Automated maze generation is a task for various algorithms, most commonly for the minimal spanning tree finders. Despite the efficiency of said algorithms, each algorithm is inherently limited to a specific maze classification set, namely in the Routing and Texture categories. This paper presents the foundations for a granular-control, rule-based CHR generator as an alternative maze generator. This generator is pioneering focus on describing the process of maze generation in terms of rules rather than strictly defining every step of the process. Additionally, another machine learning component will be presented to aid with proper input selection for the generator. Its goal is simulating the maze generation process and the different decisions taken along the way and feeding its output as input to the CHR Generator. The main learning technique used is Unsupervised Reinforcement Learning. A case study will also be presented to demonstrate how each component behaves.