Service computing traditionally includes a centralized control structure. However, multi-agent theory and technology provide another way to realize service computing. In this paper a web service is viewed as an agent, called service agent, and thus service computing is regarded as a decentralized and distributed computing mechanism. Service agents can recognize and provide services for the service requests then aggregate together for the request satisfying. Based on a function ontology and automated mechanism design, the solution of request is obtained. Moreover, how to choose a solution as a 0-1 integer programming problem is formulated, and this problem is an NP-complete problem. A negotiation framework is provided to resolve this problem. In this framework, service requester and service agents come to an agreement by negotiating within finite rounds. A simulation is given to evaluate the negotiation framework.

A feature extraction approach for printed Tibetan character based on fractal moments is proposed. After analyzing the existing feature extraction approaches, the image projection approach and the directional line element approach, the fractal moments theory is applied in the proposed approach. The extracted features using the proposed method can describe the local and global properties of the character, and decrease the adverse influence caused by the change of pixel position in the image. The method can overcome high probability of misrecognition of numerous Tibetan characters, and solve the low arithmetic speed problem caused by more feature vector dimensions. Experimental results on 592 categories of printed Tibetan characters show the effectiveness of the proposed method. Moreover, the recognition rate of the proposed approach is about 2.48％ higher than that of any other existing feature extraction approach.

A gait recognition method based on silhouette areas of multi-regions is proposed. Firstly, the detected human silhouette is divided into five sub-regions in the video sequence. Then, the silhouette area features of each sub-region are extracted and the changing features in the gait sequence are computed. Therefore, the gait feature vectors can be constructed. Finally, experimental results show that the proposed approach is valid and has encouraging recognition performance.

An image segmentation method is presented based on the partial differential equation to construct a good watershed structure for gradient level image and improve watershed segmentation. Firstly, the gradient level of the original image is obtained by edge detecting. Then the edge map of the gradient image is gradually diffused and the noises are removed by using 1D-GVF (gradient vector flow) partial differential equation. Thus the processed gradient level image has a good watershed structure. Furthermore, the local minima of the processed gradient level image are detected. Then the adjacent local minima are automatically merged by morphological dilation operation and consequently the image is favorable to the segmentation of watershed algorithm. Finally, the processed gradient level image is segmented by using watershed algorithm. Experimental results show the proposed method significantly decreases the over-segmentation and provides a reliable basis for further processing.

Computing the expected statistics is the main bottleneck in learning Bayesian networks. Firstly, a parallel expectation-maximization (PL-EM) algorithm for leaning Bayesian network parameters is presented. The PL-EM algorithm parallelizes the E-step and M-step and the greatly reduces the time complexity of the parameter learning. Then PL-EM algorithm is applied to learning Bayesian networks structure, and a parallel learning algorithm is proposed for learning Bayesian networks based on an existing structural EM algorithm (SEM), called PL-SEM. PL-SEM exploits PL-EM algorithm to compute the expected statistics at the structural E_Step. Thus, it can implement the parallel computation of the expected statistics and greatly reduce the time complexity of learning Bayesian networks.

Inspired by the transmission of seeds in nature, an evolutionary algorithm, seed optimization algorithm (SOA), is proposed. The algorithm is designed by simulating the self-adaptive phenomena of plant and it can be used to resolve complex optimization problems with the evolution of plant. The global convergence analysis of SOA is made by using the Solis and Wets'research results. Finally, SOA is applied to three function optimization problems and compared with particle swarm optimization (PSO) algorithm. The experimental results show that SOA has stable and robust behaviour and it can be used as a promising alternative to existing optimization methods for engineering design.

Ensemble diversity is considered as a key factor in ensemble learning. There are many methods for constructing clustering collection or ensemble, but a few of them focus on the production of high ensemble diversity. Two methods are proposed for generating clustering ensembles with high diversity—constructing clustering ensemble by adding noise (CEAN) and improved CEAN (ICEAN). By adding artificial data, they can obtain clustering ensembles with high diversity. Compared with other commonly used methods for generating clustering ensembles, CEAN and ICEAN increase the ensemble diversity, and thus they get better clustering integration results with the same average ensemble member accuracy.

An unsupervised learning approach for content based image retrieval system is presented, which combines the low-level vision feature and the high-level semantics using the memorized SVM relevance feedback. The proposed approach fully explores the similarities among images in database by using the improved dominant set clustering to optimize the “relevance” feedback results from SVM. The experimental results show that the proposed method can be convergent to user's retrieval concept rapidly, and it has the superior precision and total relevance feedback times in image retrieval system.

A measurement of evaluating the diversity of non-dominated solutions in the objective space is introduced. It constructs alterable neighborhoods of solutions and the sizes of these neighborhoods change with the density of solution sets. The diversity relations among these neighborhoods are computed, and a metric is build. The metric can be used to compare the performance of different multi-objective optimization techniques. In particular, it can adapt to uniform test problems and non-uniform test problems. Experimental results show the proposed measurement is effective.

Since computer users are different at degrees of familiarity with keyboards and keystroke habits, each user has his particular keystroke characteristics. To one user, his keystroke characteristics are normal classes, and all the other users' are abnormal. Thus this problem can be resolved by one-class classifier. The keystroke verification based on support vector data description (SVDD) is proposed. Through experiments, SVDD is compared with BP, RBF and SOM, and the results show SVDD has better performance. It decreases impostor pass rate (IPR) from 28.9% to 0.28%. Finally, an password & keystroke characteristics identity verification system is presented.

Inspired by the idea of combining subspace learning and regularization techniques, an algorithm called locally discriminant projection of regularized least squares is proposed. To obtain projection subspace, within-class and between-class graph are constructed firstly. Then, the formula of locally discriminant projection is derived. Finally the projection subspace is worked out by regularized least squares. Compared with common algorithms, the proposed algorithm preserves the local geometrical structure of the manifold and the discriminant structure of the manifold. The experimental results on standard face database show effectiveness of the proposed algorithm.