摘要
神经网络结构和权值的联合设计一直是神经网络进化设计的一个研究方向.本文根据基本微粒群算法的特点,借鉴递阶编码的思想,构造出一种多种群协同进化微粒群算法.该算法具有种群内个体微粒自由运动特征分量与种群运动特征分量分层递阶进化的特征,克服了标准微粒群算法在多峰函数寻优时出现的微粒“早熟”现象.应用该算法进行径向基神经网络隐层结构和径向基函数参数联合自适应设计,在非线性系统辨识中显示了比较好的收敛性和训练精度,同时也使网络的泛化能力和逼近精度这一对矛盾得到了比较好的协调统一.
Combination design of neural network's structure and weights has been one of the research focuses in neural network's evolutionary design. In this paper, a multi-species cooperative particle swarm optimizer is proposed by combining the ideas in the standard particle swarm optimization and hierarchy method. In the new algorithm, the individual free movement of particles within the species and the species population's movement evolve in a hierarchy model. The developed algorithm overcomes the limitation of particle's "prematurity" in global optimization using the standard PSO. When this algorithm is used in the training of RBF neural network's structure and parameters, the neural network shows a satisfactory accuracy and convergence in nonlinear system identification. The resulting network is able to properly balance the relation between generation and approximation accuracy.
出处
《控制理论与应用》
EI
CAS
CSCD
北大核心
2006年第2期251-255,共5页
Control Theory & Applications
基金
国家自然科学基金资助项目(50274060)
湖南省教育厅科研资助项目(03C499)
湖南省杰出青年基金资助项目(02JJYB009)
关键词
微粒群算法
多种群协同进化
径向基神经网络
结构优化
particle swarm optimization
multi-species cooperative
RBF neural network
structure optimization