A robotic gym with multiple rehabilitation robots allows multiple patients to exercise simultaneously under the supervision of a single therapist. The multi-patient training outcome can potentially be improved by dynamically assigning patients to robots based on monitored patient data. In this paper, we present an approach to learn dynamic patient-robot assignment from a domain expert via supervised learning. The dynamic assignment algorithm uses a neural network model to predict assignment priorities between patients. This neural network was trained using a synthetic dataset created in a simulated rehabilitation gym to imitate a domain expert's assignment behavior. The approach is evaluated in three simulated scenarios with different complexities and different expert behaviors meant to achieve different training objectives. Evaluation results show that our assignment algorithm imitates the expert's behavior with mean accuracies ranging from 75.4% to 84.5% across scenarios and significantly outperforms three baseline assignment methods with respect to mean skill gain. Our approach solves simplified patient training scheduling problems without complete knowledge about the patient skill acquisition dynamics and leverages human knowledge to learn automated assignment policies.