Toward human-multiagent teams

One of the most fundamental challenges of building a human-multiagent team is adjustable autonomy, a process in which the control over team decisions is dynamically transferred between humans and agents. This thesis studies adjustable autonomy in the context of a human interacting with a team of agents and focuses on four issues that arise when addressing this team-level adjustable autonomy problem in real-time uncertain domains. Firstly, the humans and agents may differ significantly in their worldviews and their capabilities. This difference leads to inconsistencies in how humans and agents solve problems. Despite such inconsistencies, previous work has rigidly assumed the infallibility of human decisions. However, in some cases following the human's decisions lead to worse human-multiagent team performance. Secondly, it is desirable for the team to manage the uncertainty of action durations and plan for the optimal action at any point in time. This is a crucial challenge to address given that these human-multiagent teams are working in real-time with strict deadlines combined with the particularly uncertain duration of actions that involve a human. Thirdly, the team needs to be able to plan for the optimal time to interrupt certain actions. This is due to the fact that actions may take an uncertain amount of time and the deadline is approaching. The human-multiagent team may benefit from attempting an action for a given amount of time and interrupting the action if it does not finish in order to try another action that has a higher expected reward. Fourthly, team-level adjustable autonomy is an inherently distributed and complex problem that cannot be solved optimally and completely online. My thesis makes four contributions to the field in order to address these challenges. First, I have included, in the adjustable autonomy framework, the modeling of the resolution of inconsistencies between human and agent view. This diverges from previous work on adjustable autonomy that traditionally assumes the human is infallible and decisions as rigid, but instead puts the humans and agents on an equal footing, allowing each to identify possible team performance problems. I have developed new "resolution adjustable autonomy strategies" that recognize inconsistencies and provide a framework to decide if a resolution is beneficial. Second, in order to address the challenges brought about by dealing with time, I have modeled these new adjustable autonomy strategies using TMDPs (Time dependent Markov Decision Problems). This allows for an improvement over previous approaches, which used a discretized time model and less efficient solutions. Third, I have introduced a new model for Interruptible TMDPs (ITMDPs) that allows for an action to be interrupted at any point in continuous time. This results in a more accurate modeling of actions and produces additional time-dependent policies that guide interruption during the execution of an action. Fourth, I have created a hybrid approach that decomposes the team level adjustable autonomy problem in a separate ITMDP for each team decision. In addition, team-based logics are used to coordinate and execute the team actions that are present in the ITMDP. In addition to developing these techniques, I have conducted experimental evaluations that demonstrate the contributions of this approach. This has been realized in a system that I have constructed, DEFACTO (Demonstrating Effective Flexible Agent Coordination of Teams via Omnipresence), that incorporates this approach to team-level adjustable autonomy, along with agent coordination reasoning and a multi-perspective view of the team for the human. DEFACTO has been applied to an urban disaster response domain and used for incident commander training. The Los Angeles Fire Department has been supportive and has given valuable feedback that has shaped the system.