An Information-Theoretic Perspective on Acting Agents

Abstract Every embodied intelligent agent constantly interacts in its own way with its environment. It perceives information about the world, processes this information in its controller, e.g. its brain, and sends signals to the body, which then interacts with its surroundings. Here we give a thorough introduction to an information-theoretic approach that aims at quantifying these information flows among an agent’s body, brain and environment. Thereby we measure how much an information flow impacts the behavior of an agent. For instance, we might ask the question whether the brain is essential for a certain behavior or whether it is a reflexive stimuli response. In order to measure these interactions we model the system using the sensorimotor loop and assess the different information flows by information-theoretic measures. This includes measures related to Morphological Computation and the complexity of the controller. Morphological Computation describes the reduction of computational cost for the controller that results from the interaction with the environment. We include results from an implementation of this framework with simplistic simulated robots. Experiments with these robots show, for instance, that an agent with a well-adapted body does not rely as much on its controller. Furthermore, learning a new task requires an increased information integration in the controller.