Abstract
This work considers the generalization of the policies for reinforcement learning. The problem of generalizing from a set of examples or researching features are widely explored subjects in artificial intelligence. Thus, we propose to use generalization methods based on a language bias to generalize policies. Those are principally based on Galois lattices properties.Firstly, we propose a general algebraical framework formalizing the generalization of policies. During its learning, the agent observes a set of the environment’s states. We formalize the generalization of policies as a partitioning of this set.Usually used on the powerset, we propose to use the Galois lattices methods on the partition lattice. This allows us to propose an algorithm which produces, after a reinforcement learning, interesting concepts from the task point of view. These can be used to describe the policy or as features for other similar tasks. Finally, we propose a new way to formalized a reinforcement learning task and an associated algorithmic method called Q-Concept Learning, consisting in applying a learning step on the all set of the generalizations available considering the used language bias. In this context, we will discuss about solutions that allow the agent to generate in line policies. Our methods are implemented ans tested on academical problems.