The diverse and increasingly stringent demands on the mineral processing industry ¿ such as energy efficiency, CO2 neutrality, and product quality ¿ require innovative control solutions to ensure that the processing methods and techniques used can continue to be operated economically in the future. This thesis introduces a machine learning-based approach, specifically utilizing online reinforcement learning, to offer a dynamic and customizable control system. Focusing on key operational variables, a data-based digital twin of the grinding circuit was developed as an accurate simulation tool for operational changes. This allows for the testing of new setpoints without negatively impacting the real-world operation of the plant. A specialized training environment for reinforcement learning algorithms was also established, reflecting the actual characteristics of the grinding circuit and enabling iterative adaptation. The research conducted as part of this dissertation combines simulation and algorithmic modeling, demonstrating the potential of machine learning for improved and industrially viable control approaches. The developed control framework was rigorously tested in a semi-industrial environment, utilizing a digital twin to validate its effectiveness and readiness for real-world industrial applications. A corresponding industrial layout, designed for edge-computing deployment, adds robustness and user-friendliness to the system. Further contributions were made in exploring the Human-in-the-Loop (HITL) concept, addressing the themes of safety, legal considerations, and system scalability. The study lays a solid foundation for future academic research and offers a ready-to-implement solution for industrial applications. It brings a new dimension to grinding circuit operations by leveraging machine learning's efficiency and flexibility, thereby contributing substantively to the existing body of knowledge in the area of mineral processing.