摘要
回声状态网是一种新型的递归神经网络,仅需要训练输出权值,克服了传统递归神经网络存在训练算法过于复杂、收敛速度慢、易于陷入局部最小等问题.为进一步提高回声状态网的建模能力,特别是面临实现多个正弦函数叠加(MSO)等任务时的逼近能力,提出了由泄露积分型神经元构建一种新型的多储备池回声状态网,称为多储备池泄露积分回声状态网(MLESN).MLESN是采用由顶向下和由底向上相结合的思路构建回声状态网.首先采用由顶向下的思路构建回声状态网,假设其储备池由P类相异泄露积分型神经元构成,每一类神经元的群体将构成一个子储备池.然后采用由底向上方式构建回声状态网,生成P个相异中心神经元,分别代表P个子储备池,每个子储备池的神经元状态要与其中心神经元状态相同或相近,P个中心神经元之间通过随机稀疏连接构成了一个新的虚拟子储备池.在储备池状态更新过程中,仍需保持各子储备池内部神经元之间的差异性小,不同子储备池神经元状态之间的差异性大的特点.最后,利用Matlab仿真软件进行实现,并与泄露积分型回声状态网(Leaky-ESN)预测性能进行比较.仿真结果表示,本文提出的方法具有更高的预测精度和预测误差波动性小等特点.
The echo state network is a new type of recurrent neural network that only needs to train the output weights,which overcomes the problems of traditional recurrent neural network,such as too complex training algorithm,slow convergence speed,easy to fall into local minimum and so on.To further improve the modeling ability of the echo state network,especially the approximation ability when faced with tasks such as multiple sine function superposition(MSO),this paper proposes to construct a new type of multi-reservoirs echo state network from leaky integral neurons,which is called multi-reservoirs leaky integral echo state network(MLESN).MLESN is a combination of top-down and bottom-up ideas to build an echo-state network.First,a top-down idea is used to build an echo-state network,assuming that its reservoir consists of P-type differential leakage integral neurons,each type of neuron group will form a sub-reservoir.Then using the bottomup method to build an echo state network to generate P different central neurons,which respectively represent P sub-reservoirs.The state of neuron in each sub-reservoir is the same or similar to that of its central neuron.A new virtual sub reservoir is constructed by randomly sparse connection between P central neurons.During the state update of the reservoir,it is still necessary to keep the difference between the neurons in each subreservoir small,and the big difference between the states of different sub-reservoir neurons.Finally It is implemented using Matlab simulation software and compared with the Leaky-ESN prediction performance.Simulation results show that the proposed method has higher prediction accuracy and less volatility of prediction error.
作者
王茜
伦淑娴
WANG Qian;LUN Shuxian(College of Mathematical,Bohai University,Jinzhou 121013,China;College of Engineering,Bohai University,Jinzhou 121013,China)
出处
《渤海大学学报（自然科学版）》
CAS
2020年第1期90-96,共7页
Journal of Bohai University:Natural Science Edition
基金
国家自然科学基金项目(No:61773074).
关键词
储备池
回声状态网
泄露积分回声状态网
稀疏连接
reservoir
echo state network
leaky-integrator echo state network
sparse connection
