Mimicking Human Problem-Solving with Agents: Exploring Model Calibration

Agent-based social simulation models have a wide range of applications, and can incorporate numeric parameters of various kinds. In this paper, we use a simple agent-based simulation model of a laboratory experiment in network colouring to explore the selection of such numeric parameters. In particular, we examine two fundamental approaches to selecting model parameters (model calibration) based on empirical data: directly, comparing the data to model parameters; and indirectly, by comparing the data to model outputs. Using our model, we examine the strengths and weaknesses of the second approach. We discuss the insights provided by the model, and the extent to which confidence in these insights is justified when parameters are selected indirectly. The indirect approach to parameter selection has value in building social agent-based models, particularly when data on parameter values is unavailable, provided the number of parameters is relatively small.