In order to improve the performance of the missile reliability estimation in free flight phase, the research of the regression problem with small scale of samples is carried out. Firstly, the learning samples are obtained after feature selection and abstraction, based on which the SVM is used to estimate the reliability of the missiles. Then the estimation performance of the neural network and SVM is compared by simulation. The results indicate that the efficiency of SVM was higher than the neural network, and SVM also has good generalization ability and can improve the performance of the reliability estimation.