Variational Bayesian identification and prediction of stochastic nonlinear dynamic causal models

Area of Science:

• Computational Statistics
• Machine Learning
• Dynamical Systems Theory

Background:

• Approximate inference is crucial for analyzing complex nonlinear stochastic dynamic models.
• Existing methods often struggle with nonlinearities, unknown parameters, and model comparison.
• Accurate inference is essential for tasks like time-series prediction and model identification.

Purpose of the Study:

• To present a general variational Bayesian approach for approximate inference in nonlinear stochastic dynamic models.
• To extend existing methods to handle nonlinearities, unknown parameters, and hyperparameters.
• To enable robust model comparison and prediction under uncertainty.

Main Methods:

• Developed a variational Bayesian scheme optimizing a free-energy bound on the model evidence.
• Derived a deterministic update scheme approximating the posterior density of model variables.
• Applied the method to nonlinear stochastic dynamic hierarchical models.

Main Results:

• The proposed scheme effectively handles nonlinear evolution and observation functions, and unknown parameters/hyperparameters.
• Demonstrated capabilities in model identification, comparison, and time-series prediction.
• Computational complexity is comparable to the extended Kalman filter, suitable for high-dimensional models.

Conclusions:

• The variational Bayesian approach offers a powerful and efficient tool for inference in complex nonlinear dynamic systems.
• The method provides a unified framework for parameter estimation, model comparison, and prediction.
• Validated through Monte-Carlo simulations on chaotic systems, showing good estimation efficiency and predictive power.