Reinforcement Learning of Perceptual Classes using Q Learning Updates

S. Jodogne and J.H. Piater (Belgium)


Machine Learning, Computer Vision, Feature Selection.


We introduce a new Reinforcement Learning algorithm de signed to operate in perceptual spaces upon which it is pos sible to define features, such as visual spaces. It incremen tally constructs a classifier that maps the percepts to a per ceptual class by testing the presence or absence of highly informative features. This approach has the advantage of enhancing the generalization abilities of the autonomous agent, while reducing the influence of noise, as well as the size of the perceptual domain. The perceptual classes are iteratively refined using a statistical analysis of the updates that Q Learning would ap ply to an optimal Q function for the considered classifica tion. This process relies only on the reinforcements earned by the agent during its interaction with the environment. Thus, the distinctive features are selected interactively in a task-driven fashion, without an external supervisor.

Important Links:

Go Back