A Framework for the Comparison of Agent-based Models.

Area of Science:

• Computational Social Science
• Agent-Based Modeling
• Systems Science

Background:

• Agent-based models (ABMs) are crucial for simulating complex systems but comparing diverse models remains challenging.
• Differences in data sources and model structures can hinder direct comparison of ABMs developed for the same domain.

Purpose of the Study:

• To introduce a novel methodology for comparing agent-based models that may vary in applied datasets and structural components.
• To enable robust comparison of ABMs by analyzing their behavior across a common parameter space.

Main Methods:

• Develop a methodology to learn a response surface within the shared parameter space of different agent-based models.
• Employ an active learning algorithm to identify phase shift boundaries in contagion processes.
• Apply the methodology to compare two agent-based models of rooftop solar panel adoption across different geographical regions.

Main Results:

• Demonstrated the ability to compare agent-based models with differing data and structures.
• Successfully identified and compared regions of qualitatively different model behaviors in parameter subspaces.
• Presented results for 2D and 3D parameter spaces, indicating scalability to higher dimensions.

Conclusions:

• The proposed methodology provides a robust framework for comparing diverse agent-based models.
• This approach facilitates a deeper understanding of model behavior and its sensitivity to parameters and data.
• The technique is applicable to various domains, including the analysis of technology adoption patterns.