A. L. Thomaz, C. Breazeal

Abstract

In this paper we advocate a paradigm of socially guided machine learning, designing agents that take better advantage of the situated aspects of learning. We augmented a standard Reinforcement Learning agent with the social mechanisms of attention direction and gaze. Experiments with an interactive computer game, deployed over the World Wide Web to over 75 players, show the positive impact of these social aspects. Allowing the human to direct the agent’s attention creates a more efficient exploration strategy. Additionally, gaze behavior lets the learning agent improve its own learning environment, using transparency to steer the human’s instruction.