Reproduce “Loss-calibrated expectation propagation for approximate Bayesian decision-making” on OpenResearch

About this paper

Michael J. Morais, Jonathan W. Pillow

Loss-EP extends expectation propagation to make posterior approximations sensitive to predictive utility. It adds a utility site whose updates inform data sites, allowing them to retain decision-relevant information. In a constructed bimodal reactor example, only Loss-EP preserves high-temperature probability to choose the Bayes-optimal shutdown. In simulated Gaussian-process classification, gains over EP were very small and not statistically significant after correction.

Key insights

• The 'best' approximate posterior for decision-making is dependent on the specific utility function, even though the true posterior distribution is not.
• For decision-making purposes, capturing utility-relevant features of the posterior may be more critical than achieving high fidelity to the overall posterior density.
• The factorized nature of Expectation Propagation provides a natural mechanism to integrate utility functions as additional factors, thereby calibrating the approximation to the decision problem.

Summary by alphaXiv.