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....

2008 Vol.21 Issue.1, Published 2008-02-01 Papers and Reports    Researches and Applications    Surveys and Reviews        Papers and Reports 1 Robust Prediction Model of Least Squares Support Vector Machine Based on Sliding Window ZHAO YongPing, SUN JianGuo In this paper, the mathematical model of weighted least squares support vector machine (WLSSVM) is introduced. Based on the algorithms of heuristic learning and sliding window, a mathematical model of robust prediction of least squares support vector machine (LSSVM) using sliding window is proposed. with the modified heuristic learning algorithm, the strategy of iterative computing matrix inverse is employed to reduce the predicted time without loss of accuracy. Finally, two examples have proved that the proposed model can eliminate the outliers, realize robust prediction and achieve good results. 2008 Vol. 21 (1): 1-5 [Abstract] ( 365 ) [HTML 1KB] [ PDF 500KB] ( 536 ) 6 An Algorithm for Mining Closed Frequent Patterns Based on Projection Sum Tree YANG ChuanYao, ZHANG ChengHong, HU YunFa In this paper, a new algorithm for mining closed frequent patterns is presented based on a projection sum frequent items tree. This algorithm projects the transaction base into a projection sum frequent items tree and stores the patterns compactly with the help of tiers. When mining, it can make full use of the existing computational result which has been done without repeat computation. It traverses the projection tree only once and does not need to generate the conditional FP trees dynamically and recursively and it avoids much timeconsuming I/O. The experiment shows that it has a high efficiency on dense datasets. 2008 Vol. 21 (1): 6-11 [Abstract] ( 278 ) [HTML 1KB] [ PDF 395KB] ( 591 ) 12 Study of Stability of Text Classification Evaluation GONG BiHong, PENG Bo Macro average precision, macro average recall and macro average F1 are usually used to evaluate classification technique. But those measures are sensitive to the datasets which means the measures are only valid for specific dataset but invalid for the others. To solve this problem, three factors are proposed to describe how datasets affect the classification result. Then a new evaluation method of categorization called newmacroF1 is presented according to the three factors. Experimental results show that the new measure remains stable on different datasets and through the performance of an algorithm on one dataset, the precision of other datasets could be estimated with the help of new measure. 2008 Vol. 21 (1): 12-17 [Abstract] ( 304 ) [HTML 1KB] [ PDF 606KB] ( 621 ) 18 Semantics and Reasoning of Hybrid Terminological Cycles in Description Logic εL with RVM JIANG YunCheng, WANG Ju, ZHOU ShengMing, TANG Yong The current research progresses and problems of terminological cycles in description logics are analyzed in this paper. Based on the research of Baader F and Brandt S, the semantics and reasoning of hybrid terminological cycles in description logic εL with RVM is further studied. The syntax and semantics of hybrid terminological cycles in description logic εL with RVM are given. Aiming at the requirement of subsumption reasoning of hybrid terminological cycles in description logic εL with RVM, TBoxcompletion is presented and the description graph is redefined. The subsumption reasoning algorithms of hybrid terminological cycles in description logic εL with RVM w.r.t. greatest fixpoint semantics and descriptive semantics are presented using TBoxcompletion and description graph. The correctness of reasoning algorithms has been proved. And it is also proved that the subsumption reasoning w.r.t. greatest fixpoint semantics and descriptive semantics can be computed in polynomial time. 2008 Vol. 21 (1): 18-27 [Abstract] ( 241 ) [HTML 1KB] [ PDF 396KB] ( 455 ) 28 Speaker Verification Based on GMM Multidimensional Likelihoods and SVM LIU MingHui, DAI BeiQian, XIE YanLu In this paper, a textindependent speaker verification system based on GMM multidimensional likelihoods and SVM is proposed, which combines the advantages of both generative model and discriminative model. In this method, the GMM multidimensional likelihoods for the test speech are regarded as new features for SVM. Experiment results of textindependent speaker verification on NIST'05 8conv4w1conv4w database show effectiveness of the proposed system. 2008 Vol. 21 (1): 28-33 [Abstract] ( 268 ) [HTML 1KB] [ PDF 405KB] ( 481 ) 34 Cluster Validity Function Based on Fuzzy Degree CHEN Duo, LI Xue, CUI DuWu, FEI Rong Construction of cluster validity function is a commonly used method to determine the optimal partition and optimal number of clusters for fuzzy partitions. Based on the basic theory of fuzzy set, the notion of cluster fuzzy set is suggested, which is subjected to the constraint conditions of fuzzy Cmeans cluster algorithm. The cluster fuzzy degree and the lattice degree of approaching for cluster fuzzy set are defined and their functions in validation process of fuzzy clustering are deeply analyzed. A new cluster validity function is presented, in which two factors, the cluster fuzzy degree and the lattice degree of approaching, are taken into account comprehensively. Furthermore, the detailed steps are given to apply the cluster validity function to the clustering validity for the fuzzy Cmeans cluster algorithm. The experimental results indicate the effectiveness and robustness of the proposed cluster validity function. 2008 Vol. 21 (1): 34-41 [Abstract] ( 266 ) [HTML 1KB] [ PDF 647KB] ( 448 ) 42 SubBand Optimization with Criterion of Maximum Weighting Entropy and Its Application in Pattern Classification BAO Ming, GUAN LuYang, LI XiaoDong, TIAN Jing Power spectral subband analysis with the criterion of maximum weighting entropy is derived as a new signal analysis method in this paper. The maximum information is obtained by optimizing the subbands allocated in frequency. Based on this method, a algorithm of feature extraction for classification, maximum weighting entropy cepstrum coefficients (MECC), is proposed and applied to ground vehicle recognition system. Experimental results show that MECC has better classification performance than the traditional methods. 2008 Vol. 21 (1): 42-48 [Abstract] ( 253 ) [HTML 1KB] [ PDF 1074KB] ( 517 ) Surveys and Reviews 49 A Review of Constructive Neural Networks WANG LunWen, ZHANG Ling Constructive neural network, a new type of neural network model, can process the largescale data. It has been widely used and deeply studied in recent years. In this paper, the principle of the neural networks is introduced, its property is analyzed and compared with that of other networks, and the research development is summarized. Furthermore, the reasons why the networks have so many advantages are analyzed, the further research is discussed and the application prospect is presented. 2008 Vol. 21 (1): 49-55 [Abstract] ( 334 ) [HTML 1KB] [ PDF 398KB] ( 637 ) Researches and Applications 56 Target Recognition Based on Fuzzy Automata WU QingE , WANG Tuo, LIU WeiLiang, HUANG YongXuan, LI JiSheng, SHU Lan A target recognition system based on fuzzy automata is presented. The system first performs image processing, then accomplishes the target recognition. The system consists of four parts: image preprocessing, feature extraction, target matching and experiment. Compared with the existing approaches, both global and local features of target image are utilized in this system, and fuzzy automata is used for target recognition. Simulation results show that the result of the target recognition based on fuzzy automata is better than that of other approaches. The recognition rate, as high as 94.59%, is achieved. 2008 Vol. 21 (1): 56-61 [Abstract] ( 371 ) [HTML 1KB] [ PDF 464KB] ( 434 ) 62 Research on Offline Handwritten Chinese Characters RecognitionBased on Biomimetic Recognition WANG JianPing, LI WeiTao, WANG JinLing, WANG XiHui, CHENG Yu In this paper, biomimetic pattern recognition is employed to construct double weighted elliptical neuron sequence which is used to extract basic stroke segments to cover the handwritten Chinese character image. The topological property of the stroke segment neurons is analyzed. An image of handwritten Chinese characters is transformed into geometric figures composed by six styles of Chinese character strokes with fault tolerance. To imitate typing methods of human Chinese characters font code, the style, the number, the position, the number of joint and the crossover of stroke neurons with redundant fault tolerant shapes are counted. Data structures of characteristic knowledge of handwritten Chinese characters are built. Handwritten Chinese characters from SCUTIRAC HCCLIB are tested and the results confirm the proposed method has the ability of cognizing handwritten Chinese characters. 2008 Vol. 21 (1): 62-71 [Abstract] ( 319 ) [HTML 1KB] [ PDF 1615KB] ( 449 ) 72 An Adaptive Polyclonal Clustering Algorithm and Its Convergence Analysis MA Li, JIAO LiCheng, BAI Lin, CHEN ChangGuo Based on a simple description of the basic principle of biology immune and clonal process, a polyclonal clustering algorithm with selfadaptive feature is put forward. The main idea of the algorithm is to put various operators in artificial immune system into clustering process and adjust clustering numbers automatically by affinity function. The recombination operator is introduced to increase the diversity of antibody group so as to broaden the search scope of the global optimization solution and avoid early mature phenomenon of the group. And the nonconsistent mutation operator is introduced to enhance the adaptability and optimize the performance of local solution seeking, meanwhile convergence of the algorithm is speeded up. The experimental result shows that reasonable clustering could be realized by the proposed algorithm. 2008 Vol. 21 (1): 72-81 [Abstract] ( 244 ) [HTML 1KB] [ PDF 797KB] ( 407 ) 82 An Unsupervised Color Image Segmentation Algorithm Based on Context Information GUO Lei, HOU YiMin, LUN XiangMin An unsupervised color image segmentation method based on image context information is proposed. According to the traditional markov random field (MRF) potential function, the method involves intensity Euclidean distance and spatial position information of pixels in the neighborhood of the image. Therefore, the traditional potential function of MRF segmentation method is improved. The segmentation is transformed into the problem of maximum a posteriori (MAP) which is solved by the iterative conditional model. And Kmeans is used to initialize the classification in the range of the specified classification numbers. The optimal class number is chosen according to the minimum message length (MML) criterion to complete an unsupervised segmentation. In the experiments, synthetic and real images are employed in segmentation procedure. Compared with other methods, the proposed algorithm is proved to be more effective. 2008 Vol. 21 (1): 82-87 [Abstract] ( 395 ) [HTML 1KB] [ PDF 1442KB] ( 885 ) 88 A Classification Algorithm for RBFNN Based on Cooperative Coevolution TIAN Jin, LI MinQiang, CHEN FuZan A new algorithm is presented to improve the classification ability of the radial basis function neural network (RBFNN). It attempts to construct RBFNN based on a cooperative coevolutionary algorithm. The Kmeans method is employed and the initial hidden nodes are divided into modules to represent the species of the coevolutionary algorithms. The good individuals in all species are found and then combined to form the whole structure of RBFNN. A matrixform mixed encoding scheme with a control vector is adopted in this algorithm. The weights between the hidden layer and the output layer are calculated by pseudoinverse algorithm. The proposed algorithm is tested on UCI datasets and the results show it outperforms the other existing methods with higher accuracy and simpler network construction. 2008 Vol. 21 (1): 88-97 [Abstract] ( 260 ) [HTML 1KB] [ PDF 675KB] ( 505 ) 98 An Algorithm for Hierarchical Policy Search Based on PSO PENG ZhiPing, LI ShaoPing In order to overcome drawbacks in hierarchical policy gradient reinforcement learning algorithm (HPGRL), such as problem of local optimum, a new algorithm for searching hierarchical policies is proposed, named Hierarchical Policy Search Based on PSO (PSOHPS). The designers create the task decomposition graph according to the hierarchical theory of MAXQ, one of the classical hierarchical reinforcement learning techniques. Then the hierarchical parameterized policies of all compound subtasks are evolved in process of direct interaction with the environment by utilizing a particle swarm to acquire the optimized action policies. Experimental results demonstrate the algorithm is valid and its performance outperforms that of HPGRL remarkably. 2008 Vol. 21 (1): 98-103 [Abstract] ( 324 ) [HTML 1KB] [ PDF 403KB] ( 567 ) 104 A Pitch Detection Algorithm Based on Linear Prediction Residual Cepstrum JIN XueCheng, WANG ZengFu An algorithm based on the linear prediction (LP) residual cepstrum for pitch detection is presented. The cepstrum of linear prediction residual of the speech signal is used to be the information for pitch determination. Voicing decisions are made based on a decision function consisting of prediction residual cepstral peak, energy and zerocrossing rate of shorttime segments of the speech signal. By this decision function the procedure of voicing decision is greatly simplified and the accuracy of voiced/unvoiced classification is improved significantly. Based on the consecution of pitch, a peak relocation method is introduced in the process of pitch determination to resolve the problems of pitch doubling and pitch halving. The results of the contrast experiment show that the proposed algorithm not only obtains a considerable improvement compared with the conventional cepstrum method, but also performs better than YIN estimator and multiscale wavelet method, which are effective admittedly. 2008 Vol. 21 (1): 104-110 [Abstract] ( 416 ) [HTML 1KB] [ PDF 737KB] ( 840 ) 111 Multiagent Cooperative Learning Based on Coordination of Boundary Samples HAN Wei Aiming at the large statespace caused by the slow convergence of Q learning, a kind of multiagent cooperative learning is proposed by the coordination of boundary samples. Each agent is specialized in its subspace, and the agents coordinate through Boolean functions in boundary states. Simulation results have proved that the proposed method performs better than the traditional global learning. 2008 Vol. 21 (1): 111-115 [Abstract] ( 307 ) [HTML 1KB] [ PDF 393KB] ( 450 ) 116 Evolutionary Algorithm Based Pattern Discovery in Graphical Databases CHANG XinGong, LI MingQiang, KOU JiSong The greedy search is often used in some existing prevalent graphical data mining systems which often ends up with suboptimal solutions. To overcome its limits, an evolutionary algorithms based system is developed to perform data mining on databases represented as graphs. New operators of mutation and crossover on graphical databases are defined, and the way of collecting instances of a certain substructure is improved. In addition, a variant of hillclimbing is integrated into the design of mutation operator to improve the capability of local search of evolutionary algorithm. Experimental results show that these measures successfully improve the searching capability of the algorithm and the qualities of solutions. 2008 Vol. 21 (1): 116-121 [Abstract] ( 266 ) [HTML 1KB] [ PDF 471KB] ( 489 ) 122 Offline Signature Verification Based on Optimal Rules of Fuzzy Modeling TIAN Wei, QIAO YiZheng, MA ZhiQiang A new offline signature verification system based on fuzzy modeling of multiple rules is proposed. In this system, both static and pseudodynamic features are extracted to make up for the loss of dynamic information and their variation is described by fuzzy sets. Then the new weight coefficients by the membership functions are devised to reflect the contribution of different fuzzy rules to verification results. In addition, the optimal selection of multiple rules by the reliable estimate of Kfold crossvalidation is presented to reduce the computational complexity of the entire fuzzy system. Databases of Chinese and English signatures are applied to the experiments and the average error rates of 9.52% and 12.67% are obtained. Thus the effectiveness of the proposed system is validated. 2008 Vol. 21 (1): 122-128 [Abstract] ( 239 ) [HTML 1KB] [ PDF 369KB] ( 452 )

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