Crowd dynamics is the study of behaviour in crowd movement, and can have critical applications such as urban planning and evacuation simulations, as well as creative utilization in game and VR/AR design. Using a graph network, crowd dynamics can be simulated by representing the agents and their interrelationships, which allows for the usage of data-driven learning methods to construct a improved locomotion model. Random feature methods are integrated using the SWIM algorithm, offering a computationally efficient solution to traditionally cumbersome training times from neural network- based approaches. Such models are then trained and validated on simulated data from knowledge-based models, generated using the software Vadere. Prediction results are visualized, both statically and dynamically, as well as evaluated to observe output characteristics. Experimentation with real-world data as training input is also conducted, where the outcome is compared against models trained exclusively with simulated data. Finally, further refinement potentials and application fields are explored to expand on the future possibilities of this study.