When a multi-sensor data fusion system is used to handle a classification problem, it is necessary to incorporate the reliability coefficients of all the sensors into the fusion process. Within the framework of evidence theory, this paper proposes a new method for evaluating the reliability coefficient of a sensor based on the training data. In this method, the distance between power-set-distribution betting commitments is used to quantify the dissimilarity between the sensor reading and the reality, which can be served as a one-sided discounting factor. Then, the optimization approach is put forward to obtain an all-sided discounting factor from plenty of one-sided discounting factors. The advantages of the proposed method are analyzed comparatively. Numerical examples are also presented to demonstrate its performance by comparing it with other supervised evaluation methods.