用于神经网络直接逆控制的连续空间遗传算法

Design of Neural Network Direct Inverse Controllers Based on Genetic Algorithm in Continous Space

为解决神经网络直接逆控制中训练样本的获取问题 ,提出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 ,验证了方法的有效性 .

A generalized design method of neural controllers was proposed for the acquisition of training data in the neural network direct inverse control. The genetic algorithm in continous space was improved in coding and genetic operators by using linear sequencing selection, arithmetic crossover and multiplier mutation, which overcomes its premature convergence and make it posess high search precision and efficiency. The control imputs of nonlinear dynamic systems were optimized by the modified genetic algorithm. A neuro controller was trained with the obtained desirable response trajectory. A synchronous machine was used as a test bed to demonstrate the effectiveness of the proposed design method, and the simulation results were also given.