Abstract: The problem of improving the fault-tolerance of feedforward neural networks that suffers open fault is transferred to a minimum optimization problem, and a genetic algorithm is used to evolve networks that have good fault-tolerance and strong generalization ability. Since neither extra network redundancy is introduced nor training algorithm is modified, the independence of the network architecture, training algorithm, and fault-tolerant process are well kept. Experiments show that this method achieves good results in two benchmark tests.