Learning to Play with Intrinsically-Motivated Self-Aware Agents

摘要: Infants are experts at playing, with an amazing ability to generate novel structured behaviors in unstructured environments that lack clear extrinsic reward signals. We seek mathematically formalize these abilities using a neural network implements curiosity-driven intrinsic motivation. Using simple but ecologically naturalistic simulated environment which agent can move and interact objects it sees, we propose "world-model" learns predict the dynamic consequences of agent's actions. Simultaneously, train separate explicit "self-model" allows track error map its own world-model, then uses self-model adversarially challenge developing world-model. demonstrate this policy causes explore informative interactions environment, leading generation spectrum complex behaviors, including ego-motion prediction, object attention, gathering. Moreover, world-model supports improved performance on dynamics detection, localization recognition tasks. Taken together, our results initial steps toward creating flexible autonomous agents self-supervise physical environments.