WS型多层前向小世界神经网络结构自适应模型

Adaptive structure model for WS multilayer feedforward small-world artificial neural networks

WS型多层前向小世界神经网络模型中捷径生成具有随机性,这种随机的删除且向后跨层连接产生捷径很可能使小世界体系神经网络丢失重要的信息,导致网络性能变差。针对这个问题提出了一种WS型多层前向小世界神经网络结构自适应模型,该模型借鉴了复杂动态网络研究中和谐统一的混合择优思想,通过以确定的连接权值矩阵为指导有目的性地选择权值大的边产生捷径,生成小世界体系结构。将新网络应用于函数逼近,在设定精度相同情况下对不同跨层概率下的收敛次数作比较,仿真发现该小世界网络在概率0.05附近时比同规模的网络有更好的收敛速度,与WS小世界人工神经网络在相同概率的情况下达到相同精度所需的平均迭代次数相比较,收敛速度得到提高。

The model of Watts-Strogatz （WS） multilayer feedforward small-world artificial neural networks generate short- cuts randomly, the random delete and cross-layer link backwards likely make small-world neural network loss important information, and lead to network performance becoming poor. According to this problem the model of WS multilayer feedforward small-world adaptive structure artifical neural network was put forward. The model bases on Harmonious Unifying Hybrid Preferential Model （HUHPM） in complex dynamic network through selecting the edges with large weights to generate short-cuts guided by deterministic connection weight matrix, constructs small-world artificial neural networks. The new network was used to function approximation. The simulation results show that the new network converges faster, in other words, has the smaller iterations compared with the WS small-world artificial neural network, and converges the fastest near the above probability of 0.05.