Applying reinforcement learning algorithms in real-world domains is challenging because relevant state information is often embedded in a stream of high-dimensional sensor data. This paper describes a novel algorithm for learning task-relevant features through interactions with the environment. The key idea is that a feature is likely to be useful to the degree that its dynamics can be controlled by the actions of the agent. We describe an algorithm that can find such features and we demonstrate its effectiveness in an artificial domain.
展开▼