Interactive selection of visual features through reinforcement learning

We introduce a new class of Reinforcement Learning algorithms designed to operate in perceptual spaces containing images. They work by classifying the percepts using a computer vision algorithm specialized in image recognition, hence reducing the visual percepts to a symbolic class. This approach has the advantage of overcoming to some extent the curse of dimensionality by focusing the attention of the agent on distinctive and robust visual features. The visual classes are learned automatically in a process that only relies on the reinforcement earned by the agent during its interaction with the environment. In this sense, the visual classes are learned interactively in a task-driven fashion, without an external supervisor. We also show how our algorithms can be extended to perceptual spaces, large or even continuous, upon which it is possible to define features.