Poster Spotlights: Situation Dependent Spatial Abstraction in Reinforcement Learning Based on Structural Knowledge

State space abstraction reduces the size of a representation by factoring out details that are not relevant for solving a task at hand. But even in abstract representations not every detail is relevant in any situation. In cases where the structure of the environment only allows for one particular action selection, all information that does not relate to the structure can be omitted. We present a method to identify such cases in a reinforcement learning setting and abstract from non-structural details when appropriate to shrink the state space and allow for knowledge reuse. A significant performance improvement of this approach is demonstrated in a goal-directed robot navigation task.