A0866
Title: Sequential Bayesian design via Laplace policies
Authors: Emma Rowlinson - University of Manchester (United Kingdom)
Tim Waite - University of Manchester (United Kingdom) [presenting]
Abstract: Policy-based methods for sequential Bayesian experimental design aim to learn a mapping from the current knowledge state to optimal future experiments, maximizing a utility such as expected information gain. A key consideration is how to represent the knowledge state. The proposal is to use the Laplace approximation to the posterior as this representation, enabling efficient training of neural network policies. The technical foundations of the Laplace policy framework are introduced, and its performance is illustrated across a range of design problems. To enable gradient-based optimization, differentiable computation of the posterior mode is ensured via custom gradient methods based on the implicit function theorem. The framework is further extended to binary response models using concrete relaxation, which enables approximate simulation of discrete random variables while preserving the differentiability of the widely used reparameterization technique for simulation of continuous random variables. New results are presented, including comparisons with state-of-the-art methods such as deep adaptive design (DAD) and policy gradient sequential optimal design (PG-SOED). These highlight the effectiveness of Laplace policies, particularly in settings where the posterior is approximately Gaussian.
