Evidential Reasoning with Expert-Guided Machine Learning

Abstract

Evidential reasoning aims to infer hidden causes from observed effects. In the context of fault detection, it is possible to trace the cause of anomalies by combining evidential reasoning with physical knowledge. However, experts may not have full knowledge of the physical systems, and data may not be large enough to learn from scratch. We develop an online evidential reasoning algorithm that blends machine learning with expert-provided physical knowledge about causal structure and functional form. The expert first represents possible causes and effects in the structure of a Bayesian Network, then provides physics-informed priors about the model relating failure modes to effects, allowing inference in the absence of strong supervision. As data are sampled from the physical system, predictions are generated using a quasi-Bayesian mixture of the expert's judgment and a data-driven estimate. With simulated datasets, we evaluate the conditions under which the system converges to correct causal inferences under weak supervision, and small amounts of strong supervision from the expert. We find that the approach is able to make accurate inferences with little or no data unless the expert's physical model is very incorrect or the signal to noise ratio across error modes is small.

Tasks

Reproductions