机构地区: 华北理工大学信息工程学院,河北唐山063210 南京邮电大学计算机学院,江苏南京210023
出 处: 《南京邮电大学学报(自然科学版)》 2017年第4期97-102,共6页
摘 要: 针对储备池的适应性问题,提出了小世界递归小波神经网络。首先基于复杂网络理论构建了具有小世界效应的稀疏储备池结构,代替原来的随机拓扑结构,为避免孤立节点的产生,该结构通过在最近邻耦合网络中随机加边来实现。其次,引入了具有良好时频局部特性的小波神经元,包括Morlet小波、Mexican hat小波、Gaussian小波和B-spline小波,并与传统的Sigmoid神经元结合,建立了储备池神经元的混合激励模式。最后,实验仿真结果表明:对比传统的小世界回声状态网络,该模型能够有效地提高对非线性系统的逼近能力。 Small-world recurrent wavelet neural networks are proposed for solving the problems about reservoir adaptation. A sparse reservoir with small-world feature is built instead of the original random one,and generated from the nearest-neighbor coupled network by randomly adding edges to avoid isolated nodes. Then,wavelet neurons are introduced,considering Morlet wavelet,Mexican hat wavelet,Gaussian wavelet,and B-spline wavelet. Combined with Sigmoid neurons,a hybrid activation scheme is used for the reservoir. Experimental results show that the model can achieve a more significant enhancement in nonlinear approximation capacity compared with traditional small-world echo state networks.