Calibrating Agent-Based Models of Innovation Diffusion with Gradients

Abstract

Consumer behavior and the decision to adopt an innovation are governed by various motives, which models find difficult to represent. A promising way to introduce the required complexity into modeling approaches is to simulate all consumers individually within an agent-based model (ABM). However, ABMs are complex and introduce new challenges. Especially the calibration of empirical ABMs was identified as a key difficulty in many works. In this work, a general ABM for simulating the Diffusion of Innovations is described. The ABM is differentiable and can employ gradient-based calibration methods, enabling the simultaneous calibration of large numbers of free parameters in large-scale models. The ABM and calibration method are tested by fitting a simulation with 25 free parameters to the large data set of privately owned photovoltaic systems in Germany, where the model achieves a coefficient of determination of R 2 ≃ 0.7.

Data assimilation for online model calibration in discrete event simulation

“…8 Another work of McCulloch et al 9 developed an agent-based model calibration framework, which is based on approaches from the fields of uncertainty quantification and model optimization. An example of calibrating agent-based models of innovation diffusion is presented in Kotthoff and Hamacher 10 which uses a gradient-based calibration method. Other examples of offline calibration include building energy simulation calibration 11–13 and traffic simulation calibration.…”