Home      About Journal      Editorial Board      Instructions      Ethics Statement      Contact Us                   中文

Judgement and Disposal of Academic Misconduct Article Copyright Transfer Agreement Proof of Confidentiality Requirements for Electronic Version More.... Chinese Association of Automation National ResearchCenter for Intelligent Computing System Institute of Intelligent Machines,Chinese Academy of Sciences More....

2006 Vol.19 Issue.3, Published 2006-06-25 Papers and Reports    Researches and Applications        Papers and Reports 281 ptimization Technique Based on Evolutionary Algorithm and Dynamic Bayesian Network XIAO QinKun, GAO XiaoGuang A new optimization technique for dynamic system is proposed to achieve autonomous control under complicated environment. Firstly, Dynamic Bayesian Network (DBN) is incorporated into evolutionary algorithm as a transfer network from t to t+1 generation. Through DBN, the original static optimization process of evolutionary algorithm based on Bayesian optimization algorithm (BOA) is effectively changed into the dynamic process. Using this scheme, the DBN transfer network can reestablish optimization direction for system to adapt to various changes of environment. The scheme can help agent to complete a series of complex tasks without intervention from users. The experimental results clearly demonstrate the accuracy and effectiveness of method. Secondly, new concepts are introduced to increase optimization speed and meet realtime requirement. One is Restriction Function, which is used to cut off unnecessary nodes during evolutionary computation, and the other is Replacement, which is used to inherit part of good results of former generation evolutionary. The new concepts are used to make the evolutionary optimization process more efficient . 2006 Vol. 19 (3): 281-288 [Abstract] ( 257 ) [HTML 1KB] [ PDF 594KB] ( 686 ) 289 An Effective High Attribute Dimensional Sparse Clustering ZHAO YaQin, ZHOU XianZhong, HE Xin, WANG JianYu Clustering analysis is one of the most important techniques in data mining with scale, dimension and sparseness of dataset being three key factors that influence accuracy of clustering. An effective clustering algorithm for the high attribute dimension sparse data is proposed in this paper. Definitions are given, such as sparse similarity, similarity between equivalence relations and generalized equivalence relation. Based on these definitions, the theory of equivalence relation is applied to form initial clusters. Initial equivalence relations are modified in terms of the similarity between two equivalence relations in order to obtain more reasonable clustering results. High dimensional sparse data is effectively compressed and expressed as sparse feature vector whose dimension is far lower than that of original data. As a result, the proposed approach can handle an array of high dimensional sparse data with high efficiency, and be independent of sequence of the objects. 2006 Vol. 19 (3): 289-294 [Abstract] ( 215 ) [HTML 1KB] [ PDF 378KB] ( 490 ) 295 Segment Procedure Neural Network LIANG JiuZhen Segment procedure neural network model and its learning algorithm are studied in this paper. A novel procedure neural network is proposed based on procedure neural networks. The algorithm is also proposed in the cases of knowing and unknowing the subsection desire outputs, which aims at simulating the procedure of subsection object programming and estimation system. Finally, an application example of undergraduate integration estimation system is presented and simulation results illustrate the efficiency of the model and the algorithm. 2006 Vol. 19 (3): 295-299 [Abstract] ( 229 ) [HTML 1KB] [ PDF 462KB] ( 431 ) 300 A Novel Voice Conversion Method Based on Codebook Mapping with PhonemeTied Weighting WANG ZiXiang, DAI LiRong, WANG YuPing, WANG RenHua The voice conversion system framework is introduced in this paper. Further, the conventional codebook mapping method for voice conversion is discussed. This paper point out that the conventional codebook mapping method, which calculates the weighting coefficients based on whole codebooks, tends to generate overly smoothed effect on converted speech spectrum. So the converted speech quality is decreased greatly. To address this problem, a novel voice conversion method based on codebook mapping with phonemetied weighting is presented. And a new decision tree based prosodic conversion method is also proposed. The experiments show that the proposed methods can effectively convert speaker's individuality while maintaining high speech quality with only a small amount of training data. 2006 Vol. 19 (3): 300-306 [Abstract] ( 275 ) [HTML 1KB] [ PDF 681KB] ( 321 ) 307 A Genetic Algorithms Based Mercer Kernel Clustering Method ZHOU LinFeng, DING YongSheng Based on Mercer kernel function theorem and genetic algorithms, a novel genetic algorithms based mercer kernel clustering method is proposed. Using Mercer kernel function, an elegant way to map the input space nonlinearly to the high dimensional feature space is gotten, which manifests the feature differences of samples, so that it not only gets more accurate clustering results but also speeds up the convergence rate of the algorithm. And, integrated with genetic algorithm, it can overcome the problem of local minima. Simulation experiments and an application to the objective evaluation of textile products verify its feasibility and effectiveness. 2006 Vol. 19 (3): 307-311 [Abstract] ( 220 ) [HTML 1KB] [ PDF 399KB] ( 458 ) 312 An Optimal Method on Uncorrelated Discriminant Vectors Based on Perturbation Analysis WANG WeiDong, ZHENG YuJie, YANG JingYu This paper, on the basis of uncorrelated image projection discriminant analysis, focuses on perturbation features of eigenvalue and eigenvector, pointing out eigenvector of morbid eigenvalue may be perturbed to a great degree. So, if the eigenvector is used as a projection axis to project, the feature vectors achieved cannot provide valid discriminant information. Therefore, an optimal method to uncorrelated discriminant vectors is proposed in this paper. And the method is tested on ORL face database. The experimental results indicate the method can simplify projection matrix, improve the efficiency of features extraction and then make the recognition ratio robust. Moreover, this paper suggests the optimal method based on perturbation analysis is suitable for optimizing other linear discriminant vectors. 2006 Vol. 19 (3): 312-317 [Abstract] ( 197 ) [HTML 1KB] [ PDF 337KB] ( 323 ) 318 Attribute Adjustment Strategies of Entity Agents in a Computer Aided Urban Planning System JIANG YunLiang, ZHUANG YueTing, XU CongFu, LIU Yong Computer Aided Urban Planning System (CAUPS) is an object and agent based computer aided urban planning system that uses the entity agents as its basic units during the process of urban planning. The evaluation factors of multiple entity agents system, i.e. the communication mode and attribute adjustment strategies of multiple entity agents, are analyzed, then a KQML based multiple agents communication mode is introduced. Based on this communication mode, the attribute adjustment function between multiple entity agents is realized. The reciprocity relationships of agents in CAUPS are divided into three basic types, i.e. single entity agents attribute adjustment, dual entity agents attribute adjustment and transitive entity agents attribute adjustment. Finally, the weak synchronization algorithm is presented to solve the problems of the single entity agents attribute adjustment and the dual entity agents attribute adjustment. And the strong synchronization algorithm is proposed to deal with the transitive entity agents attribute adjustment. 2006 Vol. 19 (3): 318-324 [Abstract] ( 187 ) [HTML 1KB] [ PDF 675KB] ( 374 ) 325 Design of MultiObjects RealTime Tracking System Based on Genetic Algorithms YANG ShuYing,HE PeiLian The detection and recognition about moving objects in video tracking system are presented in this paper. Moving objects are separated from the background by multiframe difference. According to the average value in different area, the dynamic threshold can be found, and the image is transformed into binary one. Then some invariant characteristics are extracted and used for recognition by genetic algorithm. The experiment result shows the method can detect moving targets and cluster them quickly. 2006 Vol. 19 (3): 325-330 [Abstract] ( 208 ) [HTML 1KB] [ PDF 769KB] ( 559 ) 331 The Optimization for Traffic Signal Based on Improved Immunogenetic Algorithm GU Rong, CAO LiMing, WANG XiaoPing In this paper the basic principle of immunogenetics is described, and the traditional immunogenetics algorithm is improved. The mechanism that the antibody twice responds to the antigen is simulated. Information entropy is utilized to compute the affinity between antigens and the antibodies that have high affinity and low similarity are inherited to next generation. Best antibodies are kept in memory set then participate in evolution, which can make the algorithm avoid losing in the local optimal solution. A new phase timing optimization algorithm is proposed to discuss the problem of traffic signal control. An experiment for the traffic model at a fourphase single intersection is designed with this algorithm, and the simulation results show its feasibility and effectiveness. 2006 Vol. 19 (3): 331-337 [Abstract] ( 270 ) [HTML 1KB] [ PDF 418KB] ( 446 ) 338 Properties of Homomorphism Rough Groups LIN RenBing, WANG JiYi Let φ be a surjective homomorphic mapping from group G to group K. Through the congruence relation ρ from K to G by φ, the property, that the surjective homomorphic mapping φ can commutate with the upper (lower) approximation of the congruence relation, is obtained. It is shown that a homomorphic mapping keeps upper (lower) rough subgroups, and upper (lower) rough left (right, twosided) ideally unchanged. 2006 Vol. 19 (3): 338-341 [Abstract] ( 254 ) [HTML 1KB] [ PDF 230KB] ( 304 ) Researches and Applications 342 The Improved Pattern Recognition Method for Process Optimization and Application YAN JingYu, SHEN ZhiYu, XUE MeiSheng, SUN DeMin Aiming at difficulties in applying pattern recognition optimization method to the process optimization, the concepts of sample weight and class gradient are proposed to improve the existing algorithms according to the feature weight and gradient theory. Simulation results in a nonlinear, multivariable and strongly coupled system illustrate that the proposed method outperforms the conventional pattern recognition optimization method and adaptive optimization method. The proposed method has been applied in the practical online operation condition optimization of an ammonia synthesizer (80,000 tons per year) and the net value of ammonia has raised 0.38% with considerable economic benefit. 2006 Vol. 19 (3): 342-348 [Abstract] ( 241 ) [HTML 1KB] [ PDF 822KB] ( 419 ) 349 A Fast Watershed Algorithm Based on Drain Simulations ZHANG LiDong, BI DuYan In this paper, a new fast algorithm for computing watershed in digital grayscale images is introduced. The presented algorithm is based on drain process. The flooding process of traditional algorithm is replaced by finding regional maximum on curve of section .The results show that the proposed one has lower memory requirement and is much faster than the traditional one. Moreover, it combines with recognition which can control the complexity of the segmentation to reduce the over segmentation and computing complexity. 2006 Vol. 19 (3): 349-356 [Abstract] ( 240 ) [HTML 1KB] [ PDF 1330KB] ( 687 ) 357 A Method of Speech Enhancement Using Weighting Factors PIAO ChunJun, CUI ShuangXi In this paper, a weighting factor for alleviating the speech distortion and reducing the residual noise in the speech enhancement system is derived. The function of weighting factor is to keep the energy of residual noise lower than the noise masking threshold and the speech distortion smaller than the residual noise. With Matlab language, the emulational results show the enhancement speech using the proposed method sounds more natural and leads to a significant reduction of the unnatural structure of the residual noise while alleviating the speech distortion. 2006 Vol. 19 (3): 357-361 [Abstract] ( 228 ) [HTML 1KB] [ PDF 1311KB] ( 455 ) 362 3D PolarRadiusInvariantMoments and Their Application to 3D Model Retrieval LI ZongMin,YU GuangBin, LIU YuJie, LI Hua A novel moment, 3D polarradiusinvariantmoment, is proposed for the 3D object recognition and classification. Some properties of the new moments including the invariance on shift, rotation and scale transforms are studied and proved. Eighteen moment invariants are derived and tested. Examples are presented to illustrate the performance and invariance of these moments. With the help of these moment invariants, the 3D models are distinguished accurately. 2006 Vol. 19 (3): 362-367 [Abstract] ( 219 ) [HTML 1KB] [ PDF 1552KB] ( 614 ) 368 Automated Classification for Celestial Spectra Based on Cover Algorithm YANG JinFu, WU FuChao, LUO ALi, ZHAO YongHeng Automated classification for large numbers of celestial spectra is one of the most important problems which are urgent to be solved in large survey projects. In this paper, an automated classification method for celestial spectra based on cover algorithm is presented. In the training procedure, some representative spectra of the training spectrum set can be obtained by utilizing a certain cover rule. Then, in the classification phase, the classifier computes the distances between each test spectrum and the representatives, and the class of the test spectrum is determined by the representative spectrum closest to the test spectrum. The experimental results show that this method is better than SVM both in training speed and classifying accuracy over Normal Galaxies(NGs), Stars, Active Galaxies(AGs) and Active Galactic Nucleus(AGNs) datasets. And the proposed method is promising for the automated classification of large numbers of celestial spectra collected by large survey projects. 2006 Vol. 19 (3): 368-374 [Abstract] ( 268 ) [HTML 1KB] [ PDF 426KB] ( 446 ) 375 Quantitative Rule Lattice and Its Incremental Construction LI Yun, LIU ZongTian, CHEN Ling, CAI JunJie The key for extracting the minimal nonredundant rule is to obtain the least itemsets in the Set of Frequent Itemsets with the Same Tidset(SFIST) which corresponds to the frequent closed itemset. To extract such rules using the concept lattice conveniently, the Quantitative Rule Lattice(QRL) with new data structure is presented, the way obtaining the set of the least itemsets in the SFIST of the node in QRL is discussed and the relevant algorithm is provided. Due to incremental formation of QRL and the set of the least itemsets, the QRL is very suitable for both extracting the minimal nonredundant rules from the dynamic database and realizing the rule updated incrementally. 2006 Vol. 19 (3): 375-381 [Abstract] ( 222 ) [HTML 1KB] [ PDF 354KB] ( 253 ) 382 The Analysis of CMAC Generalization LIN XuMei, MEI Tao, LUO MinZhou, SONG YanFeng Generalization is very important in Cerebellar Model Articulation Controller (CMAC). If CMAC has good generalization, it will have high precision. In this paper, the principle, structure and learning algorithm of CMAC are described. The relationship between the quantification precision and sampling precision that influences the generalization is discussed theoretically. Simulation results show the correctness of relationship between the quantification and sampling precision, and the conclusion that the quantification precision should be higher than the sampling precision is gotten. Moreover, a new kind of optimization based on pareto genetic algorithm (PGA) about generalization parameter and quantification precision is proposed. Experimental results show the correctness of the new method. 2006 Vol. 19 (3): 382-387 [Abstract] ( 274 ) [HTML 1KB] [ PDF 498KB] ( 444 ) 388 Research on Target Recognition Based on Classification IntervalWavelets Network ZHU ZhiYu, ZHANG Bing, LIU WeiTing A method to construct intervalwavelets is introduced in this paper, which is combined with neural network. A model of intervalwavelets neural network is proposed to classify signals, which can solve the problem that basic space does not match the space of learnt signals. Analogue annealing strategy is introduced into the model, and adaptive varied learning coefficient is applied to train the network. The experimental results show that the number of neurons can be decreased, the convergence rates also can be improved and “dimension disaster ” problem is solved with better classification effect as well when applying intervalwavelets neural network in Radar Target Recognition. 2006 Vol. 19 (3): 388-392 [Abstract] ( 290 ) [HTML 1KB] [ PDF 369KB] ( 279 ) 393 Uncertain Information Clustering Based on DempsterShafer Theory and HCM CAO KeJin, ZHAO ZongGui, JIANG Han When analyzing multisource information, it is necessary to cluster the information according to their sources. In this paper, a problem of clustering multisource information denoted by evidence is investigated, and an evidence clustering standard is given. In addition, an idea of transformation from the evidence interspaces to Euclidean interspaces is presented in this paper, then the HCM is used to cluster the multisource information. Based on the theory, a simple example of passive sensors ESM tracking aerial target is demonstrated. 2006 Vol. 19 (3): 393-399 [Abstract] ( 208 ) [HTML 1KB] [ PDF 421KB] ( 336 ) 399 Discrete Particle Swarm Optimization Algorithm for Independent Task Assignment Problem ZHONG YiWen, YANG JianGang A discrete particle swarm optimization algorithm is designed to tackle the independent task assignment problem in heterogeneous computing systems. Based on the characteristics of discrete variable, particle’s position, velocity and their operation rules are redefined in this paper. In order to restrain premature stagnation, individual diversity of particle and microdiversity of particle swarm are defined. A repulsion operator is designed to keep the diversity of particle swarm, and a learning operator is defined to improve intensification ability of the algorithm. The proposed algorithm gets good balance between exploration and exploitation using those operators. The simulation results show the proposed algorithm has good performance comparing with a hybrid genetic algorithm and a list scheduling, both typical from the literature. 2006 Vol. 19 (3): 399-405 [Abstract] ( 275 ) [HTML 1KB] [ PDF 487KB] ( 731 ) 406 A Reinforcement Learning Method for LQR Control Problem WEN Feng, CHEN ZongHai, ZHOU GuangMing, CHEN ChunLin Current convergence analyses of reinforcement learning method are mainly applied to discrete state problems. Analyses of continuous state reinforcement learning method are limited to simple LQR control problems. After analyzing two convergent reinforcement learning methods for LQR control problem, a new method only requiring partial model information is proposed to make up for the defects of these two methods. In this method, a recursive leastsquares TD method is used to estimate parameters of value function and a recursive leastsquares method is used to estimate the greedily improved policy. In theoretical analysis, a convergence proof is presented for the proposed policy iteration method in ideal case. Simulation result shows that this method converges an optimal control policy. 2006 Vol. 19 (3): 406-411 [Abstract] ( 353 ) [HTML 1KB] [ PDF 335KB] ( 951 ) 412 A Global Discretization Method Based on Rough Sets SHI Hong Since rough sets theory unveils the dependency of data and implements data reduction, it has attracted much attention from more and more fields. Moreover, discretization of continuous attributes plays an important role in rough sets theory and other induction learning systems. Because discretization is viewed as a process of information generalization (or abstraction) and data reduction, a global discretization algorithm is proposed based on rough sets theory. It modifies the criterion of selecting the best cut points, and introduces inconsistency checking to preserve the fidelity of the original data,which changes the MDLP method into a global one. Then the reduction of cut points is performed to lead to small size learning model while keeping the consistency level. The proposed algorithm is tested on several data sets with ID3 and ROSETTA. Experimental results show that this method performs better than MDLP and it is also superior to processing continuous data directly without discretization. 2006 Vol. 19 (3): 412-416 [Abstract] ( 335 ) [HTML 1KB] [ PDF 310KB] ( 567 ) 417 Interactive Evolutionary Computation Using an Absolute Rating DataTrained Predictor WANG ShangFei, XUE Jia, WANG XuFa Predicting IEC users’ evaluation characteristics is an effective way of reducing users’ fatigue. However, users’ relative evaluation depresses the performance of the algorithm which learns and predicts the users’ evaluation characteristics. The idea of “absolute scale” is introduced to reduce the noise and improve the performance of predicting users’ subjective evaluation characteristics in IEC, thus it accelerates EC convergence and reduces users’ fatigue. Simulation experiments of six benchmark functions are presented to prove the effectiveness of the proposed algorithm. This algorithm is also used in individual emotion fashion image retrieval system. Subjective experimental results of sign tests demonstrate that the proposed algorithm can alleviate users’ fatigue and has a good performance in individual emotional image retrieval. 2006 Vol. 19 (3): 417-421 [Abstract] ( 245 ) [HTML 1KB] [ PDF 653KB] ( 304 ) 422 A Mixture Curvature Flow Model for Image Noise Removal WANG HongYuan, SHI ChengXian, XIA DeShen Image denoise is a preliminary process in image processing with PDE based approach being a hot topic in this field in recent years. Through analyzing the performance of curve evolution equation with the theory of geometry and anisotropic diffusion, a improved curve evolution modelmixture curvature flow model for image noise removal is proposed in this paper. To improve performance, edge information is introduced and a new geometry flow is constructed in this model. The experimental results show the model has advantages over other filterings in image noise removal. 2006 Vol. 19 (3): 422-427 [Abstract] ( 217 ) [HTML 1KB] [ PDF 1543KB] ( 605 ) 428 ForemostPolicy Reinforcement Learning Based ART2 Neural Network FAN Jian, WU GengFeng A foremostpolicy reinforcement learning based ART2 neural network (FPRLART2) and its learning algorithm are proposed in this paper. To fit the requirement of real time learning, the first awarded behavior based on present states is selected in our ForemostPolicy Reinforcement Learning (FPRL) in stead of the optimal behavior in 1step QLearning. The algorithm of FPRL is given and it is integrated with ART2 neural network. The stored weights of classified pattern in ART2 is increased or decreased by reinforcement learning. The FPRLART2 is successfully used in collision avoidance of mobile robot and the simulation experiment indicates that the times of collision between robot and obstacle is effectively decreased. The FPRLART2 makes favorable result of collision avoidance. 2006 Vol. 19 (3): 428-432 [Abstract] ( 208 ) [HTML 1KB] [ PDF 591KB] ( 350 )

Supervised by China Association for Science and Technology Sponsored by Chinese Association of Automation NationalResearchCenter for Intelligent Computing System Institute of Intelligent Machines, Chinese Academy of Sciences Published by Science Press

Copyright © 2010 Editorial Office of Pattern Recognition and Artificial Intelligence Address: No.350 Shushanhu Road, Hefei, Anhui Province, P.R. China Tel: 0551-65591176 Fax:0551-65591176 Email: bjb@iim.ac.cn Supported by Beijing Magtech 　Email:support@magtech.com.cn