The recent advances in the computational area allowed the microscopic simulation of vehicular traffic to become one of the main tools used by transportation professional for traffic design and analysis purposes (e.g. evaluating the performance and level of safety of highways, the impact of different traffic management strategies, etc.). On the other hand, because the calibration of micro-simulation models is a complex and resource intensive process one might resort to using the default parameter values when, for instance, repetitive simulation runs are needed to reutilize an already calibrated model. However, this practice rarely leads to valid results. To address this problem we propose an automated calibration procedure by integrating two soft-computing techniques, neural-network and genetic-algorithm. This methodology is independent of the simulation platform used and in this study it is applied using VISSIM, a commercial microscopic simulation, and MATLAB. First, a Latin Hypercube Sampling method is used to select representative sets of values for VISSIM's main calibration parameters. Next, the effect of each set of parameter values on the simulated traffic stream speed is recorded. A neural-network is trained to determine the relationship between the input parameter values and the output traffic stream speed. Finally, a genetic-algorithm uses the trained neural-network as a fitness function to determine the appropriate set of values for the calibration parameters of the collected speed data from a real-world test networks. The proposed methodology allows calibrating microscopic traffic models with fewer computational resources than commonly used methods.
展开▼
