| Transportation agencies often use microscopic traffic simulators to evaluate the impact on traffic performance of changes in traffic volume,road geometric design,and traffic control schemes.In performing simulations,a premise is to have appropriately calibrated parameter values.This however is often computationally expensive as it requires repeatedly running simulations in search for the best parameter value set.In this paper,we propose a machine learning(ML)-based methodology for calibrating microscopic traffic simulator parameters which avoids repeated running of simulations,thus significantly improving the computational efficiency of calibration.The methodology first develops machine learning models that use the parameters to be calibrated as inputs to predict simulation outputs.Four machine learning models: decision tree,support vector machine,Gaussian process regression,and artificial neural networks are considered.The best-performing model is embedded in heuristic algorithms(HA)to seek the set of parameter values that minimizes the difference between the predicted simulation output using the embedded ML model and the observations from the field data.Three machine learning models: particle swarm optimization,genetic algorithm,and Tabu search are considered.To demonstrate their use,the ML+HA calibration methodology is applied to Trans Modeler calibration of highway links in Shanghai,China.The experiment results indicates that artificial neural network model yields the best prediction accuracy.We also find that particle swarm optimization with embedded artificial neural networks shows superior computational efficiency,and Tabu search shows higher accuracy of the proposed methodology for calibration. |
