国家重点研发计划(2022YFE0112400); 国家自然科学基金(61972181, U20A20228); 江苏省教育厅高校自然科学重点项目(22KJA520009)
基于宽度学习的动态模糊推理系统(broad-learning-based dynamic fuzzy inference system , BL-DFIS)能自动构建出精简的模糊规则并获得良好的分类性能. 然而, 当遇到大型复杂的数据集时, BL-DFIS因会使用较多模糊规则来试图达到令人满意的识别精度, 从而对其可解释性造成了不利影响. 对此, 提出一种兼顾分类性能和可解释性的模糊神经网络, 将其称为特征扩展的随机向量函数链神经网络(FA-RVFLNN). 在该网络中, 一个以原始数据为输入的RVFLNN被作为主体结构, BL-DFIS则用作性能补充, 这意味着FA-RVFLNN包含具有性能增强作用的直接链接. 由于主体结构的增强节点使用Sigmoid激活函数, 因此, 其推理过程可借助一种模糊逻辑算子(I-OR)来解释. 而且, 具有明确含义的原始输入数据也有助于解释主体结构的推理规则. 在直接链接的支撑下, FA-RVFLNN可利用增强节点、特征节点和模糊节点学到更丰富的有用信息. 实验表明: FA-RVFLNN既减缓了主体结构RVFLNN中过多增强节点带来的“规则爆炸”问题, 也提高了性能补充结构BL-DFIS的可解释性(平均模糊规则数降低了50%左右), 在泛化性能和网络规模上仍具有竞争力.
The broad-learning-based dynamic fuzzy inference system (BL-DFIS) can automatically assemble simplified fuzzy rules and achieve high accuracy in classification tasks. However, when BL-DFIS works on large and complex datasets, it may generate too many fuzzy rules to achieve satisfactory identification accuracy, which adversely affects its interpretability. In order to circumvent such a bottleneck, a fuzzy neural network called feature-augmented random vector functional-link neural network (FA-RVFLNN) is proposed in this study to achieve excellent trade-off between classification performance and interpretability. In the proposed network, the RVFLNN with original data as input is taken as its primary structure, and BL-DFIS is taken as a performance supplement, which implies that FA-RVFLNN contains direct links to boost the performance of the whole system. The inference mechanism of the primary structure can be explained by a fuzzy logic operator (I-OR), owing to the use of Sigmoid activation functions in the enhancement nodes of this structure. Moreover, the original input data with clear meaning also help to explain the inference rules of the primary structure. With the support of direct links, FA-RVFLNN can learn more useful information through enhancement nodes, feature nodes, and fuzzy nodes. The experimental results indicate that FA-RVFLNN indeed eases the problem of rule explosion caused by excessive enhancement nodes in the primary structure and improves the interpretability of BL-DFIS therein (The average number of fuzzy rules is reduced by about 50%), and is still competitive in terms of generalization performance and network size.