Equation-Based Versus Agent-Based Models: Why Not Embrace Both for an Efficient Parameter Calibration?

In this work, we propose 2LevelCalibration, a multi-stage approach for the calibration of unknown parameters of agent-based models. First, we detail the aspects of agent-based models that make this task so cumbersome. Then, we conduct extensive research on common methods applied for this purpose in other domains, highlighting the strong points of each approach that could be explored to efficiently calibrate the parameters of agent-based models. Finally, we present a multi-stage method for this task, 2LevelCalibration, which benefits from the simplicity of equation-based models, used to faster explore a large set of possible combinations of parameters and to quickly select the more promising ones. These values are then analysed more carefully in the second step of our method, which performs the calibration of the agent-based model parameters' near the region of the search space that potentially contains the best set of parameters, previously identified in our method. This strategy outperformed traditional techniques when tested to calibrate the parameters of an agent-based model to replicate real-world observations of the housing market. With this testbed, we show that our method has highly desirable characteristics such as lightweight implementation and consistency, which are ideal for agent-based models' development process.