Safety in the Face of Unknown Unknowns: Algorithm Fusion in Data-driven Engineering Systems

Abstract

Most current machine learning algorithms make highly confident yet incorrect classifications when faced with unexpected test samples from an unknown distribution different from training; such epistemic uncertainty (unknown unknowns) can have catastrophic safety implications. In this conceptual paper, we propose a method to leverage engineering science knowledge to control epistemic uncertainty and maintain decision safety. The basic idea is an algorithm fusion approach that combines data-driven learned models with physical system knowledge, to operate between the extremes of purely data-driven classifiers and purely engineering science rules. This facilitates the safe operation of data-driven engineering systems, such as wastewater treatment plants.