Reinforcement learning is an active field in machine learning where an agent learns to perform a task by interacting with the environment and receiving positive or negative rewards depending on the chosen actions. Recently, reinforcement learning has seen some big breakthroughs in beating the best human players various tasks, such as the classic board game Go and the popular video game StarCraft II. One of the reasons why the architectures that were used are so successful is that deep learning modules are used which can perform some form of relational reasoning. This allows them to view the environment in terms of distinct objects and make use of the relations between these objects. The field of relational reinforcement learning looks at how these relations between objects can be learned and used to optimally solve the given tasks. As this field is relatively new, there are still many open research questions, such as how to best create a representation of the environment based on these objects and relations and how to improve the efficiency of these networks by learning only the important relations while ignoring the irrelevant ones. In this proposal we introduce new relational reinforcement learning architectures that will allow us to efficiently represent the environment in a relational way, improve efficiency by focusing on the important relations in this representation and increase the ability to generalize to unseen tasks.