摘要
针对回声状态网络(ESN)结构设计复杂、参数选择难度大的问题,提出一种具有small world特性的ESN(SWESN).首先采用神经元空间增长算法在平面区域生成small world拓扑网络;然后根据网络节点与基准点的Euclidean距离将网络节点进行重新排序,并将平面上的物理节点及其连接映射为SWESN的内部神经元连接矩阵,从而使动态神经元池具有small world特性.实验表明,SWESN动力学特性比常规ESN更为丰富,在鲁棒性、抗干扰能力等方面均优于常规的ESN.
For the problems of complex structure design and hard parameters selection on echo state network（ESN） frame, this paper proposes a structure design method of dynamic neurons reservoir（DNR） with the small world structure（SWESN）. Design method of SWESN is introduced in detail. Neurons space growth algorithm is used to generate network with small world topology structure on 2-D plane. Then the neurons are rearranged by Euclidean distance from network＇s nodes to the fiducial node, and the physical nodes of the plane region and their internal connections are mapped to the connection matrixes of ESN internal neurons. This design method makes the DNR have small world property. The simulation experiments show that, the proposed ESN can create more abundant dynamic behavior than conventional ESN, and both robustness and anti- interference ability of SWESN are better than that of conventional ESN.
出处
《控制与决策》
EI
CSCD
北大核心
2012年第3期383-388,共6页
Control and Decision
基金
国家自然科学基金重点项目(61034008)
国家自然科学基金项目(60873043)
北京市自然科学基金项目(4092010)
教育部博士点基金项目(200800050004)
北京市属高等学校人才强教计划项目(PHR201006103)
关键词
回声状态网络
小世界
动态神经元池
动力学特性
echo state network
small world
dynamic neurons reservoir
dynamic behavior
