Abstract：During the learning process of back-propagation (BP)networks based on gradient descent principle,a good weighting value distribution can be finally obtained from adopting certain definite rule of weights variation and then gradually adjusting it in training in ordinary method,but it usually falls into local optimal solution because of nonlinear mufti-extreme object function. Training networks combining genetic algorithm(GA)based on global optimization with back- propagation (BP) based on gradient descent in the paper make the linking weights of networks self-adaptive evolution in constantly iterative process. The practice reconstructing sonic logging under bottom hole in NH area by extrapolating seismic characteristic parameters of near-well bore traces shows that the evolutionary learning method can overcome shortcomings of traditional method,and avoid "pseudo-learning"phenomenon in the training and improve popularizing and predicting ability for networks.
