NormalizedRegressorNc was sometimes misbehaving due to trying to calculate log(0). Fixed this by adding a small error term to each training example before taking their log error.