Online Evolving Interval Type-2 Intuitionistic Fuzzy LSTM-Neural Networks for Regression Problems

Abstract

Since the existence of fuzziness in the real world, uncertainty could be always present in the reasoning results. Therefore, how to tackle with the hesitation existing in the process of reasoning is a meaningful issue for the construction of the inference system. In this paper, a novel interval type-2 intuitionistic fuzzy neural network based on long-short term mechanism is proposed (LSTM-IT2IFNN). By means of interval type-2 intuitionistic set, the hesitation of reasoning is described and involved to determine fuzzy rules in inference system, by which the chosen of fuzzy rules not only considers the membership value but also non-membership degree. In order to handle the regression problems with long-term time dependency, long-short term mechanism is introduced into fuzzy neural networks, which can effectively improve the memory ability of long-term knowledge. Furthermore, by means of intuitionistic sets, an extended metacognitive learning mechanism is proposed for the adaptively online learning of network structure, where it achieves when, what, and how to learn from the sequential data. The performance of the LSTM-IT2IFNN is evaluated on different regression problems, including synthetic and real-world data. By comparing to existing algorithms, the experimental results show the efficiency and accuracy of the proposed scheme.