The efficacy of the application of traffic models depends on a successful process of model calibration. Microsimulation models have a significant number of input parameters that can be optimized in the calibration process. This paper presents the optimization of input parameters that are difficult to measure or unmeasurable in real traffic conditions and includes parameters of the driver’s behavior and parameters of Wiedemann’s psychophysical car-following model. Using neural networks, models were generated for predicting travel time and queue parameters and were used in the model calibration procedure. This paper presents the results of a comparison of five different applications of neural networks in calibrating the microsimulation model. The VISSIM microsimulation traffic model was selected for calibration and field measurements were carried out on two roundabouts in a local urban transport network. The applicability of neural networks in the process of calibrating the microsimulation models was confirmed by comparison of the modelled and measured data of traffic indicators in real traffic conditions. Methods of calibration were validated with two sets of new measured data at the same intersection where the calibration of the model was carried out. The third validation was made at the intersection in a different location. The selection of the optimal calibration methodology is based on the model accuracy between the simulated and measured data of traveling time, as well as queue parameters. The microsimulation model provides access to the raw data of observed traffic parameters for each vehicle in the simulation. The dataset of the calibrated model simulation results of all travel times of the selected traffic flow was compared with the dataset of the measured field data to determine whether the data are statistically significantly different or not.