遗传算法在神经网络控制中的应用与实现

比较了遗传算法与神经网络的特点，并对将遗传算法用于前向神经网络的可能性进行了研究，同时阐明了遗传算法和神经网络结合的必要性，提出了一种融合遗传算法的神经网络控制方法，该方法采用多层前向神经网络作为遗传搜索表示方式的思维，以神经网络为基础，用遗传算法 习神经网络的权系数，即保留了遗传算法的强全局随机搜索能力，又具有神经网络的鲁棒性和自学习能力，将遗传算法和神经网络相结合，分析了遗传算法基本参数及神经网络结构、隐层和输出层节点非线性函的选择，设计了用遗传算法学习神经网络权系数的软件实现方法，成功地实现了机械手逆运动学求解问题及倒立摆的控制，仿真结果显示了遗传算法快速学习神经络权系数的学习效率与收敛精度，确保了快速达到全局收敛，克服了多层前向神经网络传统的BP学习算法精度低、收敛速度慢、容易陷入局部极小的缺陷，表明了该方法的可行性与有效性。

Applying and Realizing of Genetic Algorithm in Neural Networks Control

The characteristics of neural networks and genetic algorithm are described. The possibility and the method of the application of genetic algorithm to the multi-layer forward neural networks are discussed. The necessity of combining neural networks with genetic algorithm is demonstrated. A kind of neural networks control method is proposed in which genetic algorithm and neural networks are mixed. In this method, the notion of using the multi-layer forward neural networks as the representation method of the genetic searching technique is introduced, and the weighs of neural networks are trained by genetic algorithm. So the method remains the global stochastically searching ability of genetic algorithm and the robustness and self-learning ability of neural networks. After the neural networks with genetic algorithm are combined organically, the selection of the basic parameters in genetic algorithm and the structure of neural networks and the nodes of the hidden layer and the output layer are all analyzed. The software in which the weights of neural networks are learned by genetic algorithm is designed. The inverse kinematics solution of the robot manipulators and the inverted pendulum control are successfully realized by the combination of genetic algorithm and neural networks. The simulation results indicate the capability of the new method in fast learning of neural networks and guarantee a rapid global convergence. Moreover, the premature convergence in genetic algorithm is restrained effectively, and the learning efficiency and the convergent precision for the weights of the multi-layer forward neural networks are improved greatly. The motivation of this approach is to overcome the shortcomings of traditional error back propagation algorithm for updating the weights of the multi-layer forward neural networks, such as the low precision of the solutions, the slow search speed and easy convergence to the local minimum points. These results show the proposed method in this paper is feasible and effective.