Niceness assumptions for Learning Algorithms