Abstract: Genetic algorithm neural networks is a method that uses genetic algorithms to optimize connection right value to replace gradient decrease method. Joining the simulation annealing algorithm in the evolution process of the genetic algorithm, it has the global optimum and local search capability. At the same time, it can not only enhance the rate and efficiency of algorithmic constringency, but also effectively avoid appearing precocity and plunging into local optimum. The method can completely satisfy the accuracy and speed’s requirements of the real-time system. It was studied that genetic algorithm neural networks joining simulation annealing arithmetic operators after reproduction, crossover or mutation, and was compared their performance of optimize in the genetic neural networks.

Key words: Genetic algorithms, Simulated annealing algorithms, GSAANN

摘要： 遗传神经网络是利用遗传算法优化连接权值代替梯度下降法求解的方法，在遗传算法进化的过程中加入模拟退火算法，同时具有优秀的全局寻优能力和局部搜索能力，不仅能够提高运算收敛的速度和效率，而且可以有效避免出现早熟现象，防止陷入局部最优，同时性能也很稳定，完全能满足实时系统对精度和速度的要求。研究了遗传神经网络分别在复制、交叉和变异后应用模拟退火算子进行优化的方法，并且比较了三者在遗传神经网络优化中性能的优劣。

关键词: 遗传算法, 模拟退火算法, 遗传退火神经网络