Escaping brittleness: the possibilities of general-purpose learning algorithms applied to parallel rule-based systems