改善遗传神经网络收敛性的研究  

Study on improving the convergence of genetic neural networks

作　　者：李享梅[1] 赵天昀[2] 

机构地区：[1]成都信息工程学院网络工程系,四川成都610225 [2]郑州大学信息管理系,河南郑州450001

出　　处：《计算机应用》2005年第12期2789-2791,共3页journal of Computer Applications

摘　　要：针对BP神经网络中采用的梯度下降法局部搜索能力强、全局搜索能力差和遗传神经网络中采用的遗传算法全局搜索能力强、局部搜索能力差的特点,提出了一种集梯度下降法和遗传算法优点为一体的混合智能学习法(HybridIntelligencelearningalgorithm),简称HI算法,并将其应用到优化多层前馈型神经网络连接权问题。对该算法进行了设计和实现,从理论和实际两方面证明混合智能学习法神经网络与BP神经网络和基于遗传算法的神经网络相比有更好的运算性能、更快的收敛速度和更高的精度。To describe the advantage and shortcoming of gradient descent algorithm and genetic algorithm for training connection weights of neural networks, a new algorithm combined genetic algorithm with gradient descent algorithm was proposed, referred as to Hybrid Intelligence learning algorithm（HI）. Applied to the problem of optimizing the connection weight of the feedforward neural networks, the algorithm was feasible. The design and realization of HI was introduced. And it was proved that hybrid intelligence learning algorithm is better, faster and more accurate than gradient descent algorithm and genetic algorithm in theory and practice.

关 键 词：遗传算法 遗传神经网络 人工神经网络 BP神经网络 梯度下降法 混合智能学习法 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]