Using emulators to approximate predicted performance indicators of complex microsimulation and multiagent models of travel demand

Abstract

The very high run times associated with microsimulation and agent based models of travel demand preclude their application in real time policy development sessions in which discussions require fast turnaround times of model simulation results. This paper suggests and explores the use of emulators to increase the possibilities of using large scale simulation models in real time policy development sessions. An emulator is a statistical approximation of a large scale model, depicting the relationship between a set of input variables and an output variable of a large scale model. To illustrate the approach, an emulator is built for the Albatross model system with special reference to its daily distance traveled per person performance indicator. Results are satisfactory and show that simple main effects plus first order interaction effects of regression model can approximate daily distance traveled per person, simulated by the considerably more complex large scale Albatross model system, very accurately. The paper also examines the impact of the number of simulation runs on the performance of the emulator. Results suggest that the accuracy of the emulator improves, albeit in a non-linear fashion, with an increasing number of simulation runs. Implications and avenues of future research are discussed.