Efficient construction of Bayes optimal designs for stochastic process models

TL;DR

This paper introduces a fast, parallelizable algorithm for constructing Bayes optimal designs in stochastic process models, significantly reducing computational effort compared to traditional methods like Muller’s algorithm.

Contribution

The paper presents a novel, efficient scheme for optimal design construction that outperforms existing methods in speed and computational resource requirements.

Findings

Requires up to ten times fewer utility evaluations

Performs well across models of increasing complexity

Leverages parallel computing for efficiency

Abstract

Stochastic process models are now commonly used to analyse complex biological, ecological and industrial systems. Increasingly there is a need to deliver accurate estimates of model parameters and assess model fit by optimizing the timing of measurement of these processes. Standard methods to construct Bayes optimal designs, such as the well known \Muller algorithm, are computationally intensive even for relatively simple models. A key issue is that, in determining the merit of a design, the utility function typically requires summaries of many parameter posterior distributions, each determined via a computer-intensive scheme such as MCMC. This paper describes a fast and computationally efficient scheme to determine optimal designs for stochastic process models. The algorithm compares favourably with other methods for determining optimal designs and can require up to an order of magnitude fewer utility function evaluations for the same accuracy in the optimal design solution. It benefits from being embarrassingly parallel and is ideal for running on multi-core computers. The method is illustrated by determining different sized optimal designs for three problems of increasing complexity.