基于PSO算法的动态模块化神经网络结构设计

Dynamic modular neural network structure design based on PSO algorithm

针对模块化神经网络结构设计过程中子网络输出不能最优集成的问题,提出一种基于粒子群算法的动态模块化神经网络.首先,该网络采用数据密度辨识样本分布空间,并更新数据中心;然后,根据输入数据激活相应的子网络,利用PSO算法寻找子网络的最优网络贡献度,并依据贡献度计算子网络的输出权值;最后优化模块化神经网络的集成输出.通过对非线性函数和时变系统的逼近实验,验证了集成网络中子网络数目可以根据任务动态调整,网络输出的集成权值能够通过PSO算法寻找到最优值,并且训练精度和自适应能力较其他算法均有一定的提高.

In order to solve the problem the sub-network output can not be optimally integrated in a modular neural network（MNN）, this paper proposeds a dynamic MNN based on the particle swarm optimization（PSO） algorithm. Firstly,the distribution of samples can be identified and the center of datas can be updated by computing the data density. Secondly,the corresponding sub-networks are activated according to the input datas, then the output weights are calculated by the best contribution degrees which are computed via the PSO algorithm. Finally, a dynamic neural network is completed to optimize the integrated output of the MNN. Based on the approximating experiments of the non-linear function and timeseries prediction, it is proved that the number of sub-networks can be adjusted dynamically, and the integrated weights of the neural network can be optimized by using the PSO algorithm. Comparisons with other algorithms demonstrate that the proposed method is more effective in terms of the accuracy and adaptive ability.