Learning with imperfect perception

Machine learning algorithms which adopt a state space representation usually assume perfect knowledge of what state the system is currently in. This is to guarantee that rewards and penalties are correctly assigned to the responsible state. This assumption, however, does not hold in most real world learning problems due to imperfect perception. In this paper estimation and control theories are used to classify the systems depending on the observability of the system states. This observability determines whether the optimal control strategy of a particular system can be learned. A novel approach based on enhancing the observability is used to deal with perceptual aliasing problem. In order to learn to perceive, the perception actions are directly integrated into the control actions. An example is shown and further applications to robot learning is discussed