Human-Machine Teamwork: An Exploration of Multi-Agent Systems, Team Cognition, and Collective Intelligence

One of the major ways through which humans overcome complex challenges is teamwork. When humans share knowledge and information, and cooperate and coordinate towards shared goals, they overcome their individual limitations and achieve better solutions to difficult problems. The rise of artificial intelligence provides a unique opportunity to study teamwork between humans and machines, and potentially discover insights about cognition and collaboration that can set the foundation for a world where humans work with, as opposed to against, artificial intelligence to solve problems that neither human or artificial intelligence can solve on its own. To better understand human-machine teamwork, it’s important to understand human-human teamwork (humans working together) and multi-agent systems (how artificial intelligence interacts as an agent that’s part of a group) to identify the characteristics that make humans and machines good teammates. This perspective lets us approach human-machine teamwork from the perspective of the human as well as the perspective of the machine. Thus, to reach a more accurate understanding of how humans and machines can work together, we examine human-machine teamwork through a series of studies. In this dissertation, we conducted 4 studies and developed 2 theoretical models: First, we focused on human-machine cooperation. We paired human participants with reinforcement learning agents to play two game theory scenarios where individual interests and collective interests are in conflict to easily detect cooperation. We show that different reinforcement models exhibit different levels of cooperation, and that humans are more likely to cooperate if they believe they are playing with another human as opposed to a machine. Second, we focused on human-machine coordination. We once again paired humans with machines to create a human-machine team to make them play a game theory scenario that emphasizes convergence towards a mutually beneficial outcome. We also analyzed survey responses from the participants to highlight how many of the principles of human-human teamwork can still occur in human-machine teams even though communication is not possible. Third, we reviewed the collective intelligence literature and the prediction markets literature to develop a model for a prediction market that enables humans and machines to work together to improve predictions. The model supports artificial intelligence operating as a peer in the prediction market as well as a complementary aggregator. Fourth, we reviewed the team cognition and collective intelligence literature to develop a model for teamwork that integrates team cognition, collective intelligence, and artificial intelligence. The model provides a new foundation to think about teamwork beyond the forecasting domain. Next, we used a simulation of emergency response management to test the different teamwork aspects of a variety of human-machine teams compared to human-human and machine-machine teams. Lastly, we ran another study that used a prediction market to examine the impact that having AI operate as a participant rather than an aggregator has on the predictive capacity of the prediction market. Our research will help identify which principles of human teamwork are applicable to human-machine teamwork, the role artificial intelligence can play in enhancing collective intelligence, and the effectiveness of human-machine teamwork compared to single artificial intelligence. In the process, we expect to produce a substantial amount of empirical results that can lay the groundwork for future research of human-machine teamwork