Trajectory-based global sensitivity analysis in multiscale models.

Area of Science:

• Computational modeling and simulation
• Complex systems analysis
• Uncertainty quantification

Background:

• Agent-based models (ABMs) are crucial for simulating complex systems but present unique analytical challenges.
• Traditional global sensitivity analysis (GSA) methods often struggle with the multi-level structure and temporal dynamics inherent in ABMs.
• A need exists for advanced GSA frameworks tailored to the specific characteristics of ABMs.

Purpose of the Study:

• To introduce a novel GSA framework specifically designed for agent-based models (ABMs).
• To enable a more comprehensive estimation of parametric sensitivities across different scales within ABMs.
• To provide robust tools for detailed model analysis and informed decision-making in ABM development and verification.

Main Methods:

• Utilizes Grassmannian diffusion maps for dimensionality reduction of ABM output data.
• Employs sparse polynomial chaos expansion (PCE) to calculate sensitivity indices for stochastic parameters.
• Applies the framework to diverse models including system dynamics, epidemiological, and economic ABMs.

Main Results:

• The proposed GSA framework effectively handles the multi-level structure and temporal dynamics of ABMs.
• Demonstrates the ability to estimate parametric sensitivities from the micro-level (individual agents) to the macro-level (population).
• Successfully applied to various ABMs, showcasing its versatility across different dynamic systems.

Conclusions:

• The novel GSA framework enhances the understanding of complex spatio-temporal processes within ABMs.
• Encourages the adoption of manifold-based techniques in uncertainty quantification for complex models.
• Equips ABM practitioners with advanced tools for model analysis, verification, and refinement, improving routine practice.