Attributes reduction based on rough set theory is an important but difficult task under incomplete information system. For the attributes reduction which is gained by the old attributes reduction algorisms , the attribute belongs to it or not . Nevertheless in the practice when there is the probability that the attribute can discern two objects , this shows the attribute may belong to the attribute reduction. Probability discernibility matrix is defined and corresponding discernibility function is given. Then a probability algorism for attributes reduction is proposed and an example shows the algorism is effective. The probability that the attribute belong to the reduction can be know from the reduction which is gained by the algorism.