2017/05/23 by Antony M. Overstall, Overstall, Antony, David C. Woods +4
Computer Science · Decision Sciences · #Advanced Multi-Objective Optimization Algorithms #Computation (stat.CO) #FOS: Computer and information sciences #Gaussian Processes and Bayesian Inference #Optimal Experimental Design Methods
paper · pdf · doi:10.48550/arxiv.1705.08096
openalex publication_date 2017/05/23 · openalex created_date 2022/10/02 · openalex updated_date 2026/07/28
We describe the R package acebayes and demonstrate its use to find Bayesian\noptimal experimental designs. A decision-theoretic approach is adopted, with\nthe optimal design maximising an expected utility. Finding Bayesian optimal\ndesigns for realistic problems is challenging, as the expected utility is\ntypically intractable and the design space may be high-dimensional. The package\nimplements the approximate coordinate exchange algorithm to optimise (an\napproximation to) the expected utility via a sequence of conditional\none-dimensional optimisation steps. At each step, a Gaussian process regression\nmodel is used to approximate, and subsequently optimise, the expected utility\nas the function of a single design coordinate (the value taken by one\ncontrollable variable for one run of the experiment). In addition to functions\nfor bespoke design problems with user-defined utility functions, acebayes\nprovides functions tailored to finding designs for common generalised linear\nand nonlinear models. The package provides a step-change in the complexity of\nproblems that can be addressed, enabling designs to be found for much larger\nnumbers of variables and runs than previously possible. We provide tutorials on\nthe application of the methodology for four illustrative examples of varying\ncomplexity where designs are found for the goals of parameter estimation, model\nselection and prediction. These examples demonstrate previously unseen\nfunctionality of acebayes.\n