XCS is seen as the most investigated Michigan-style Learning Classifier System (LCS). It is typically integrated in Organic Computing systems as their self-adaptive learning component. Despite its attractiveness for Organic Computing due to its interpretability and maturity, XCS has well-known weaknesses in multi-step Reinforcement Learning problems. Reasons for this are the curse of dimensionality of Michigan- style LCSs, a weakness against long action chains and learning stability issues. In the meanwhile, modern Deep Reinforcement Learning methods have left classic Reinforcement Learning methods including XCS far behind. The PhD project outlined in this paper aims to create Deep Learning-XCS-hybrids leveraging the computational and representational abilities of Deep Learning to improve the Reinforcement Learning capabilities of XCS. Hereby, it aims to narrow or close the gap between Deep Reinforcement Learning and LCS. Furthermore, this paper discusses the key concepts behind this vision, resulting research questions, the results so far and the road ahead.