Fast Machine Learning (AI) Based High Fidelity Mesoscopic Modeling Tool for Traffic Simulation

Microscopic traffic simulators, such as VISSIM, use time-based step-by-step movement of vehicles to predict outcome of a proposed network configuration. This requires a full update of characteristics of each element in network every simulation time-step, which can be as frequent as every 100 ms. This process is inherently slow and impractical for use in stochastic large network optimization, which requires evaluation of large numbers of potential network configurations and/or signaling solutions. Our solution is to use a mesoscopic traffic simulator where cars are moved in a single step from one node in traffic network to another. The mesoscopic model uses a predictor function for each road segment which predicts travel time distribution for the traffic conditions using machine learning (AI) techniques. Our simulator chooses a random sample from predicted distribution as the travel time.  Travel time distributions may be sensitively dependent on parameters such as traffic and road conditions, weather, time of day, and traffic behavior patterns which may be dependent on specific road. By learning traffic behaviors rather than using a priori traffic formula, we will get a more accurate model of traffic behavior that reflects many of the real-world influences and a model that adapts to changing conditions on roadway. The fidelity of our simulator is similar to that of VISSIM for various traffic conditions. Furthermore, our simulator is more than 100 times faster than time-based microscopic simulation models, such as VISSIM, and provides network performance results that are comparable to these models.