Learning Numerical Action Models from Noisy Input Data