为解决神经网络直接逆控制中训练样本的获取问题 ,提出了一种神经控制器的设计方法。通过对连续空间遗传算法的编码和遗传算子进行适当改进 ,采用保留精英的线性排序选择 ,避免成熟前收敛 ,并给出算术交叉算子和乘法变异算子 ,使算法同时具有好的搜索精度和搜索效率 ;然后采用这种改进的遗传算法对非线性动态系统的控制进行优化 ,获得了基于一定性能指标的期望的状态轨迹及相应的最优控制序列 ,并以此训练神经网络控制器。最后给出了以同步机为控制对象的仿真结果。 To solve the problem of obtaining training samples in direct inverse control of neural networks, a neural controller design method is proposed. By improving the coding and genetic operators of continuous space genetic algorithm, the linear ordering of reserved elites is adopted to avoid the convergence before maturity. The arithmetic crossover operator and the multiplicative mutation operator are given, and the algorithm has good searching accuracy And the search efficiency. Then, the improved genetic algorithm is used to optimize the control of nonlinear dynamic system. The desired state trajectory and the corresponding optimal control sequence based on certain performance index are obtained, and then the neural network controller is trained. Finally, the simulation results of synchronous machine are given.