Optimizing Hybrid Human-AI Decision-making with Reinforcement Learning

With autonomous systems continuing to demonstrate increased capability, we can anticipate a continued increase in integration of these systems into daily life, resulting in a corresponding increase in interactions. Therefore, this thesis investigates concepts relating to the collaboration and/or coordination of teams of humans and autonomous systems. I propose the use of models accounting for the diverse, and potentially erroneous, behavior of humans and autonomous systems to maximize their collective performance via a Reinforcement Learning based manager. First, the use of cognitively inspired models is investigated to generate human-like reasoning and use of experience. Next, the scenario is shifted to a 2D driving case to generate a case with continuous state and action spaces for increased complexity. Expanding the team characteristics supported by the manager, the next aspect of the thesis considers a team with diversity in the actions/behaviors available to the team members. The final portion of the thesis is dedicated to my work which increased isolation of the manager and team models and added support for operation across multiple time steps. In this case, the manager only selects an agent for delegation of authority when the manager’s constraints indicate the need for intervention. The resulting model demonstrates strong isolation between the manager and team, with the added benefit of support for delegations which span multiple time steps.