Relevance and Redundancy as Selection Techniques for Human-Autonomy Sensor Fusion

Human-autonomy teaming through the use of physiological sensors poses a novel sensor fusion problem due to the dynamic nature of human physiological sensor models and the difficulty of quantifying their variability across subjects and time. Analytical techniques for developing these models thus depend on objective criteria for selecting and weighting sensors in the fusion process. We draw on feature selection methodologies that employ dual intuitions: (1) maximizing the relevance between sensors and the target classes will enhance overall performance within a given fusion scheme while (2) minimizing redundancy amongst the selected sensors will not harm it and may even help to increase relevance. We analyze these intuitions in the context of a human-autonomy image classification task. While we find strong evidence for the relevance hypothesis, we show that the redundancy hypothesis may be fusion-dependent. This relationship and its implications for human-autonomy sensor fusion is explored within a framework employing Naive Bayes fusion, Dempster Shafer theory (DST), and the related Dynamic Belief Fusion.