Abstract: In order to overcome the insufficiencies of premature convergence and weak extensive ability in the combination of Genetic Algorithm and Artificial Neural Networks, we proposed a new hybrid study algorithmGALM, which uses the Genetic Algorithm and Levenberg Marquardt in turn to optimize the neural network. This algorithm mainly includes two stages: First a group of solutions were obtained which approximate the global optimum through cursorily adjusting the genetic algorithm. Then these approximate solutions were taken as the initial values, the GA and LM algorithms were used to optimize the neural networktraining in turn until the satisfactory network parameters were found. Finally we compared the GALM algorithm with other relevant algorithms through experimentation. The results indicate that our algorithm can effectively overcome the problem about falling into the local optimal solutions, and remarkably improves the network learning capability and the convergence rate.

Key words: LM algorithm, genetic algorithm, neural network, GALM algorithm