between simulated and observed traffic data was less than five percent. Copyright ASCE 2006.