Multivariate emulation of computer simulators: model selection and diagnostics with application to a humanitarian...

Area of Science:

• Computational statistics
• Computer modeling
• Bayesian inference

Background:

• Multivariate output simulators are crucial for complex systems but computationally expensive.
• Existing emulation methods have limitations in handling complex, high-dimensional outputs.
• Developing robust emulation techniques is vital for efficient scientific exploration.

Purpose of the Study:

• To present a unified Bayesian emulation framework for multivariate output computer models.
• To develop novel diagnostics for assessing emulator performance.
• To compare parametric and non-parametric emulation approaches for accuracy and interpretability.

Main Methods:

• Developed a common framework for Bayesian emulation using parametric linear models and non-parametric Gaussian processes.
• Introduced novel diagnostics for multivariate covariance separable emulators.
• Applied and compared various emulators to a humanitarian relief simulator.

Main Results:

• The framework accommodates both parametric and non-parametric emulation strategies.
• New diagnostics enhance the evaluation of multivariate emulators.
• Sensitivity analysis revealed key input variable impacts on simulator outputs.

Conclusions:

• The proposed Bayesian emulation framework offers a versatile approach for multivariate simulators.
• The comparison provides insights into the strengths of different emulation methods for prediction and interpretability.
• This work advances the application of emulation in complex simulation studies.