B1423
Title: Generative AI for model selection
Authors: Jungeum Kim - Booth School of Business, University of Chicago (United States) [presenting]
Abstract: Understanding the Bayesian analogue of frequentist power analysis greatly enhances decision-making via Bayes factors. However, existing methods reliant on MCMC for estimating Bayes factors become impractical for extensive simulations required in power analysis. Furthermore, reliable Bayes factor estimators are absent when likelihood computation is unfeasible. An MCMC-free, likelihood-free method is presented for Bayes factor estimation. Leveraging the intrinsic connection between Bayes factors and classification, a deep learning classifier is trained that serves as a neural Bayes factor estimator. The method offers efficient Bayes factor estimations, reducing the computational burden of diagnostics on massive simulated data. Moreover, the approach facilitates more flexible model design beyond the constraints of standard distributions implemented in MCMC-based methods.
