Enhancing Human Predictive Understanding of AI Agents Through Variation Theory of Learning

To stay safe and effective when collaborating with a cobot or an AI agent, people must be able to predict the future behaviors of their automated partners. We propose using the Variation Theory of Learning, a theory of how humans learn new concepts, to allow people to predict agent behaviors by building conceptual models of agent policies. In this work, we explore the space of design decisions needed to operationalize Variation Theory and how to best to scaffold peoples' experiences of interacting with agents to inform their conceptual model development. We study this operationalization by analyzing a simulated highway driving environment. We find evidence that operationalizing Variation Theory can assist people in identifying a given agent's behavior in novel settings, an intermediary task en route to measure the promise of applying Variation Theory to people predict new agent behaviors.