Relate to Predict: Towards Task-Independent Knowledge Representations for Reinforcement Learning