Astra-discovered kernel reportedly outperforms the best tuned standard kernel on simulated glucose data

A researcher sharing the paper says its approach learns positive semidefinite kernels for Gaussian processes, going beyond earlier methods that combine kernels from a fixed library. The researcher says a structured kernel discovered by Astra outperformed the best tuned standard kernel at predicting glucose levels using UVA/Padova type-1 diabetes simulator data. They also describe the discovered kernel as interpretable, with 16 scalar functions of inputs specifying meal and bolus size and time.