Model of Human-Machine Trust Dynamics

Another piece that was developed as part of my Ph.D. candidacy advanced examination.  This artifact was created while addressing the dialogue that takes place between automated systems and human operators and the cycle of trust evolution that results.  The illustration presumes that trust updates based only on positive and negative experiences (as has been previously shown).  This illustration is important to highlight the varied effect that delayed feedback can have on attitudes of trust; however it is greatly simplified in that it does not consider additional factors that could influence the degree of change when receiving feedback on one’s decision to rely.  For example, I would hypothesize that the degree of risk associated with relying on the system and the perceived separation between expected performance and actual performance represent two factors that would also likely play into the degree of influence that receiving feedback on one’s decisions could yield.

Framework of Trust and Reliance in Automation

As part of my Ph.D. candidacy I was required to take an Advanced Examination where each student receives individual questions based on his or her intended dissertation topic.  The questions are intended to motivate the student to learn about areas that his or her advisers feel are gaps in the student’s understanding of the problem domain or dissertation method.  The questions are very theoretical and typically require (very long) written responses based on extensive literature reviews to formulate logical responses.

For my A-Exam, I received 5 questions (some with multiple parts) and was required to respond to them in a 2 week period (pretty standard for my department/major combination).  At the time of my A-Exam my intended focus was on dynamic trust in automated systems so one of my questions was to characterize the effects that real-world dynamics are likely to have on human trust and reliance in automated systems.  Luckily, I bounded this discussion within the domains of intelligence analysis as it served as  perfect test bed for both my intended dissertation focus and for highlighting the factors that influence trust and reliance.

(click to enlarge)

Incorporating situationally qualified human observations into a fusion process for intelligence analysis

Since my second year at UB my research assistantship has been funded by a 3 year multi-university research initiative (MURI) grant focused on network based hard/soft information fusion.  The first year of this research (for me at least) was focused on learning about how accurate people are at observing different phenomena in the world given specific contextual factors.  For example, how accurately can people judge distances?  What if it’s nighttime?  What if it’s snowing? etc.

This work was developed to allow human observations to be appropriately characterized in terms of their error/bias so that they could be integrated into a data fusion process along with hard data (which comes from things like radar sensors that are highly calibrated).  The poster below, which was created for a UB School of Engineering and Applied Sciences poster competition, describes the results of these research efforts at a high level.

Evaluating and Interpreting the Creation of Causal Influence Models

Below (flash required for viewing)  is a poster submission that highlights some of the research I helped to conduct in my first year at UB.  The research objective was to investigate users’ abilities to create Causal Influence Networks, a class of Bayesian networks, with varying complexity and types of relationships, using a causal influence model.  Funding for this research was generously provided by Charles River Analytics.

Abstract: Bayesian networks (BNs) are probabilistic models used to reason under uncertainty by graphically expressing domain knowledge in order to reason about states, causes, and effects. While BNs have many advantages, their complexity can hamper the process of knowledge elicitation and encoding. For example, BNs require the definition of a priori, conditional probabilities: as complex models increase in size, this requires eliciting exponential numbers of complex probabilities. Multiple “canonical modeling” approaches, such as Causal Influence Models (CIMs), have been developed to address these complexities. However, little progress has been made towards human-in-the-loop evaluation of such approaches – specifically, their accessibility and usability, their related user interfaces, and how they enable a user to correctly create and interpret variables and probabilistic relationships. In this study, we evaluated the CIM approach (implemented in a software application) to determine the effect on user task performance. Results indicate that the model complexity has an adverse effect on performance when users are interpreting an existing model; that semantics of a model may impact performance; and that users were generally successful in creating new models of different situations.