ML 型迁移学习模糊系统

经典模糊系统构建方法训练时通常仅考虑单一的场景,其伴随的一个重要缺陷是: 如当前场景重要信息缺失,则受训所得系统泛化能力较差.针对此问题, 以Mamdani-Larsen (ML)型模糊系统为对象,探讨了具有迁移学习能力的模糊系统, 即ML型迁移学习模糊系统. ML型迁移学习模糊系统不仅能充分利用当前场景的数据信息, 而且能有效地利用历史知识来进行学习,具有通过迁移历史场景知识来弥补当前场景信息 缺失的能力.具体地,基于经典的压缩集密度估计(Reduced set density estimator, RSDE) ML型模糊系统构建方法, 通过引入迁移学习机制提出了一种基于密度估计的ML型迁移模糊系统构建方法. 在模拟数据和真实数据上的实验研究亦验证了该迁移模糊系统在信息缺失场景下较之于 传统模糊系统建模方法的更好适应性.

Mamdani-Larsen Type Transfer Learning Fuzzy System

The classical fuzzy system modeling methods only consider the single scene, which may bring up the following weakness: if the information of partial data is missing, the fuzzy systems constructed based on this dataset will have the weak generalization abilities for this scene. In order to overcome this shortcoming, by focusing on the Mamdani-Larsen type fuzzy system (ML-FS) model, the fuzzy system with the transfer learning abilities, i.e. ML-transfer fuzzy system is proposed. The ML-transfer fuzzy system will not only make full use of the data information in the learning procedure, but also effectively learn from the existing useful historical knowledge, such as parameters of the fuzzy system obtained from the dataset of historical scene, in order to make up the information lack in the current scene. Based on this idea, a specified ML-transfer fuzzy system (ML-TFS) based on reduced set density estimation (RSDE) technology is proposed by introducing transfer learning mechanism. It has been verified by experiments on simulation data and real data that the ML-transfer fuzzy system has a better adaptability than the traditional fuzzy modeling method in the scene with information missing.