Precise control of the braking torque of an MR-brake (MRB) is possible with a control system using a Hall sensor which measures the magnetic field. Closed-loop controllers can be used to control the torque based on the Hall sensor readings. However, over time the fluid slowly leaks out of the MRB due to failure of rubber seals, which degrades the device performance and presents challenges in torque prediction. In this research, machine learning (ML) algorithms were used to model the hysteretic relationship between the torque (T) and the Hall sensor readings (B), which is called B-T mapping. Using these models, the degrading performance of the controller can be improved by updating the mapping as the fluid leaks out. This approach can extend the useful life of the device. Results showed that, in the same experimental setting, the random forest model is more accurate than neural networks in terms of torque prediction. In addition, the 2-Step-RN approach can effectively make predictions under a realistic scenario when the fluid level changes. In particular, the random forest implementation of this approach outperformed the models that were trained for and operated in a stable environment with a fixed fluid level.