报告题目：Fuzzy Discrete Event Systems with Online Supervised Learning Capability
报告人: Hao Ying
讲座人简介：Professor Hao Ying has published two fuzzy control books, 126 journal papers, and 160 conference papers. He is ranked among top 25% of the 100,000 most-cited authors across all 22 scientific fields (176 subfields) which are selected from nearly 7 million scientists worldwide. He is serving as an Associate Editor or a Member of Editorial Board for 13 international journals, including the IEEE Transactions on Fuzzy Systems and the IEEE Transactions on Systems, Man, and Cybernetics: Systems. He served as a Member of Fellows Committee of both the IEEE Computational Intelligence Society (2020 and 2021) and the IEEE Systems, Man, and Cybernetics Society (2016, 2017, 2020).
To effectively represent deterministic uncertainties and vagueness as well as human subjective observation and judgment encountered in many real-world problems especially those in medicine, we recently originated a theory of fuzzy discrete event systems (DES). The theory is unique in that it is capable of modeling a class of event-driven systems as fuzzy automata with states and event-invoked state transitions being ambiguous. We introduced fuzzy states and fuzzy event transition and generalized conventional crisp DES to fuzzy DES. The largely graph-based framework of the crisp DES was unsuitable for the expansion and we thus reformulated it using state vectors and event transition matrices which could be extended to fuzzy vectors and matrices by allowing their elements to take values in [0, 1]. We also extended optimal control of DES to fuzzy DES. This novel fuzzy DES theory is consistent with the traditional DES theory, both at conceptual and computation levels, in that the former contains the latter as a special case when the membership grades are either 0 or 1.
We further developed the FDES theory so that it possessed self-learning capability. More specifically, we use stochastic gradient descent to develop online learning algorithms for the fuzzy automata (i.e., learning the event transition matrix from data). We uncover an inherent obstacle in the initial derived algorithms that fundamentally restricts their learning capability owing to dependences of the model parameters to be learned. We develop a novel mechanism to not only overcome the obstacle but also make the learning adaptive. Our final algorithms can (1) learn an event transition matrix based on automaton’s states before and after the occurrence of a fuzzy event, and (2) learn the transition matrix and multi-dimensional Gaussian fuzzy sets yielding automaton’s pre-event states from relevant input (physical) variables and target states. Computer simulation results are presented to show learning performance of the final algorithms.