Learning Knowledge of Agency

Justin Brody, Samuel Levin, Samuel Shapiro

Abstract: In this paper, we will study the theoretical foundations for operationalizing an agent's knowledge of agency - that is an agent's knowledge of its own actions and their effects in a dynamic environment. Our main concern will be to develop theoretical foundations and algorithms which will enable a grounded knowledge of agency; these will be empirically evaluated in future work.Our approach will employ ideas from enactive cognition [27] and predictive coding [8]. Specifically, we will equate learning knowledge of agency with making successful predictions about the expected sensory feedback that will result from specific actions.