Specification-Aided Trajectory Prediction with RNNs
Simulations of cyber-physical systems are used to avoid cost-intensive physical experiments. Even though the overall structure of the system might be well-studied, the exact dynamics can depend on undetermined parameters. Thus, modeling the system manually leads to additional expenses. Instead, the dynamics can be learned from existing trajectories. Learning and predicting the highly non-linear dynamics of the full system, however, requires an expressive model such as recurrent neural networks and large amounts of training data. Expert knowledge of the system should be incorporated into the learning process to improve the learning rate and the quality of the model. Even though neural networks are in general considered to be black-box systems, feature preprocessing provides a convenient way to introduce additional knowledge into the learned model. We study the use of runtime monitoring specifications to calculate such features. In this thesis, we aim to model cyber-physical systems with recurrent neural networks. We study the integration of RTLola specifications as a means to introduce expert knowledge into the model and aid the learning process. We evaluate our approach on a trajectory prediction task for quadcopters.