Tom Lymburn - Challenges in optimising large scale traffic simulation models

In order to monitor and predict complex phenomena, such as traffic on a large region of road network, operators often use theory-driven simulation models, which have the advantage of extensibility and interpretability when compared to machine learning methods. However, the optimisation of these models is often restricted to a "black-box" approach, due to high complexity and difficulty in measuring the parameters in the real world (assuming the model parameters correspond to their physical counterparts at all). In this seminar I will discuss the optimisation of driver behaviour parameters for an agent-based simulation covering an extended region of the Perth network. This model is employed in real-time for the prediction of the next hour of traffic. The key challenges in optimising this model arise from its high-dimensional nature, the presence of noise, computational limitations, as well as the assumption that the optimal parameters themselves are dynamic. In order to address these issues we use techniques and shortcuts to reduce the scale of the problem, such as Bayesian optimisation and a nearest neighbour method that efficiently reuses simulation outputs for multiple tasks. Finally, I will discuss attempts at model emulation, which have been unsuccessful so far, but could be used for true real-time optimisation.