A Principled Method for the Creation of Synthetic Multi-fidelity Data Sets

Multifidelity and multioutput optimisation algorithms are an area of current interest in many areas of computational design as they allow experimental and computational proxies to be used intelligently in the search for optimal species. Characterisation of these algorithms involves benchmarks that typically either use analytic functions or existing multifidelity datasets. Unfortunately, existing analytic functions are often not representative of relevant problems, while many existing datasets are not constructed to easily allow systematic investigation of the influence of characteristics of the contained proxies functions. To fulfil this need, we present a methodology for systematic generation of synthetic fidelities derived from a reference ground truth function with a controllable degree of correlation.