Modellheuristiken für effizientes forward model learning

Forward model learning, i. e., learning forward models from data, finds application in prediction-based control. This involves observing inputs and outputs of the system to build a transition model and make predictions about future time steps. In particular, complex state spaces require the use of specialized search and model building techniques. In this work, we present abstraction heuristics for high-dimensional state spaces, which allow to reduce the model complexity and, in many cases, yield an interpretable result. In the context of two case studies, we demonstrate the effectiveness of the presented procedure in the context of artificial intelligence in games and motion control scenarios. The transfer of these methods enables promising applications in automation engineering.