Empowerment and State-dependent Noise : An Intrinsic Motivation for Avoiding Unpredictable Agents

Abstract

Empowerment is a recently introduced intrinsic motivation algorithm based on the embodiment of an agent and the dynamics of the world the agent is situated in. Computed as the channel capacity from an agent’s actuators to an agent’s sensors, it offers a quantitative measure of how much an agent is in control of the world it can perceive. In this paper, we expand the approximation of empowerment as a Gaussian linear channel to compute empowerment based on the covariance matrix between actuators and sensors, incorporating state dependent noise. This allows for the first time the study of continuous systems with several agents. We found that if the behaviour of another agent cannot be predicted accurately, then interacting with that agent will decrease the empowerment of the original agent. This leads to behaviour realizing collision avoidance with other agents, purely from maximising an agent’s empowerment