Model-based Empowerment Computation for Dynamical Agents

摘要

Empowerment is defined as the channel capacity between action sequences and sensor information and has been studied as a kind of intrinsic rewards for survival. It is based on the mutual information conditioned by a current state, but generally calculating it needs heavy computation. This paper points out weak points of the previous methods to compute empowerment and proposes an improved method for sensorimotor environments, where their states are observed as continuous values, under the model-based setting. Then, we introduce a planning problem, whose goal is to maximize cumulative empowerment value on the articulated robots, and the behaviors after training are discussed while referring to other theories of intrinsic rewards and the control theory.