Abstract
The overall safety and reliability of critical systems may be improved if interfaces can be tailored to the current cognitive states of their operators. For this to be realized, online measures of cognitive workload need to be developed. This dissertation proposes that cognitive measures based on physiological indicators provides the most potential in real world environments where task performance is difficult to quantify and operators may not be able to periodically self-report their workload. Here, the primary aim is the development and evaluation of algorithms for identifying cognitive workload from multiple relatively unobtrusive physiological measures using wavelet decomposition and machine learning. To support his primary aim, a tracking task was developed that allowed workload difficulty to be subtly and continuously manipulated in a systematic fashion. This manipulation was validated against subjective ratings of workload as well as with a secondary random number generation task. After establishing a means of controlling workload difficulty pupil diameter, skin conductance, heart rate, and heart rate variability were recorded and used in conjunction with machine learning to build classifiers of workload difficulty. Applying discrete wavelet decomposition to physiological measures and training classifiers to specific individuals yielded algorithms that could classify workload difficulty and derivative workload difficulty with enough accuracy to support practical applications (> 90% accuracies).