According to the topological characteristics of BP networks, a Genetic Algorithm based on the structural formula binary-coding has been designed in this paper. By means of fractionalizing the large-scale solution-space and performing the GA operations to the fractionalized subspaces, the GA's global-convergence and parallelism can be utilized to search the subspace for the optimal solution in the whole solution-space, thus definitude the starting point and narrow the domain for the next BP's local-search. Testing shows that the two-step algorithm (GA-BP) can solve the existed problems in the NN's training such as local minimum, tardy convergence and so on.