(Return to ICNN97 Homepage) (Return to ICNN'97 Agenda)
ICNN'97 FINAL ABSTRACT
CI: COMPUTATIONAL INTELLIGENCE
ICNN97 Computational Intelligence Session: CI1A Paper Number: 367 Oral
Fuzzy rule networks and its applications to decision-making under uncertainty
Zhang Xinghu and How Kee Yin
Keywords: Fuzzy rule networks decision-making uncertainty
Abstract:
This paper first proposes a new method for rule representation by considering the restriction conditions and supporting conditions separately. Then it introduces a new rule inference mechanism, winner-take-all strategy, which can follow the human decision-making process more properly. This paper introduces a concept of negative evidence, a body of evidence to deny a hypothesis, which enables us to express possibility distribution over the decision space with values from -1 to 1, instead of from 0 to 1. To jointly handle positive and negative evidence, a new combination model of positive or negative evidence is proposed. On basis of the new rule representation and rule inference mechanism this paper proposes a new architecture of fuzzy neural networks, to be called Fuzzy Rule Networks (FRN) in this paper, as the structural model of rule base. This paper also derives a Delta-learning rule for the FRN networks through mathematical derivation, and provides some criteria for rule combination in the FRN. By using these criteria we can reduce the number of rules, and therefore simplify the architecture of the FRN networks. A simulation is given to show that the learning algorithm and the criteria for rule combination developed in this paper is effective.
_____
ICNN97 Computational Intelligence Session: CI1B Paper Number: 359 Oral
Evaluation ordering of rules extracted from feedforward networks
Ismail Taha and Joydeep Ghosh
Keywords: Evaluation ordering feedforward networks rules extraction
Abstract:
Rules extracted from trained feedforward networks can be used for explanation, validation, and cross-referencing of network output decisions. This paper introduces a rule evaluation and ordering mechanism that orders rules extracted from feedforward networks based on three performance measures. Detailed experiments using three rule extraction techniques as applied to the Wisconsin breast cancer database, illustrate the power of the proposed methods.
Moreover, a method of integrating the output decisions of both the extracted rule-based system and the corresponding trained network is proposed. The integrated system provides further improvements.
_____
ICNN97 Computational Intelligence Session: CI1C Paper Number: 416 Oral
A fuzzy neural network based on hierarchical space partitioning
Chu Kwong Chak, Gang Feng and Marimuthu Palaniswami
Keywords: fuzzy neural network hierarchical space partitioning adaptive fuzzy system
Abstract:
Abstract: A self-organized and adaptive fuzzy system implemented in the framework of sigmoid function neural networks is proposed. The proposed fuzzy neural network adopts the hierarchical space partitioning method so that it can generate its rules and optimize its membership functions by its hybrid algorithm. Simulation is presented to demonstrate the performance of the proposed scheme.
_____
ICNN97 Computational Intelligence Session: CI1D Paper Number: 401 Oral
A neural fuzzy approach for fuzzy system design
M. Figueiredo and F. Gomide
Keywords: neural fuzzy approach fuzzy system design fuzzy rule base
Abstract:
M. Figueiredo and F. Gomide
A new class of neural fuzzy network based on a general neuron model is introduced in this paper. The network encodes a fuzzy rule base in its structure and process data according to fuzzy reasoning schemes. It learns membership functions for each input variable and rules covering the whole input/output space. These are important decisions when designing fuzzy systems. Due to its structure the fuzzy rules encoded are trivially and explicitly recovered. The network is also shown to have the universal approximation capability. Simulation results are included to compare its features with alternative approaches and to show its usefulness as well.
For the function approximation problem, the neural fuzzy network developed here has shown to be superior from the accuracy, complexity, and system design point of view.
Keywords: Neural Fuzzy Modeling, Fuzzy System Design, Function Approximation, learning strategies
_____
ICNN97 Computational Intelligence Session: CI1E Paper Number: 150 Oral
Comparative analysis of fuzzy inference systems implemented on neural structures
Sinan Altug and Mo-Yuen Chow
Keywords: fuzzy inference systems neural structures motor fault detection
Abstract:
This paper presents comparative analysis of two popular neural fuzzy inference systems, namely, Fuzzy Adaptive Learning Control/decision Network (FALCON) and Adaptive Network based Fuzzy Inference System (ANFIS), and their application to an induction motor fault detection problem. The fault detectors are analyzed with respect to architectural and fuzzy inference system specifications, and the results for motor fault detection are evaluated in terms of fault detection accuracy, knowledge extraction capability, and computational complexity. The advantages and disadvantages of using these two architectures are also discussed. The experimental results suggest a promising future for using neural fuzzy inference systems for incipient fault detection in induction motors.
_____
ICNN97 Computational Intelligence Session: CI2A Paper Number: 128 Oral
A co-evolutionary algorithm for neural network learning
Qiangfu Zhao
Keywords: evolutionary algorithm neural network learning optimal system design
Abstract:
Title : A Co-Evolutionary Algorithm for Neural Network Learning Author : Qiangfu Zhao
Abstract :
Usually, the evolutionary algorithms (EAs) are considered more efficient for optimal system design because EAs can provide higher opportunity for obtaining the global optimal solution. However, in most existing EAs, an individual corresponds directly to a candidate of the solution, and a huge amount of computations are required for designing large-scaled systems. This paper introduces a co-evolutionary algorithm (CEA) based on the concept of divide and conquer. The basic idea is to divide the system into many small homogeneous modules, define an individual as a module, find many good individuals using existing EAs, and put them together again to form the whole system. To make the study more concrete, we focus the discussion on the evolutionary learning of neural networks for pattern recognition. Experimental results are provided to show the procedure and the performance of the CEA.
_____
ICNN97 Computational Intelligence Session: CI2B Paper Number: 647 Oral
Feedforward neural networks configuration using evolutionary programming
Manish Sarkar and B. Yegnanarayana
Keywords: artificial neural networks evolutionary programming contract bridge game
Abstract:
This paper proposes an evolutionary programming based neural networks construction algorithm, that efficiently configures feedforward neural networks in terms of optimum structure and optimum parameter set. The proposed method determines the appropriate structure, i.e. an appropriate number of hidden nodes, in such a way that locally optimal solutions are avoided. While choosing the number of hidden nodes, this method performs a trade-off between generalization and memorization. In this method, the network is evolved so that it learns an optimum parameter set, i.e. weights and bias, without being trapped into a locally optimal solution. Efficiency of this method is further enhanced by incorporating the concepts of adaptive structural mutation. Finally, efficacy of the proposed scheme is demonstrated on a Contract Bridge game opening bid problem.
_____
ICNN97 Computational Intelligence Session: CI2C Paper Number: 381 Oral
Neuro-fuzzy modeling of complex systems using genetic algorithms
Wael A. Farag and V. H. Quintana G.Lambert-Torres
Keywords: Neuro-fuzzy modeling complex systems genetic algorithms
Abstract:
In this paper, a genetic-based neuro-fuzzy (NF) approach is proposed to build and optimize fuzzy models. The learning algorithm of the Fuzzy-Neural Network (FNN) is divided into three phases. The first phase is used to find the initial membership functions of the fuzzy model. In the second phase, a new algorithm is developed to find the linguistic fuzzy rules. In the third phase, a new technique is used to apply a Genetic Algorithm (GA) to tune the membership functions of the fuzzy model optimally. A well known example is used to investigate the performance of the proposed modeling approach, and compare it with the other modeling approaches.
_____
ICNN97 Computational Intelligence Session: CI2D Paper Number: 234 Oral
A parallel hybrid genetic algorithm simulated annealing approach to finding most probable explanations for Bayesian belief networks
Ashraf M. Abdelbar and Sandra M. Hedetniemi
Keywords: parallel hybrid genetic algorithm simulated annealing most probable explanations for Bayesian belie
Abstract:
Bayesian belief networks are an important knowledge structure for reasoning under uncertainty. In the Most Probable Explanation (MPE) problem, also known as the maximum a-posteriori (MAP) assignment problem, the objective is to assign truth values to network variables in a way that will maximize their joint probability conditioned on the evidence to be explained. This problem has recently been shown to be NP-hard for general belief networks and for large networks, exact solution methods are not practical. In this paper, we present a parallel processing technique, particularly suitable for loosely-coupled multicomputers, which combines genetic algorithms with simulated annealing. This method is applied to the MPE problem on Bayesian belief network and is found to be superior on the MPE problem to either genetic algorithms or simulated annealing separately.
_____
ICNN97 Computational Intelligence Session: CI2E Paper Number: 672 Oral
An evolutionary programming-based probablistic neural network construction technique
Manish Sarkar and B. Yegnanarayana
Keywords: probablistic neural networks Gaussian mixture evolutionary programming fuzzy
Abstract:
Quick learning ability of a probabilistic neural network (PNN) has made it a popular alternative to feedforward neural networks trained by backpropagation algorithm. However in real life applications, where training set size is quite large, the PNN model suffers from a huge memory overhead and a long testing time. Moreover, its generalization capability critically depends on the choice of certain architectural parameters, which are presently chosen in an adhoc basis. Since the PNN can be viewed as a Gaussian mixture model, above drawbacks can be avoided if the PNN is configured based on a Gaussian mixture model with an optimum parameter set. In this paper, an evolutionary programming based clustering technique is employed to determine the optimum parameter set of the Gaussian mixture model. The efficacy of the proposed scheme is demonstrated on a Contract Bridge game opening bid problem.
_____
ICNN97 Computational Intelligence Session: CI3A Paper Number: 149 Oral
An optimizing FAMLB rules algorithm for traffic control in ATM communication networks
Zhifeng Jiang and Zemin Liu
Keywords: FAMLB rules traffic control ATM communication networks
Abstract:
_____
ICNN97 Computational Intelligence Session: CI3B Paper Number: 616 Oral
Neural fuzzy agents that learn profiles and search databases
Sanya Mitaim and Bart Kosko
Keywords: fuzzy sets database Neural fuzzy agents fuzzy rules
Abstract:
This paper shows how a neural fuzzy system can help learn an agent profile of a user. The fuzzy system uses if-then rules that store and compress the agent's knowledge of the user's likes and dislikes.
A neural system uses training data to form and tune the rules. The profile is a preference map or a bumpy utility surface over the space of search objects. Rules define fuzzy patches that cover the bumps as learning unfolds and as the fuzzy agent system gives a finer approximation of the profile. The agent system searches for preferred objects with the learned profile and a new fuzzy measure of similarity. We derive a new supervised learning law that tunes this matching measure with new sample data. Then we test the fuzzy agent profile system on object spaces of flowers and sunsets and test the fuzzy agent matching system on an object space of sunset images.
_____
ICNN97 Computational Intelligence Session: CI3C Paper Number: 649 Oral
Neuro-Fuzzy Approaches to collaborative Scientific computing
N. Ramakrishnan, Anupam Joshi, Elias N. Houstis, John R. Rice
Keywords: Learning Supervised classification
Abstract:
Rapid advances in High Performance Computing (HPC) and the Internet are heralding a paradigm shift to network--based scientific software servers, libraries, repositories and problem solving environments. According to this new paradigm, vital pieces of software and information required for a computation are distributed across a network and need to be identified and `linked' together at run time; this implies a `net--centric' and collaborative scenario for scientific computing. This scenario requires the application to dynamically choose the best among several competing resources that can solve a given problem. For these systems to become ubiquitous, efficient mechanisms for collaboration and automatic inference of the abilities of multiple `compute servers' need to be established. In this paper, we demonstrate a methodology to facilitate collaborative scientific computing. Our idea comprises of (i) a concept of `reasonableness' to automatically generate exemplars for learning the mapping from problems to `servers' and (ii) a neuro--fuzzy technique developed earlier by the authors that conducts supervised classification on the exemplars generated. Our techniques work in an on--line manner and cater to mutually non--exclusive classes which are critical in the collaborative networked computing landscape.
_____
ICNN97 Computational Intelligence Session: CI3D Paper Number: 19 Oral
A new fuzzy classifier with triangular membership function
Yong-Sheng Yang, Francis H. Y. Chan, F. K. Lam, Hung Nguyen
Keywords: Fuzzy classifier Membership functions Clustering
Abstract:
Fuzzy logic is widely applied in control and modeling for its robustness, simplicity and clarity. It is also applied in classifier design with rules directly generated from numerical data. Some available rule generation methods, however, are either too complicated to implement or impractical for high dimensions. In this paper, we propose a new fuzzy classifier architecture. At the very beginning the training data is clustered at the input space. Fuzzy sets are then defined based on these clusters with triangular membership function. The outputs in the rule conclusion are initially determined by the "normalized vote" in the corresponding cluster. Fuzzy sets and conclusions can be adjusted through training. The proposed fuzzy system is simple in structure, and can be fast trained and easily implemented. Its classification performance is generally better than artificial neural network.
_____
ICNN97 Computational Intelligence Session: CI3E Paper Number: 376 Oral
Stability analysis of neural networks
Thomas Feuring and Andreas Tenhagen
Keywords: Stability analysis neural networks fuzzy network
Abstract:
Neural networks can only be trained with crisp and finite data set. Therefore stability analysis seams to be impossible. In this article we propose a new method how stability for neural networks can be proven. Here we use fuzzy input and output data for the training process. After the learning phase the fuzzy network will be defuzzified. Using special properties of fuzzy neural networks the output behaviour can be estimated. This gives us the ability of proofing stability for neural networks.
_____
ICNN97 Computational Intelligence Session: CI3F Paper Number: 82 Oral
Task Allocation for multiple-network architectures
Thorsten Drabe, and Wolfgang Bressgott
Keywords: Task allocation Multiple network architectures Genetic algorithms
Abstract:
Modular neural architectures pose the problem to find those subtasks of a complex task which can be efficiently trained together on the same network. We attack the involved combinatorial optimization problem by a genetic algorithm. For comparison a monolithic network and a modular architecture with random task distribution are considered. Letter recognition experiments show that the proposed method yields considerably better results concerning final convergence speed, generalization and completeness of solutions.
_____
ICNN97 Computational Intelligence Session: CIP1 Paper Number: 22 Poster
Retaining diversity of search point distribution through breeder genetic algorithm approach
D. Popovic, K. C. S. Murty
Keywords: Genetic algorithms Training Optimization
Abstract:
_____
ICNN97 Computational Intelligence Session: CIP1 Paper Number: 186 Poster
Knowledge-based fuzzy MLP with rough sets
Mohua Banerjee, Sushmita Mitra and Sankar K. Pal
Keywords: Knowledge-based networks fuzzy MLP rough sets classification
Abstract:
A new scheme of knowledge encoding using rough set-theoretic concepts is proposed. Knowledge collected from a data set is initially encoded among the connection weights of a fuzzy MLP. The network is then refined during training. Results on real data demonstrate that the speed of learning and classification performance of the proposed scheme are better than that obtained with the fuzzy and conventional versions of the MLP (involving no initial knowledge).
_____
ICNN97 Computational Intelligence Session: CIP1 Paper Number: 189 Poster
Fuzzy neurons and fuzzy multilinear mappings
K. S. Abdulkhalikov, C. Kim and H. S. Cho
Keywords: fuzzy neurons fuzzy multilinear mappings fuzzy sets
Abstract:
There are several mathematical models of fuzzy neurons. Usually, input values of them are fuzzy numbers(that is, fuzzy sets in one-dimensional space) with triangular membership functions. The aggregating operations may be one of the T-norms or T-conorms. There are currently few, if any, learning methods proposed in the literature. We propose to consider fuzzy linear spaces as fuzzy inputsof fuzzy neuronsand offer mathematical thoery to work with this notions. In particular, we study fuzzy multilinear maps of fuzzy linear spaces. If model of neuron would carry out only linear operations, mathematical attractiveness of this one can be disappeared. We can overcome this obstacle by considering multidimensional linear space.
_____
ICNN97 Computational Intelligence Session: CIP1 Paper Number: 212 Poster
An intelligent sales forecasting system through fuzzy neural network
R. J. Kuo and K. C. Xue
Keywords: sales forecasting fuzzy neural network fuzzy logic
Abstract:
_____
ICNN97 Computational Intelligence Session: CIP1 Paper Number: 615 Poster
Adaptive joint fuzzy sets for function approximation
Sanya Mitaim and Bart Kosko
Keywords: fuzzy sets function approximation
Abstract:
This paper presents a new method to create and tune joint fuzzy sets. Multidimensional fuzzy sets define the if-part fuzzy sets of rules in feedforward fuzzy function approximators. These joint set functions do not factor into a product of scalar fuzzy sets (such as triangles or bell curves) and so they do not ignore the correlation structure among the input components. The joint set functions transform a scalar distance measure that preserves the correlation structure. Supervised learning tunes the metrical joint set functions and tunes the scalar set functions that make up factorable joint set functions. Factorable joint set functions tend to collapse to spikes in high dimensions. This holds for all joint set functions that combine factors with product or minimum or other t-norms. Simulations suggest that some metrical joint set functions may offer a practical tool for fuzzy function approximation in higher dimensions and in $L^p$ function spaces.
_____
ICNN97 Computational Intelligence Session: CIP1 Paper Number: 552 Poster
A neuro-fuzzy model reduction strategy
G. Castellano and A. M. Fanelli
Keywords: neuro-fuzzy model reduction simplification fuzzy system
Abstract:
A neuro-fuzzy model reduction strategy
This paper presents an approach to obtain simple fuzzy models. The simplification strategy involves structure reduction of a neural network modeling the fuzzy system and is carried out through an iterative algorithm aiming at selecting a minimal number of rules for the problem at hand. The selection algorithm allows manipulation of the neuro-fuzzy model to minimize its complexity and to preserve a good level of accuracy. Experimental results demonstrate the algorithm's effectiveness in identifying reduced neuro-fuzzy networks with no degradation in the original performance.
_____
ICNN97 Computational Intelligence Session: CIP1 Paper Number: 344 Poster
Competitive hybrid neuro-fuzzy models for supervised classification
Nicola Giusti, Francesco Masulli and Alessandro Sperduti
Keywords: Competitive hybrid neuro-fuzzy models supervised classification fuzzy basis function networks
Abstract:
Neuro-fuzzy systems are often very complex and may require long training times. In the context of supervised classification, we propose a competitive and a hybrid model based on Fuzzy Basis Function networks. These models are fast to train and still hold very good generalization performances. Experimental results on the classification of handwritten digits are presented.
_____
ICNN97 Computational Intelligence Session: CIP1 Paper Number: 293 Poster
Extraction method of rules from reflective neural network architecture
Takumi Ichimura, Noboru Matsumoto, Eiichiro Tazaki and Katsumi Yoshida
Keywords: reflective neural network network architecture learning procedure
Abstract:
Reflective Neural Network is a new architecture with a learning procedure for systems composed of many networks based on network module concept. To learn a subset of the complete set of training data, each module has two kinds of feed-forward networks; a monitor network and a worker network. A monitor network estimates how good a worker network is for distributed training data. In this paper, we propose a extraction method of fuzzy rules from the modified network based on Reflective Neural Network. To verify the validity and the effectiveness of the proposed method, we develop a medical diagnostic system for thyroid diseases.
_____
ICNN97 Computational Intelligence Session: CIP1 Paper Number: 588 Poster
A training technique for fuzzy number neural networks
James Dunyak and Donald Wunsh
Keywords: training fuzzy number neural networks
Abstract:
A new technique is discussed for the training of fuzzy neural networks with general fuzzy number inputs, weights, and outputs. Fuzzy number neural networks are difficult to train because of the many alpha-cut constraints implied by the fuzzy weights. In this paper, an unconstrained representation is used for the fuzzy weights, allowing application of a standard backpropagation approach. The technique is demonstrated on a moderately large problem.
_____
ICNN97 Computational Intelligence Session: CIP1 Paper Number: 507 Poster
Fuzzy logic, neural networks, and brain-like learning
Asim Roy and Raymond Miranda
Keywords: Fuzzy logic neural networks brain-like learning
Abstract: