Towards real-time predictions using emulators of agent-based models

The use of Agent-Based Models (ABMs) to make predictions in real-time is hindered by their high computation cost and the lack of detailed individual data. This paper proposes a new framework to enable the use of emulators, also referred to as surrogate models or meta-models, coupled with ABMs, to allow for real-time predictions of the behaviour of a complex system. The case study is that of pedestrian movements through an environment. We evaluate two different types of emulators: a regression emulator based on a Random Forest and a time-series emulator using a Long Short-Term Memory neural network. Both emulators perform well, but the time-series emulator proves to generalise better to cases where the number of agents in the system is not known a priori. The results have implications for the real-time modelling of human crowds, suggesting that emulation is a feasible approach to modelling crowds in real-time, where computational complexity prohibits the use of an ABM directly.