Evaluating Classication Reliability for Combining Classiers