AI-assisted calibration of real-time Perth traffic simulation

The Transport predictive solution Stage 2: AI and real-time simulation project, conducted with iMOVE partners Main Roads Western Australia, Aimsun, and the University of Western Australia has been completed, and its final report is available for download below.

The aim of this work was to improve the prediction traffic congestion on the Perth road network and quickly assess the impact of unplanned events and evaluate the mitigation potential of several possible responses

This research looked to improve model calibration and the accuracy of 24 hour/ 7-day models (live and offline) for not just the AM and PM peaks but any hour of any day. It did so by employing the use of the Perth Live Aimsun Model, a real-time predictive traffic management solution.

This research looked to improve model calibration and the accuracy of 24 hour/ 7-day models (live and offline) for not just the AM and PM peaks but any hour of any day. It did so by employing the use of the Perth Live Aimsun Model, a real-time predictive traffic management solution.

Key outputs of this project were:

1. The development of methods for the offline calibration of the model.
2. Identification of patterns in both traffic volume and optimal model parameters.
3. Comparison of metrics for evaluating the quality of the model prediction.

The Perth Live Model

The Perth Live Model consists of a hybrid macroscopic-mesoscopic model, wherein the larger Perth region is modelled by a static macroscopic model with a smaller region around the CBD modelled by a mesoscopic discrete-event based simulation.

Traffic in the macroscopic region is modelled at the aggregate level of vehicle flow according to the principle of user equilibrium, origin-destination (OD) demand, and travel times associated with each segment and intersection.

Data

The Aimsun Perth Live Model operates online, continuously adapting its predictions based on incoming data from the real world.
This project used loop detector data as observations of the real-world traffic state. The data consists of volume observations at a 5-minute resolution.

From that data, the research team applied the following techniques:

1. AI-assisted calibration module: Development and implementation of statistical learning techniques to estimate optimal driver behaviour and supply parameters in different conditions.
2. Pattern refinement module: Use of clustering techniques to identify patterns in driver behaviour and supply via the optimised parameter values
3.  Prediction confidence module: Identification of the best statistical measures for model confidence, in both training and evaluation contexts.

The final report goes into some detail about the AI-assisted calibration, along with various optimisation algorithms used to fine-tune and increase the prediction confidence of the model.

Consclusions

Using machine learning and AI techniques, this project successfully overcame the issues of a large-scale, noisy simulation model through the development of methods for calibrating the driver behaviour and supply-side parameters for Aimsun Live.

These methods also eliminated the need for use of massive computational resources by “… reducing the dimensionality of the problem with sensitivity analysis, and using Bayesian optimisation, which is designed to be efficient with function evaluations.

The key contribution of this project is the establishment of a methodology for the systematic calibration of driver behaviour and supply-side parameter.

Future work

The methods presented in this research use optimisation with respect to historical data to recommend parameter values for running the model in practise. This approach is limited in the degree to which it can respond to unexpected changes in traffic dynamics, only allowing flexibility via matching with patterns in the historical data.

True real-time optimisation involves selecting the parameter values for the model based on current information about the state of the network, alongside the historical data. This advantage of the real-time approach is that the model is more capable of responding to unusual and unseen behaviours in the system.

However, due to the computational requirements of the simulation model, basic attempts at real-time optimisation are not feasible and more advanced methods are required.

Looking ahead, to build on this work the researchers propose the following suggestions:

• Apply similar techniques to calibrate offline models.
• Based on the results in Section 5, root-mean-squared error (RMSE) should be adopted as the cost function for performing optimisation.
• The source of unexpected, optimised parameter values Could be investigated.
• Model emulation could be explored to develop a simplified surrogate model that approximates the full model. This could enable
  • True real-time parameter optimisation via accelerated model evaluations.
  • o Better testing of cost functions. The ability of optimisation to recover the known ground truth parameters of the emulation indicates the most suitable cost function.
• Clustering results could be improved by considering an adaptive distance threshold across different branches of the hierarchical tree structure. For example, a finer clustering resolution may exist at lower levels for weekdays, while a coarser resolution is sufficient for Sundays.
• More data sources could be included for calibration (e.g., vehicle speeds) if available

Download the final report

Download your copy of the final report, Transport Predictive Solution – Stage 2 – R&D – WA Node: AI-assisted Model Calibration for Realtime Traffic Simulation, by clicking the button below.