The purpose of this study is to build an on-line knowledge-based computerized adaptive diagnostic testing system (CADT), combining the computer with network application. This kernel of the CADT system is applying knowledge structure for the item select strategies in an adaptive diagnostic test. By doing that, we can effectively decrease the number of testing item according to the student’s ability. Based on the attribute of on-line testing, we add the conditions of class, time, and place, to help to avoid cheating on testing, besides randomly selecting items. In order to decrease the impact of students’ guessing out in the exam, we apply the different random items for doing similar items in a concept node of the on-line adaptive diagnostic test.