Since BP neural network possesses disadvantages of slow convergence and local minimum, a global optimum algorithm simulated annealing (SA) for modifying the weight values is proposed instead of local gradient descend. The adaptive steps of state accepting and temperature - lowering processes are applied to avoiding the losing of current optimization solution and to enhancing search efficiency. By the comparison in solving XOR problem, the predominant capability of SABP algorithm is presented on global optimization and convergence precision. Finally, a SABP model of slope stability evaluation is built based on practical engineering data, and it has success fully solved the slope stability evaluation problem that has the highly nonlinear character. 