This dissertation deals with the coupling of reinforcement learning (RL) algorithms with digital twins based on the discrete element method (DEM). These digital twins are developed to act as environments to solve RL problems of machine control and parameter optimization. Due to the remarkable performance of modern RL algorithms and the versatility of DEM simulation, the coupling of these two fields opens up possibilities for solutions to many problems of modern machines or processes. In order to achieve a suitable coupling and handle the computationally
slow DEM simulations, appropriate methodologies are developed. By applying these methodologies and state-of-the-art RL algorithms to two specific applications, the applicability of the entire approach is presented. In the first application, RL is used to solve the single- and multi-actuation task of the novel peristaltic sortation machine. Therefore, a DEM based digital twin is developed to properly represent the complex interaction of the individual parts of this machine. The second
application deals with the problem of the DEM input parameter optimization which is always required to research new materials are researched with the DEM. A newly developed approach to optimize the parameters using RL leads to remarkable results and lower computation times. The developed approaches and methodologies of this dissertation are generally adaptable to other problems and contribute to the usage of the combination of RL and DEM in many other research
fields.