多函数激活的拉普拉斯深度回声状态网络

结合可变激活函数、降维算法和深度回声状态网络针对新的神经网络模型进行了研究。其中可变激活函数是多函数的线性组合，可以通过调整系数来改变激活函数的非饱和区；拉普拉斯特征映射降维算法通过降低储层状态矩阵的维度来改善原网络面临的病态、不适定问题；还使用了遗传算法来寻找最佳目标子空间维度。仿真分析从扰动影响、转换稳定性、时序预测和记忆容量四个方面进行，从仿真结果（新模型的记忆容量是深度回声状态网络的两倍，均方根误差比回声状态网络小42%）来看，新模型的记忆容量、预测精度都得到了显著改善。

Laplace deep echo state network with multiple activation functions

This paper studied on a novel neural network model, which combined variable activation function, dimensionality reduction algorithm and deep echo state network. The variable activation function was a linear combination of multiple functions, and adjusting the coefficients could change the unsaturated region. Laplace feature mapping dimensionality reduction algorithm could improve the ill-posed problems by reducing the dimensions of the reservoir state matrix. It used genetic algorithm to find the optimal target subspace. It carried out simulation from four aspects: disturbance influence, transformation stability, time series prediction and memory capacity. The simulation results(the memory capacity of the novel model is twice than that of DESN, and the root-mean-square error is 42% smaller than ESN) show that the memory capacity and prediction accuracy of the new model are improved compared with other models.