EFFECTIVE TASK ALLOCATION FOR AD HOC HUMAN-AGENT TEAMS

Abstract

Due to the rapid advancement in autonomous agent capabilities, mixed human-agent teams will be increasingly deployed for new applications in our personal as well as professional spheres. Researchers, educators, application developers, and policymakers are evaluating the design, effectiveness, and usability of such human-agent teams and their role in our lives. We envision humans and agents working together in collaborative environments to enrich human lives, support human self-efficacy, increase satisfaction and well-being, improve decision making, and enhance team performance and productivity. With increased connectivity and the transition to knowledge economies, new collaboration environments are emerging. Often, such environments are ad hoc in nature, where humans and autonomous systems have to collaborate without pre-coordination. Such scenarios require the use of new collaboration strategies, particularly those that support rapid adaptation and learning. Therefore, a thorough examination of factors that affect collabo ration efficacy in these types of environments is warranted. A critical part of effective team collaboration is the process and information used to allocate tasks to team members. Effective allocation of tasks to them is particularly challenging in ad hoc teamwork, as prior knowledge about teammate capabilities is either absent or minimal but is critical for the viability and success of such teams. We study task allocation scenarios in which ad hoc and virtual human-agent teams collaborate and adapt over a few interactions to find allocations to improve team effectiveness, measured in terms of team performance and human participant satisfaction. We designed and implemented CHATBoard, a collaborative environment to facilitate task allocation in teams. We developed task allocation protocols in which humans and agents play varying task allocation roles in harnessing the potential of team members with different expertise distribution over task types. We compare team performance and participant satisfaction in experiments with human and agent task allocators in diverse scenarios involving intellective tasks. We vary team composition in terms of the number of team members and if the human believes that the teammate is another human or an autonomous agent. We also investigate team effectiveness for different distributions of agent expertise, including those that are complementary and similar to their human teammates. We develop and evaluate associated research hypotheses through experiments with human participants recruited on the Amazon Mechanical Turk crowdsourcing platform. Our investigation of ad hoc human-agent teams showed that agent task allocators can better leverage the potential of the team, compared to human task allocators, producing higher performance. Humans, however, are less satisfied with the protocol when the agent allocates tasks. To improve human satisfaction without undermining team performance, we investigate alternative allocation mechanisms and approaches. These include allocation protocols providing more control and input for human teammates and approaches that vary agent characteristics and strategies, such as agents providing explanations and guidance. This dissertation addresses a novel area in the literature: how to allocate tasks without pre-coordination so that existing expertise in team members is maximally realized while ensuring the satisfaction of human team members. Our findings herein provide novel and significant insights for researchers and designers developing ad hoc and general human-agent teams.