A method for designing and training artificial neural network based on genetic algorithm(NNDT) was presented.NNDT trains both architectures and weights of networks simultaneously.The problem that the one-to-one relationship between the topology and the encoding can not be kept was solved effectively.Heuristic method was used to constraint the probability of topology mutation and the trend of the choice of the kind of mutation.Also,the niching mechanism was used to protect the mutation of network topologies.The experiments results show the efficiency and rapidity of NNDT.