LEARNING SCAN PATHS FOR OBJECT RECOGNITION WITH RE LATIONAL REINFORCEMENT LEARNING

Combining first order logic and reinforcement learning is known as relational reinforcement learning. In this work we examine the applicability of relational reinforcement learning in a computer vision task. We propose a state representation for reinforcement learning which is based on the perceived appearance of an object only, which especially makes the explicit encoding of world coordinates unnecessary. This enables computer vision researchers to endow their object recognition systems with additional flexibility and applicability, because they become independent of any knowledge about the camera parameters. In addition, we present results of an implementation of this approach. Our implementation is supported by a simple but effective image retrieval system. The image retrieval system is used to generate the reward during the learning episodes and it is described in this work as well.