论文部分内容阅读
提出一个模块化神经网络的广义定义,它包含了几乎所有多神经网络(系统)。简要分析了模块化神经网络子网集成的相关概念和问题。针对一类模块化神经网络,提出了5种基于“分而治之”原理和自适应组合的新型动态集成方法。它们之间的主要区别在于:距离测度(绝对距离测度和相对距离测度);个体数目(有些全部参与集成,有些则是部分参与);集成策略和规则(数据驱动和数据/知识驱动)。仿真实验证实了这些方法的有效性。同时,还提出了一种基于“一专多能”思想的子网训练方法。
Proposed a broad definition of modular neural network, which contains almost all multi-neural network (system). The related concepts and problems of modularization neural network subnet integration are briefly analyzed. For a class of modularized neural networks, five new dynamic integration methods based on “divide and rule ” principle and adaptive combination are proposed. The main differences between them are: distance measures (absolute distance measure and relative distance measure); number of individuals (some fully involved, some partially involved); integrated strategies and rules (data driven and data / knowledge driven). Simulation results confirm the effectiveness of these methods. At the same time, a sub-network training method based on the idea of “multi-specialty” is also proposed.