A fuzzified neural fuzzy inference network for handling both linguistic and numerical information simultaneously
Introduction
In the real world, information usually comes in one of two types, numerical information from measurement and linguistic information according to human opinion. Linguistic information could be represented by fuzzy theory. This paper considers linguistic information represented in the form of fuzzy input and output data pairs. When linguistic information comes from a very large batch of fuzzy input and output pairs, automatic construction of a small fuzzy system from these fuzzy data is necessary. Moreover, for a complex system, it is usually difficult even for an expert to directly describe the solution by linguistic if–then rules, so numerical information from measurement is also necessary. Therefore, a neural fuzzy system that can construct automatically and simultaneously from both numerical and linguistic information is worth studying. However, most current studies focus on the construction of fuzzy rules from numerical information only [1], [4], [7], [8], [9], [11], [15], [17]. To deal with the aforementioned problems, this paper proposes a fuzzified Takagi–Sugeno–Kang (TSK)-type neural fuzzy inference network (FTNFIN) that can learn from both numerical and fuzzy input–output data pairs.
For a neural network to process linguistic information with fuzzy inputs and outputs, several approaches have been proposed [3], [6], [10]. In studies [6], [3], the authors propose an architecture of multilayer feedforward neural networks that can deal with fuzzy input vectors with fuzzy/crisp outputs. In this method, the inputs and outputs of the neural network are fuzzified using fuzzy numbers represented by α-cuts. In addition to the fuzzified neural networks above, fuzzy systems with neural learning for processing fuzzy information have also been proposed [14], [16]. In [14], input fuzzy data are viewed as fuzzy numbers being represented by α-level sets. An unusual fuzzification operation is used, where the fuzzification operation maps level sets onto a real number by measuring total differences of upper and lower limits of the level sets between input fuzzy numbers and fuzzy numbers in the antecedent part, and then uses an activation function to map the difference to the range [0,1]. In [16], fuzzy inputs are assumed to be Gaussian fuzzy sets, and in the fuzzification process, a fuzzy mutual subsethood measure is adopted to define the activation. However, the measurement is performed directly on the Gaussian functions, which is too complex and inefficient. Overall, the neural fuzzy systems above have fuzzy sets used in consequent parts, and the whole network is constructed off-line. In the proposed FTNFIN, fuzzification based on the general sup-min operation [13] is derived and the consequence is of TSK-type. For learning characteristics, FTNFIN is constructed on-line and can learn both numerical information and linguistic information simultaneously.
In the proposed FTNFIN, the input–output crisp/fuzzy data relationship is computed by using α-cut technique. The α-cut arithmetic is a general technique for fuzzy arithmetic computation. Many studies on applying α-cut technique to different fuzzy arithmetic computations have been proposed. In [2], α-cut is applied to rule interpolation. Different methods for fuzzy weighted average operation using α-cut are proposed in [12], [5]. In [18], fuzzy number neural networks using α-cut technique for network output computation and fuzzy chaotic neuron are proposed. As stated above, studies [3], [6], [10] use α-cut to compute fuzzy inputs and fuzzy/crisp outputs of neural networks. Since the fuzzy arithmetic functions in the FTNFIN are different from functions studied in the papers above, the computation process using α-cut differs and a new computation method is proposed this paper. In summary, contributions of the proposed FTNFIN are threefold. First, the proposed FTNFIN can handle fuzzy input–output data or both numerical and fuzzy data simultaneously. Second, derivation of fuzzy input–output mapping of a TSK-type fuzzy system using α-cut is proposed. The result enables FTNFIN to handle fuzzy input/output numbers with any shape. Third, FTNFIN is characterized by simultaneous structure and parameter learning ability, which enables FTNFIN to be constructed on-line. For parameter learning, a simplified learning algorithm is proposed to ease the antecedent parameter update equation and reduce computation time.
This paper is organized as follows. Section 2 introduces the basic structure and mathematical equations of FTNFIN. Structure and parameter learning algorithms of FTNFIN are presented in Section 3. In Section 4, FTNFIN is applied to several examples and comparisons with fuzzified neural networks are made. Conclusions are presented in the last section.
Section snippets
Structure of FTNFIN
This section introduces the structure of the FTNFIN, which deals with linguistic/crisp input values and outputs linguistic/crisp output values. Suppose there are two external fuzzy linguistic inputs, one fuzzy linguistic output, and three rules in the network, then this five-layered network realizes where and are fuzzy numbers, and aij is crisp number. The consequent part for the external output y is of TSK-type and
Learning of FTNFIN
Two types of learning, structure and parameter learning are used concurrently for constructing the FTNFIN. There are no rules in the FTNFIN initially, and they are created dynamically as learning proceeds upon receiving training data. Detailed structure and parameter learning algorithms are described as follows.
Simulations
To verify the performance of FTNFIN, four examples are presented in this section. In all of these examples, the teaching signal contain fuzzy inputs and output, so fuzzy neural networks that can only learn from numerical data, like [7] and [8], cannot be applied to these problems. In FTNFIN, learning parameters are set at β=0.4, ρ=0.8 and η=0.005 for all examples and learning is performed for 1000 iterations.
Example 1 (Learning from fuzzy numbers): Consider the following fuzzy if–then rules If x1 is S
Conclusions
This paper proposes FTNFIN that can process both numerical and linguistic information. Based on the technique of α-cuts, FTNFIN can deal with fuzzy inputs and fuzzy outputs with arbitrary shapes. Based on structure learning, flexible partition of the input space is achieved, so the total number of rules and fuzzy sets in FTNFIN can be reduced. For parameter learning, the proposed learning algorithm eases the antecedent parameters update equation. The simulations in Example 1, Example 2, Example
Chia-Feng Juang received the B.S. and Ph.D. degrees in control engineering from the National Chiao-Tung University, Hsinchu, Taiwan, R.O.C., in 1993 and 1997, respectively.
In 2001, he joined the faculty of Department of Electrical Engineering, National Chung Hsing University, Taichung, Taiwan, R.O.C., where he is currently a Professor. His current research interests are computational intelligence, intelligent control, computer vision, speech signal processing, and FPGA chip design.
References (18)
- et al.
Fuzzy number neural networks
Fuzzy Sets and Systems
(1999) Fuzzy weighted averages revisited
Fuzzy Sets Syst.
(2002)- et al.
Neural network implementation of fuzzy logic
Fuzzy Sets Syst.
(1992) - et al.
An efficient algorithm for fuzzy weighted average
Fuzzy Sets Syst.
(1997) - et al.
Structure identification of generalized adaptive neuro-fuzzy inference systems
IEEE Trans. Fuzzy Syst.
(2003) - P. Baranyi, D. Tikk, Y. Yam, L.T. Koczy, L. Nadai, A new method for avoiding abnormal conclusion for α-cut based rule...
- et al.
Online adaptive fuzzy neural identification and control of a class of MIMO nonlinear systems
IEEE Trans. Fuzzy Syst.
(2003) - et al.
Neural networks that learn from fuzzy if-then rules
IEEE Trans. Fuzzy Syst.
(1993) ANFIS: adaptive-network-based fuzzy inference system
IEEE Trans. Syst., Man, Cybern.
(1993)
Cited by (11)
Rapid assessment of building collapse based on sequential dynamic fusion of multi-source disaster information from news media
2020, International Journal of Disaster Risk ReductionCitation Excerpt :Therefore, to reflect the real situation of buildings collapse as much as possible, the assessment results must be updated with new information. The common methods to address the first issue are fuzzy mathematics [40–42] and neural networks [43,44]. Although fuzzy mathematics can deal with the fuzzy linguistic term, it cannot solve the randomness of fuzzy words.
An incremental support vector machine-trained TS-type fuzzy system for online classification problems
2011, Fuzzy Sets and SystemsReinforcement self-organizing fuzzy control using ant colony optimization
2020, International Journal of Computational Intelligence in ControlA real-time robust fuzzy-based level control using programmable logic controller
2015, Elektronika ir ElektrotechnikaMonitoring and Control of Information-Poor Systems: An Approach Based on Fuzzy Relational Models
2012, Monitoring and Control of Information-Poor Systems: An Approach Based on Fuzzy Relational Models
Chia-Feng Juang received the B.S. and Ph.D. degrees in control engineering from the National Chiao-Tung University, Hsinchu, Taiwan, R.O.C., in 1993 and 1997, respectively.
In 2001, he joined the faculty of Department of Electrical Engineering, National Chung Hsing University, Taichung, Taiwan, R.O.C., where he is currently a Professor. His current research interests are computational intelligence, intelligent control, computer vision, speech signal processing, and FPGA chip design.
Chun-I Lee received the B.S. and Ph.D. degrees in electrical engineering from the National Chung-Hsing University, taiwan, Taiwan, R.O.C., in 2003 and 2004, respectively. He is currently in military service, Taiwan, R.O.C. His research interests are neural networks and fuzzy systems.