How to make bias and variance errors insensitive to system and model complexity in identification