An Infomax Controller for Real Time Detection of Social Contingency

摘要

We present a model of behavior according to which organisms react to the environment in a manner that maximizes the information gained about events of interest. We call the approach "Infomax control" for it combines the theory of optimal control with information maximization models of perception. The approach is reactive, not cognitive, in that it is better described as a continuous "dance" of actions and reactions with the world, rather than a turn-taking inferential process like chess-playing. The approach however is intelligent in that it produces behaviors that optimize long-term information gain. We illustrate how Infomax control can be used to understand the detection of social contingency in 10 month old infants. The results suggest that, while lacking language, by this age infants actively "ask questions" to the environment, i.e., schedule their actions in a manner that maximizes the expected information return. A real time Infomax controller was implemented on a humanoid robot to detect people using contingency information. The system worked robustly requiring little bandwidth and computational cost