论文部分内容阅读
本文提出的层次式多子网级联神经网络是一个新的神经网络自结构方案，它通过不断地加入新的子网，逐一地分解复杂的任务为多个简单的子任务，每个子任务为一专有的子网所处理，从而达到分而治之的目的，使问题得以求解．它的优势性能在于它实现了复杂任务的自动分解和模块化训练策略，降低了全局最优搜索的复杂性，提高了训练速度，改善了网络性能．从模拟结果看，层次式多子网级联神经网络不仅在性能上优于ＢＰ网络，而且，在网络的泛化能力方面也优于级联相关学习网络．
The hierarchical multi-subnet cascaded neural network proposed in this paper is a new neural network self-structure scheme. By continuously adding new subnets, the complex tasks are decomposed into multiple simple sub-tasks one by one, and each sub-task is A proprietary subnet to deal with, so as to achieve the purpose of divide and rule to solve the problem. Its advantage lies in that it realizes the automatic decomposition of complex tasks and modular training strategies, reduces the complexity of the global optimal search, improves the training speed and improves the network performance. From the simulation results, hierarchical multi-subnet cascaded neural network is not only superior to BP network in performance, but also superior to cascade-related learning network in the generalization ability of the network.