The technology of automated short answer grading (ASAG) can efficiently process answers according to human-prepared grading examples. Computer-assisted acquisition of grading examples uses a computer algorithm to sample real student responses for potentially good examples. The process is critical for optimizing the grading accuracy of machine learning models given a budget of human effort and the appeal of ASAG to online learning providers. This paper presents a novel method called short answer grading with active learning (SAGAL) that features a unified formulation comprising the heuristics for identifying potentially optimal examples of representative answers, borderline answers, and anomalous answers. The method is based on active learning, which iteratively samples good examples and queries for annotation to increase the sampling accuracy. SAGAL has been evaluated with three different public datasets of distinctive characteristics. The results show that the resulting models generally outperform the baseline semi-supervised learning methods on the same number of grading examples.