Assessing the Quality of Learned Local Models