A hybrid modeling approach for parking and traffic prediction in urban simulations

Urban simulations are an important tool for analyzing many policy questions relating to the usage of public space, roads, and communal transportation; they can be used to predict the long-term impact of new construction projects, traffic restrictions, and zoning laws. However, it is unwise to rely upon predictions from a single model since each technique possesses different strengths and weaknesses and can be highly sensitive to the choice of parameters and initial conditions. In this article, we describe a hybrid approach for combining agent-based and stochastic simulations (Markov chain Monte Carlo, MCMC) to improve the accuracy and reduce the variance of long-term predictions. In our proposed approach, the agent-based model is used to bootstrap the proposal distribution for the MCMC estimator. To demonstrate the applicability of our modeling technique, this article presents a case study describing the usage of our hybrid simulation method for forecasting transportation patterns and parking lot utilization on a large university campus. A comparison of our simulation results against an independently collected dataset reveals that our hybrid approach accurately predicts parking lot usage and performs significantly better than other comparable modeling techniques. Developing novel architectures for combining the predictions of agent-based models can produce insights that are different than simply selecting the best model.