Abstract: In this paper, the conception of train-running limit state under earthquake is proposed, and the corresponding optimum reliability is studied. In order to optimize the total cost during service life, which is the function of reliability, an approximate method is developed, and the Genetic Algorithm is used for optimization. The damage-repair process is described by Renewal Process. The damage index of a member is expressed in terms of the ratio of maximum curvature to the ultimate one at the position of plastic hinge. Coefficient of derail is used to judge whether a train is in the normal running state or in derail state. The failure probability is calculated by the improved Response Surface Method. A numerical example on optimum reliability is given.