Investigating of machine learning's capability in enhancing traffic simulation models

The development of agent-based modeling in traffic simulation allows for the modeling of traveler movement and decision making using predefined rules and variables. Nonetheless, the computational cost of agent-based modeling is high, and it takes a long time to generate new scenarios using these models. To address this, this study proposes a new approach to predict the results of new simulations using machine learning techniques. This paper focuses on the reproduction of the models that simulate variables reflecting traveler decision making, such as mode choice, travel distance and duration, and waiting time. A variety of data-driven techniques have been employed in this regard to model these features resulting from unanticipated activities in a dynamic environment. The proposed approach will be based on synthetic data generated from various simulation scenarios, that will be followed by a data preparation process. Therefore, the robustness of the built machine learning models was tested and assessed in different and new situations in order to evaluate their capability to reproduce the models responsible for generating the stated variables. Experiments show that the suggested solution has a high level of robustness, implying that it can replicate the final results of these models. Further, Extreme Gradient Boosting outperformed other machine learning techniques in terms of predicting simulation variables when comparing prediction accuracy and running time.

Uncontrolled Keywords

• transport simulation
• machine learning
• data analysis
• agent based model