Abstract: The relevance vector machines (RVM) is very efficient in solving regression problems. A RVM-based generic model adaptive to the characteristic of the given data set is used for software failure time prediction. The RVM learning scheme is applied to the failure time data, forcing the network to learn and recognize the inherent internal temporal property of software failure sequence in order to capture the most current feature hidden inside the software failure behavior. We also compare the prediction accuracy of software reliability prediction models based on RVM, support vector machine (SVM) and artificial neural network (ANN). Experimental results by four data sets show that our RVM-based software reliability prediction model could achieve higher prediction accuracy than that of the ANN-based or SVM-based models. It is also shown that recent failure data could give more accurate prediction of near future failure event. The reasonable values of m and r can be achieved by experiments.