On The Dynamical Modeling With Model-Free Estimators

• Main Authors: Feng-Yu Yang, 楊豐毓
• Other Authors: 蘇順豐
• Format: Others
• Language: en_US
• Published: 2001
• Online Access: http://ndltd.ncl.edu.tw/handle/43122354557464014609

碩士 === 國立臺灣科技大學 === 電機工程系 === 89 === Abstract Traditional statistical approaches for system modeling require mathematical description about how the outputs functionally depend on the inputs. Recently, neural networks and fuzzy systems have been widely used for modeling nonlinear systems. Those approaches are often referred to as model-free estimators. In fact, they are also proven to be universal approximators. Those networks are all static ones. A dynamical system is a system of which outputs are functions of not only current input, but also past inputs and past outputs. Thus, modeling a dynamical system can be viewed as a temporal or associative memory problem. When neural networks or fuzzy systems model dynamical systems, the often-used approach is to consider all necessary past inputs and outputs as explicit inputs. It is clear that if we want to model a nonlinear dynamical system accurately using static models, one must know the order of the considered system in advance. In the literatures, researchers introduced delay feedback networks, which can be regarded as networks with “internal memories”, and claimed that those methods could have good performance on modeling dynamical systems. In the thesis, we have studied those delay feedback networks and proposed a new version of delay feedback networks. From simulation, the proposed networks have the best performance among those existing delay feedback networks. We also showed by examples that those delay feedback networks can only reach the accuracy of static systems with order two and proved that the number of delays in delay feedback networks plays the same role as the order in static networks. Finally, we attempted to embed prior knowledge into the used neural fuzzy system. With such inclusion of knowledge, the learning performance can be improved.