Cognitive Diagnostic Computerized Adaptive Testing is a new testing mode which combines Computerized Adaptive Testing with Cognitive Diagnosis. It has the characteristics of adaptive and cognitive diagnosis, which can use less items and less time to evaluate knowledge state, it is applicable to classroom instruction because of timely and immediately evaluation. It is the basis of remedial teaching and students’ self-learning. It was presented by Thompson that CD-CAT had five important parts, they were parameter-calibrated and attribute-identified item bank, initial stage item selection methods, an item selection strategy, a knowledge stage estimation method and a stopping rules. Of which initial stage item selection methods can influence the pattern classification correct rate. It was only found through literature review that Tu, Cai, & Dai (2013) studied the initial stage item selection method, and presented “T matrix-method”, which selected initial items from R-matrix. The other initial stage item selection method was random method, which means that select first stage items from item bank randomly, this method does not consider item parameters and attributes. It is easy and fast. The research showed that T-matrix method had higher PCCR than random method. It is presented four initial stage item selection methods based on the discrimination of Cognitive Diagnosis, they are CTTID method, CDI method, CTTIDR* method and CDIR* method. 1) CTTID calculates discrimination based on identification index of Classical Test Theory, the fundamental question is “How well does this item help me to differentiate between respondents who have mastered “more” attributes and respondents who have mastered “fewer” attributes?” In DINA model, high discriminatory items are the items slipping and guessing parameters smaller. 2) CDI method is presented by Henson and Douglas (2005), which is a weighted average of the elements in D where the KLI values associated with attribute profiles that differ by one attribute have the highest weight and the KLI values associated with attribute profiles that differ by all A attributes have the smallest weight. 3) Reachability-matrix (R-matrix) was presented by Tasuoka (1995), it describes the direct, indirect and own relationships between attributes, and it is a matrix composed with K line * K column (K is the number of attributes). R* matrix is the necessary condition for testing to achieve diagnosing each attribute. Item parameter and R matrix are the important elements which can influence the efficiency of testing, so here we consider discrimination and R matrix simultaneously to select first stage items, CTTIDR* method is from this consideration. 4) A “good” item is one that discriminates well between respondents with different levels of ability, disposition, or another set of latent characteristics that are measured by a diagnostic assessment whose response data are modeled with a CDM. CDI is a good discrimination of CDM which can use in different models, CDIR* method is combined CDI with R* matrix. A simulation research was used to verify these methods of CD-CAT. CDM was DINA model, the number of examinee was 1000, item bank was 300, the number of attributes was 6. PCCR was the evaluation indices. The result showed that PCCR of CTTIDR* method was higher than existing T-matrix method with 0.2697 under fixed-length CD-CAT at the end of initial stage. When the testing finished, CTTIDR* method was also the best. The average testing length was also smallest under the variable-length CD-CAT.