Please use this identifier to cite or link to this item:
Title: Multistage fuzzy neural networks : architectures, algorithms and analysis
Authors: Duan, Ji-cheng
Degree: Ph.D.
Issue Date: 1999
Abstract: In the past couple of years, there have been increasing interests in the integration of neural networks and fuzzy logic, aiming at deriving intelligent systems that combine their corresponding strengths and eliminate their individual weaknesses. The resulted hybrid systems correspond to one of the most active and fruitful research area, i.e., fuzzy neural network (FNN). Most of the existing FNN models have been proposed to implement different types of single-stage fuzzy reasoning mechanisms where the consequence of a rule can not be used as a fact to another rule. It is well known that single-stage fuzzy reasoning suffers from the dimensionality problem indicating that the number of rules increases exponentially with the number of inputs. Moreover, human beings usually take use of more sophisticated reasoning mechanisms when handling complex decision making problems. Compared with the fruitful developments of single-stage FNN models, the integration of neural networks with high-level reasoning strategies has never been systematically studied. FNN modeling based on multistage fuzzy reasoning (MSFR), where the consequence of a rule is passed to another rule as a fact, is pursued in this dissertation. It has been pointed out in the literature that MSFR is essential to effectively build up a large-scale system with a high degree of intelligence. We propose to incorporate MSFR into FNN construction and study three basic multistage fuzzy neural network (MSFNN) architectures, namely, incremental, aggregated and cascaded architectures. Three MSFNN models, namely, incremental FNN (IFNN), aggregated FNN (AFNN), and cascaded FNN (CFNN) have been designed. The resulted new models implement comprehensive fuzzy inference and correspond to three typical ways of reasoning carried out by human beings. IFNN is to reason incrementally by considering some important factors (input variables) first, making an approximate decision (intermediate variable), and then fine tuning it by considering more factors until a final decision (output variable) is made. AFNN is to reason in a mixture-of-expert manner, i.e., first considering some sets of correlated/coupled factors independently and then combining the decisions to form a more judicious one. CFNN corresponds to human being's syllogistic fuzzy inference and chain of reasoning mechanisms. The incremental and aggregated architectures can address the dimensionality problem fundamentally by assigning inputs in hierarchical manners. However, it is not trivial to determine an optimal or sub-optimal incremental (aggregated) architecture when there is no a priori knowledge available (which is usually the case in most of the practical applications). The existing input selection methodologies, e.g., genetic algorithms, heuristic searching and complete/exhaustive searching strategies, are all computation prohibitive when they are applied to solving a high-dimensional problem. We have particularly addressed this problem of distributing input variables to different reasoning stages for IFNN and AFNN. In this regard, we propose two fast and systematic input selection approaches to determine appropriate incremental and aggregated architectures, respectively. The proposed approaches are distinctive by the properties that no a priori expert knowledge is required and the computation complexity is greatly simplified compared with the other existing input selection methods. Specifically, the incremental architecture is determined by distributing inputs among the first, intermediate and final reasoning stages according to their individual importance degrees (contributions) to the output. For the aggregated one, inputs are only allowed to pass to the first reasoning stage consisting of a number of independent sub-stages. The inputs of each sub-stage are selected by analyzing the coupling magnitudes among them. The outputs from the first stage form the inputs to the successive stage and such an arrangement can be extended for more stages. The last type of MSFNN architecture, i.e., cascaded one, exempts from input selection since it has all the inputs being fed to the first reasoning stage whose inference results, the intermediate ones, are then fed to the subsequent stages in a stage-by-stage manner. Those three MSFNN models are not simple extensions of conventional single-stage FNN's but implementing comprehensive MSFR in more flexible-connected neural network architectures on a higher level. Hybrid learning algorithms have been developed to extract corresponding MSFR rule sets from stipulated data pairs. Mamdani and TSK fuzzy reasoning mechanisms are adopted in IFNN and AFNN for their popularity. In Mamdani type IFNN and AFNN, a rough fuzzy rule set is firstly derived using competitive learning and the system parameters are then fine-tuned by back-propagation. The consequent and premise parameters of TSK type IFNN and AFNN models are trained by LSE and back-propagation, respectively. CFNN only implements Mamdani fuzzy reasoning. Its learning algorithm is also a hybrid one containing genetic algorithm and back-propagation. Several benchmarking problems have been simulated in this dissertation. The simulations show that IFNN and AFNN models are superior to their single-stage counterparts in used resources (e.g., fuzzy rules, fuzzy terms, fuzzy operations, etc.), convergence speed, robustness and generalization ability. Besides implementing comprehensive Mamdani syllogistic fuzzy reasoning, CFNN is distinctive by its learning abilities and robustness, particularly when the number of inputs is not large. Those three MSFNN models are not exclusive to each other and the hybridization of them are discussed in the dissertation.
Subjects: Neural networks (Computer science)
Fuzzy systems
Computer network architectures
Hong Kong Polytechnic University -- Dissertations
Pages: x, 160 leaves : ill. ; 30 cm
Appears in Collections:Thesis

Show full item record

Page views

Last Week
Last month
Citations as of Jun 4, 2023

Google ScholarTM


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.