How to Train AI on a New Candy Bar

When something changes on the machine, which hasn't been foreseen, the code needs to be adjusted. Online learning implies that you must do some trade-off between exploration and exploitation. This means if you occasionally need to try out random actions or control concepts - this is not a good idea for conveyor systems running with speeds of 1000 millimeters per second. Our AI can only deal with those or similar situations as it has seen from the simulation during the training process. And after training, the weight matrices of our model simply become freest for safety reasons.