designing a reinforcement learning controller from scratch.

The trained algorithm predicts the ground truth value with a normalized RMSE of 0.05 on a synthetic dataset. This accuracy is sufficient for diagnostic support to the physician. With sufficient robustness, this is suitable for automated triggering of a medical ventilator, as well. On robustness, the algorithm outperforms the incumbent traditional algorithm, maintaining high accuracy despite artifacts in the measured signal. On top of that it has lower latency. The algorithm is tested on data from measurements with volunteers, as well, with excellent performance.

explore more cases.

synthetic data for food sorting.

Sorting of food is inherently complex because of the great variation in food and specifically in organic produce. Machine learning is a powerful tool and very fitting to this task.

patient ventilator asynchrony in respiratory support.

Critically ill patients who are unable to breathe (sufficiently) on their own receive respiratory support. It is essential that this support is synchronized (in time) with the patient’s spontaneous breaths.

suppress medical pump generated flow disturbances with RL.

Pumps in medical applications require a steady and controllable flow. We designed a medical pump consisting of valves and pistons, but due to the limited length of the pistons and the fluid displacement by the valves, the dynamics of this pump are discontinuous and non-linear.