Bagged filters for partially observed interacting systems.

Area of Science:

• Computational statistics
• Stochastic modeling
• Epidemiological modeling

Background:

• Bagging (bootstrap aggregating) combines multiple estimators for improved inference.
• Stochastic dynamic systems with interacting units pose computational challenges, especially in high dimensions.
• Monte Carlo filtering methods struggle with the curse of dimensionality in nonlinear, non-Gaussian systems.

Purpose of the Study:

• To introduce a novel bagged filter (BF) methodology for inference in interacting stochastic dynamic systems.
• To address the curse of dimensionality in complex systems, particularly in spatiotemporal modeling.
• To evaluate the performance of BF against existing filtering techniques.

Main Methods:

• Ensemble of Monte Carlo filters combined using spatiotemporally localized weights.
• Selection of successful filters at each unit and time point.
• Application to coupled population dynamics models for infectious disease transmission.

Main Results:

• Bagged filter (BF) methodology effectively handles inference for noisy or incomplete data in interacting systems.
• BF can overcome the curse of dimensionality under specific conditions and demonstrates applicability even when these conditions are not met.
• BF outperformed an ensemble Kalman filter in a coupled population dynamics model.

Conclusions:

• Bagged filter (BF) provides a robust and scalable approach for inference in complex interacting stochastic systems.
• BF offers advantages over traditional methods by mitigating the curse of dimensionality and preserving system properties.
• The methodology shows significant potential for applications in fields like epidemiology.