The quantification of human error probability in CREAM was improved based on available meterials. First, the basic method of CREAM was introduced. Second, two methods to determine probabilistic control modes were discussed, which are Bayesian nets and fuzzy logic. The necessity to quantify human error probability in probabilistic control modes was emphasized. Last, the method was proposed to quantify the human error probability in probabilistic control modes by theoretical inference. In order to improve the computing effectiveness, the Monte Carlo algorithm was provided. An example was presented and the validity of the proposed method was proved.