A proper calibration of trust in automation is imperative to achieve optimal overall performance in human-machine systems. Previous research has suggested that human operator trust could be influenced by various situational and dispositional factors, as well as operator self-confidence. It is critical to examine what traits and factors will influence how likely a person is to trust autonomous vehicles as they become more prevalent on today's roadways. The goal of this study was to further examine the relationship between individuals' level of self-confidence in their own driving abilities and their reported trust in automation when driving semi-autonomous cars. It was hypothesized that self-confidence and level of automation would be significant predictors of participants' trust. A total of 314 participants read through a series of vignettes describing several driving scenarios and completed an online assessment that measured both their trust and self-confidence in relation to autonomous driving functions. A series of multiple regression analyses showed that driving self-confidence was a significant predictor of operator trust when using level 1 automation. Results also indicated that gender was found to be a significant predictor across all levels of automation. These results suggest that self-confidence could be good a predictor of how individuals will respond to an automated system, which may have the potential to be generalized for implementation in training and selection environments. A series of repeated measures ANOVAs were conducted to determine the effect level of automation had on trust responses. Results indicated that trust levels significantly decreased as the automation levels increased. Theoretical and practical implications are discussed. These results can inform future research that aims to determine what makes an individual more likely to accept new technologies and help those creating autonomous vehicles design features and functionality that is more likely to be trusted and effectively utilized in on-road environments.

Image compression is a technique to reduce overall data size, but its effects on human perception have not been clearly established. The purpose of this effort was to determine the most effective psychophysical method for subjective image quality assessment, and to apply those findings to an objective algorithm. This algorithm was used to identify the minimum level of texture compression noticeable to the human, in order to determine whether compression-induced texture distortion impacted game-play outcomes. Four experiments tested several hypotheses. The first hypothesis evaluated which of three magnitude estimation (ME) methods (absolute ME, absolute ME plus, or ME with a standard) for image quality assessment was the most reliable. The just noticeable difference (JND) point for textures compression against the Feature Similarity Index for color was determined The second hypothesis tested whether human participants perceived the same amount of distortion differently when textures were presented in three ways: when textures were displayed as flat images; when textures were wrapped around a model; and when textures were wrapped around models and in a virtual environment. The last set of hypotheses examined whether compression affected both subjective (immersion, technology acceptance, usability) and objective (performance) gameplay outcomes. The results were: the absolute magnitude estimation method was the most reliable; no difference was observed in the JND threshold between flat textures and textures placed on models, but textured embedded within the virtual environment were more noticeable than in the other two presentation formats. There were no differences in subjective gameplay outcomes when textures were compressed to below the JND thresholds; and those who played a game with uncompressed textures performed better on in-game tasks than those with the textures compressed, but only on the first in-game day. Practitioners and researchers can use these findings to guide their approaches to texture compression and experimental design.

There exists considerable research concerning how humans attribute fault to each other, both in cases of accidents and those instances of intentional harm. There also exist studies involving blame attribution towards robots, when such robots have caused harm through operational failure or lack of safety features. However, relatively little work has, to date, examined the ways in which fault is attributed to self-driving vehicles involved in collisions, despite many newspaper and popular articles which both report past incidents and warn of future risk. This dissertation examined fault attribution in collisions involving autonomous vehicles by conducting three separate experiments. The first experiment placed participants in the roles of witnesses to a collision, and compared fault attributed to an autonomous vehicle to fault attributed to a regular, manually-operated vehicle, when those cars were involved in identical collisions. The second, and third experiments explored the fault that operators attributed to both themselves and autonomous vehicles when involved in a collision, whether they were the operator of the autonomous vehicle or the operator of a regular car that shared the road with automated ones. Results showed that, across experiments, perceived avoidability of the collision was the largest predictor of fault regardless of whether the participant was a witness or a driver. Additionally, participants in all three experiments thought themselves in general to be better than average drivers.

This work is designed to address the questions as to what drives and collapses trust between a human and a robot. Such information is needed to properly design automated decision aids. Human-robot trust (HRT) has traditionally been measured by questionnaires, which can be subject to lack of participant understanding, disengagement, and dishonesty. Therefore, implicit measures of trust are needed to measure HRT. The goal here is to identify neuro-physiological underpinnings (implicit measures) for HRT to assist designers in the development of automated robotic aids. More specifically, experiment one, looked to determine the effects of witnessing robot error on skin conductance response (SCR) and heart rate variability (HrV). The second experiment complemented this first procedure by determining the effects of witnessing robot error on Event Related Potentials (ERPs). Each experiment employed situations which previously have been empirically demonstrated to elicit a trust change in human participants. Both studies included two different robot reliability rates in a within subject design. Reliability consisted of each robot identifying civilians at either 95% reliability or 75% reliability. Self-reported dependent measures were perceptional robot reliability, trust questionnaires, a stress measure and a cognitive workload measure. Neurological and physiological dependent variables included SC, HrV, and ERPs. Heart rate variability did not demonstrate any evident changes based on robot reliability. In addition, SC demonstrated mixed changes based on robot reliability. However, ERP measures showed predictable changes based on robot reliability. None of the measures significantly correlated to changes in trust.

Multiple stress pathways can impact economic decision making. Two of these stress systems are the SAM axis and HPA axis. Prior research suggested that these pathways have the potential to exert independent alterations to economic decision performance, each with its own distinctive time course. In addition, stress has been found in general to alter salience (how important information is), but how this effect impacts specific economic dimensions (e.g. magnitude, loss or gain) has yet to be tested. Finally, the literature does not use the temporal profiles of the SAM and HPA axes to determine economic performance differences related to specific stress pathway responses. To address how the salience of different dimensions of financial information changes under stress and the impact of the SAM and HPA axes on economic decision making, the current study aimed to: 1) Examine the effects of acute stress on economic decision making based on the specific temporal profiles of each stress pathway, and 2) Examine the qualitatively distinct effects of stress on decision making from each pathway. The established temporal pattern of the SAM axis first, HPA axis later determined neural responses to an acute stressor during a given time block. In Experiment 1, the salience of individual dimensions of financial information was directly tested (e.g., magnitude, domain) to determine if basic components lead to performance differences, and/or were stress pathway dependent. While economic dimensions impacted behavior, stress did not affect how economic dimensions are responded to. Experiment 2 investigated the impact of each pathway on more complex financial behaviors (e.g. loss aversion and risk aversion). Risk aversion was decreased under SAM axis activation only. Loss aversion increased as an effect of time but was not dependent on stress. The results indicated economic dimensions and timing, including immediate neural stress pathway activation, have significant impacts on financial behaviors.

The goal of this dissertation was to study the effects of state and trait anxiety on explicit and implicit category learning. It was hypothesized that participants with higher state anxiety scores would require more trials to learn the explicit rule learning task compared to participants with lower state anxiety scores. On the other hand, high state anxiety participants were expected to excel in the implicit rule learning task relative to participants with low state anxiety scores. The hypotheses were informed by two theories, COVIS and ACT. The ACT theory states that there are three major mechanisms of executive functions that worsen with increasing anxiety. The COVIS theory states that explicit and implicit category learning rely on separate structures of the brain and, therefore, differently affected by anxiety. In experiment 1, participants completed implicit and explicit category learning tasks in either the control condition or the pressure condition. In the pressure manipulation group, participants completed a mortality salience writing task and were told they had a partner relying on their success in learning the categorization rule for both to receive a reward to induce anxiety. While the control participants completed a neutral writing task and were offered a reward solely based on their performance. In experiment 2, the study design was same as experiment 1 except for the addition of neuroimaging during category learning. Manipulating pressure during category learning replicated earlier research showing worsened performance in explicit rule learning under pressure, but no effect for implicit rule learning. In general, there was evidence that category learning was better in participants with high state anxiety scores, contradicting predictions based on ACT theory.

Learning to categorize visual stimuli is a fundamental cognitive skill underlying both everyday functioning and professional competencies in domains such as radiology and airport security screening. Categories may be very simple or highly complex, with accurate categorization dependent on multiple interacting features. General recognition theory (GRT) models uniquely allow examination of feature dimension interactions, but basic questions remain about the applicability of such models and the 2x2 categorization tasks (four-alternative forced choice) employed in studies which use them. Findings in several studies that factorially combine 2 levels of 2 stimulus dimensions indicate a common pattern of perceptual advantage for the category that is high on both dimensions, despite examining stimuli as diverse as simulated human faces, baggage x-rays, and mammograms. Because of the ambiguous ground truth of these applied studies, their conclusions are limited by the inability to rule out the influence of task artifacts on their results. The present work fills this gap in the literature and seeks to disambiguate such findings by examining the contributions of task artifacts such as response mapping and assessing the sensitivity of the modeling paradigm using simple stimuli. Participants learned categories of simple two-dimensional stimuli produced by various manipulations of a basic category construction, and GRT-wIND models were fit to their responses. Results indicate that the model is sensitive to manipulations of the perceptual space and category structures. Further, the previously observed pattern advantaging one of four categories is observed here despite the absence of such a relationship between the feature dimensions in the objective category constructions. The effect is largely mitigated, however, by altering the response locations such that they are no longer orthogonally mapped to their corresponding categories. These findings further evidence the utility and sensitivity of the GRT-wIND model and suggest updates to best practices in applying the four-alternative forced choice task.

As gesture-based interactions with computer interfaces become more technologically feasible for educational and training systems, it is important to consider what interactions are best for the learner. Computer interactions should not interfere with learning nor increase the mental effort of completing the lesson. The purpose of the current set of studies was to determine whether natural gesture-based interactions, or instruction of those gestures, help the learner in a computer lesson by increasing learning and reducing mental effort. First, two studies were conducted to determine what gestures were considered natural by participants. Then, those gestures were implemented in an experiment to compare type of gesture and type of gesture instruction on learning conceptual information from a computer lesson. The goal of these studies was to determine the instructional efficiency – that is, the extent of learning taking into account the amount of mental effort – of implementing gesture-based interactions in a conceptual computer lesson. To test whether the type of gesture interaction affects conceptual learning in a computer lesson, the gesture-based interactions were either naturally- or arbitrarily-mapped to the learning material on the fundamentals of optics. The optics lesson presented conceptual information about reflection and refraction, and participants used the gesture-based interactions during the lesson to manipulate on-screen lenses and mirrors in a beam of light. The beam of light refracted/reflected at the angle corresponding with type of lens/mirror. The natural gesture-based interactions were those that mimicked the physical movement used to manipulate the lenses and mirrors in the optics lesson, while the arbitrary gestures were those that did not match the movement of the lens or mirror being manipulated. The natural gestures implemented in the computer lesson were determined from Study 1, in which participants performed gestures they considered natural for a set of actions, and rated in Study 2 as most closely resembling the physical interaction they represent. The arbitrary gestures were rated by participants as most arbitrary for each computer action in Study 2. To test whether the effect of novel gesture-based interactions depends on how they are taught, the way the gestures were instructed was varied in the main experiment by using either video- or text-based tutorials. Results of the experiment support that natural gesture-based interactions were better for learning than arbitrary gestures, and instruction of the gestures largely did not affect learning and amount of mental effort felt during the task. To further investigate the factors affecting instructional efficiency in using gesture-based interactions for a computer lesson, individual differences of the learner were taken into account. Results indicated that the instructional efficiency of the gestures and their instruction depended on an individual's spatial ability, such that arbitrary gesture interactions taught with a text-based tutorial were particularly inefficient for those with lower spatial ability. These findings are explained in the context of Embodied Cognition and Cognitive Load Theory, and guidelines are provided for instructional design of computer lessons using natural user interfaces. The theoretical frameworks of Embodied Cognition and Cognitive Load Theory were used to explain why gesture-based interactions and their instructions impacted the instructional efficiency of these factors in a computer lesson. Gesture-based interactions that are natural (i.e., mimic the physical interaction by corresponding to the learning material) were more instructionally efficient than arbitrary gestures because natural gestures may help schema development of conceptual information through physical enactment of the learning material. Furthermore, natural gestures resulted in lower cognitive load than arbitrary gestures, because arbitrary gestures that do not match the learning material may increase the working memory processing not associated with the learning material during the lesson. Additionally, the way in which the gesture-based interactions were taught was varied by either instructing the gestures with video- or text-based tutorials, and it was hypothesized that video-based tutorials would be a better way to instruct gesture-based interactions because the videos may help the learner to visualize the interactions and create a more easily recalled sensorimotor representation for the gestures; however, this hypothesis was not supported and there was not strong evidence that video-based tutorials were more instructionally efficient than text-based instructions. The results of the current set of studies can be applied to educational and training systems that incorporate a gesture-based interface. The finding that more natural gestures are better for learning efficiency, cognitive load, and a variety of usability factors should encourage instructional designers and researchers to keep the user in mind when developing gesture-based interactions.

Incident reporting systems are an integral part of any organization seeking to increase the safety of their operation by gathering data on past events, which can then be used to identify ways of mitigating similar events in the future. In order to analyze trends and common issues with regards to the human element in the system, reports are often classified according to a human factors taxonomy. Lately, machine learning algorithms have become popular tools for automated classification of text; however, performance of such algorithms varies and is dependent on several factors. In supervised machine learning tasks such as text classification, the algorithm is trained with features and labels, where the features here are a function of the incident reports themselves and the labels are supplied by a human annotator, whether that is the reporter or a third person. Aside from the intricacies of building and tuning machine learning models, a subjective classification according to a human factors taxonomy can generate considerable noise and bias. I examined the interdependencies between the features of incident reports, the subjective labeling process, the constraints that the taxonomy itself imposes, and basic characteristics of human factors taxonomies that can influence human, as well as automated, classification. In order to evaluate these challenges, I trained a machine learning classifier on 17,253 incident reports from the NASA Aviation Safety Reporting System (ASRS) using multi-label classification, and collected labels from six human annotators for a subset of 400 incident reports each, resulting in a total of 2,400 individual annotations. Results show that, in general, reliability of annotation for the set of incident reports selected in this study was comparatively low. It was also evident that some human factors labels were more agreed upon than others, sometimes related to the presence of key words in the reports which map directly to the label. Performance of machine learning annotation followed patterns of human agreement on labels. The high variability of content and quality of narratives has been identified as a major factor for difficulties in annotation. Suggestions on how to improve the data collection and labeling process are provided.

Phishing emails have become a prevalent cybersecurity threat for the modern email user. Research attempting to understand how users are susceptible to phishing attacks has been limited and hasn't fully explored how task factors influence accurate detection. Even further lacking are the existing training interventions that still have users falling victim to up to 90% of phishing emails following training. The present studies examined how task factors (e.g., email load, phishing prevalence) and a new form of intervention, rather than training, influence email performance. In four experiments, participants classified emails as either legitimate or not legitimate and reported on a variety of other categorizations (e.g., threat level). The first two experiments examined how email load and phishing prevalence influence phishing detection. The third experiment examined the interaction of these two factors to determine whether they have compounding effects. The last experiment investigated how performance can be improved with a novel cheat sheet intervention method. All four experiments utilized individual difference variables to examine how cognitive, behavioral, and personality factors influence detection under various task conditions and how they impact the utilization of training interventions. The results across the first three experiments indicated that both high email load and low phishing prevalence decrease email classification accuracy and sensitivity. However, performance was poor across all conditions, with phishing detection near chance performance and sensitivity values indicating that the task was very challenging. Additionally, participants demonstrated poor metacognition with over confidence, low self-reported difficulty, and low perceived threat for the emails. Experiment 4's results indicated that phishing detection could be improved by 20% with the embedded cheat sheet intervention. Overall, the present studies suggest that email load and phishing prevalence can decrease fraud detection, but that embedded phishing tips can improve performance.