Variational inference#
Variational inference (VI) approximates a Bayesian posterior by optimizing a parameterized probability distribution. ROM Tools supports mean-field and full-covariance Gaussian VI. Multifidelity VI (MF-VI) combines full-order and surrogate-model evaluations to reduce the cost of estimating the optimization objective and its derivatives.
Examples#
Analytic sine VI and MF-VI: compare diagonal and full-covariance approximations against a known Gaussian posterior.
H2-air flame VI and MF-VI: compare full-order inference with Gaussian-process surrogate-assisted inference.
For signatures and configuration options, see Inverse Problems. Detailed algorithm and restart controls are documented in VI and MF-VI implementation notes.