Learning optimal actions with imperfect images