Exploration of Gaussian processes for emulation of stochastic models

Working with Turing Research Software Engineer Louise Bowler, we have recently finished a short project that experimented with the use Gaussian Processes (GPs) to emulate agent-based models. It was developed as part of the project Uncertainty in agent-based models for smart city forecasts, funded by the Alan Turing Institute.

The results are available in full from the GitHub Repo. The emulators worked well at emulating the behaviour of deterministic agent-based models. But when a degree of randomness was introduced into the model, as is often the case in agent-based modelling, they struggled to cope with the additional uncertainty and did not emulate the behaviour of the underlying models particularly well. The project has identified a number of interesting questions that need further research.