Engineering Shared Leadership for Human and Autonomy Collaboration in Multi-Agent Systems

Abstract

Autonomous technology has advanced rapidly in recent years, with intelligent systems demonstrating increasingly sophisticated capabilities in perception, decision-making, and adaptive behavior. These advancements have positioned autonomous agents to be teammates, enabling collaboration with humans in diverse domains and prompting emergence of Human-Autonomy Teaming (HAT) systems. HAT systems increasingly involve multiple autonomous agents working alongside humans in dynamic, high-stakes environments. HAT systems are often engineered with static hierarchical structures that predefine leadership authority for a set of tasks, thereby constraining their adaptability to shifting situational demands or unanticipated conditions, resulting in unintended degradation of collaboration and task performance. For dynamic environments, HAT systems require flexible or emergent leadership structures between agents. This dissertation investigates shared leadership for enabling flexible authority distribution between human and autonomous agents to enhance collaboration and performance in multi-agent systems composed of human and autonomous agents. The objectives of this research were (1) to understand how shared leadership functions in human teams can be adapted for multi-agent HAT systems, (2) to model leadership emergence from the human's perspective and identify factors governing the temporal patterns, and (3) to compare performance and perceived team dynamics between shared leadership and centralized leadership. vspace{0.1in} newline Study 1 was a systematic literature review of shared leadership in human teams for deriving mechanisms that can be engineered into HAT. The review revealed that humans rely on interpersonal trust and performance-based competence assessments for leadership distribution, with decentralization and mutual influence as the most influential mechanisms for enabling sharing leadership. The review also identified questionnaire-based assessments and network analysis as viable measurement approaches, with the latter also a viable approach for implementing shared leadership in HAT. These findings established the theoretical and methodological foundation for operationalizing and assessing shared leadership in HAT. Study 2 was an experiment recruiting human participants to complete a series of object-recognition tasks which involved assignments of multiple unmanned aerial vehicles (UAVs) in a simulated search and rescue context. Modeling the experimental data using network analysis, specifically in how the human's trust-competence perceptions of the autonomy evolve over time, revealed temporal patterns of leadership assignment. The study included the Trust-Competence-Identity Network (TCIN) that was developed to capture the humans' perception of agents across repeated task iterations. Logistic regression at the population level demonstrated that competence functioned as a capability-based predictor, while temporal exponential random graph models at the individual levels demonstrated that trust operated as an individualized experience-driven factor for predicting leadership assignment. The results provided foundational evidence supporting TCIN in predicting leadership emergence in HAT, illustrating the co-variation of key factors in human selection of autonomous agents as the leader. Study 3 was another experiment recruiting human participants to complete a series of object-recognition tasks that included conditions of the traditional centralized leadership and shared leadership for comparison of performance in multi-agent HAT. Study 3 also included a newly developed shared leadership questionnaire for HAT, adapted from validated instruments in human teams to measure leadership dynamics in HAT. Shared leadership demonstrated superior performance compared to centralized leadership, suggesting that distributing authority between humans and autonomous agents produces better outcomes than concentrating authority. Logistic regression at the population level demonstrated that trust moderated the rate at which complementary claiming-granting increased, while temporal exponential random graph models at the individual levels demonstrated that participants ultimately adopted complementary patterns. The shared leadership questionnaire also revealed that participants perceived more leadership distribution, team collaboration, and deference to expertise under shared leadership than the centralized leadership condition. These findings demonstrate that shared leadership in HAT involves both temporal learning processes and recognition of functional benefits that transcend individual differences in agent evaluation, establishing shared leadership as a viable organizational structure for multi-agent teams.