Comparing agent-based model versions using Approximate Bayesian Computation

Publication Name
Social Simulation Conference 2018
SSC 2018

Structural uncertainty in modelling social systems is a perennial problem, since there may be several ways to create formal descriptions and characterizations of the system itself and the dynamics operating therein. In this paper, we apply Approximate Bayesian Computation to the comparison of several versions of a model of domestic energy demand in north east Scotland. We demonstrate that these versions, which switch 'on' and 'off' various features in the software used to implement the model, show considerable differences in relative performance. In the case of the model itself, we are able to show which versions are more credible with respect to the data. More generally, however, the methodology described here shows how issues with structural uncertainty can be addressed in empirical agent-based modelling.

Book Chapter