Nesterov-aided Stochastic Gradient Methods using Laplace Approximation for Bayesian Design Optimization

TL;DR

This paper introduces Nesterov-aided stochastic gradient methods combined with Laplace approximation estimators to efficiently solve Bayesian experimental design problems, significantly reducing model evaluations needed for complex models.

Contribution

The paper develops a novel combination of accelerated stochastic gradient methods with Laplace-based estimators for Bayesian design optimization, enabling efficient handling of complex models.

Findings

Accelerated stochastic gradient descent with MCLA converges faster than traditional methods.

Laplace-based estimators enable the use of complex models like PDEs.

Significant reduction in model evaluations, up to five orders of magnitude.

Abstract

Finding the best setup for experiments is the primary concern for Optimal Experimental Design (OED). Here, we focus on the Bayesian experimental design problem of finding the setup that maximizes the Shannon expected information gain. We use the stochastic gradient descent and its accelerated counterpart, which employs Nesterov's method, to solve the optimization problem in OED. We adapt a restart technique, originally proposed for the acceleration in deterministic optimization, to improve stochastic optimization methods. We combine these optimization methods with three estimators of the objective function: the double-loop Monte Carlo estimator (DLMC), the Monte Carlo estimator using the Laplace approximation for the posterior distribution (MCLA) and the double-loop Monte Carlo estimator with Laplace-based importance sampling (DLMCIS). Using stochastic gradient methods and Laplace-based estimators together allows us to use expensive and complex models, such as those that require solving partial differential equations (PDEs). From a theoretical viewpoint, we derive an explicit formula to compute the gradient estimator of the Monte Carlo methods, including MCLA and DLMCIS. From a computational standpoint, we study four examples: three based on analytical functions and one using the finite element method. The last example is an electrical impedance tomography experiment based on the complete electrode model. In these examples, the accelerated stochastic gradient descent method using MCLA converges to local maxima with up to five orders of magnitude fewer model evaluations than gradient descent with DLMC.