PRICAI 2006: Trends in Artificial Intelligence

This paper explains an architecture for a BDI agent that can learn based on its own experience. The learning is conducted through explicit procedural knowledge or plans in a goal-directed manner. The learning is described by encoding

abductions

within the deliberation processes. With this model, the agent is capable of modifying its own plans on the run. We demonstrate that by abducing some complex structures of plan, the agent can also acquire complex structures of knowledge about its interaction with the environment.

This paper presents an approach that uses a Constructive Interactive Activation and Competition (CIAC) neural network to model a situated agent’s experience. It demonstrates an implemented situated agent and its learning mechanisms. Experiments add to the understanding of how the agent learns from its interactions with the environment. The agent can develop knowledge structures and their intentional descriptions (conceptual knowledge) specific to what it is confronted with – its experience. This research is presented within the design optimization domain.

This paper provides a framework based on temporal defeasible logic to reason about deliberative rule-based cognitive agents. Compared to previous works in this area our framework has the advantage that it can reason about temporal rules. We show that for rule-based cognitive agents deliberation is more than just deriving conclusions in terms of their mental components. Our paper is an extension of [5,6] in the area of cognitive agent programming.

Boolean games, introduced by [15,14], allow for expressing compactly two-players zero-sum static games with binary preferences: an agent’s strategy consists of a truth assignment of the propositional variables she controls, and a player’s preferences is expressed by a plain propositional formula. These restrictions (two-players, zero-sum, binary preferences) strongly limit the expressivity of the framework. While the first two can be easily encompassed by defining the agents’ preferences as an arbitrary

n

-uple of propositional formulas, relaxing the last one needs Boolean games to be coupled with a propositional language for compact preference representation. In this paper, we consider generalized Boolean games where players’ preferences are expressed within two of these languages: prioritized goals and propositionalized CP-nets.

In this paper, we propose a new agent-based flexible videoconference system (AVCS) by modifying videoconference manger (VCM) agent in a conventional flexible videoconference system (FVCS). The proposed AVCS can more flexibly cope with changes in working conditions during videoconferencing than the conventional FVCS. It is because an automatic parameter tuning algorithm is imbedded to VCM dynamically adapt QoS (Quality of Service) parameters in the sense that the current working condition can meet with the desired working condition of the user, which can change in time during videoconferencing. In the experimental section, we design a new structure of the VCM with the automatic parameter tuning module, imbed to the prototype of FVCS and implement the new AVCS. Also, it is shown that the proposed AVCS outperforms the existing FVCS in the experiment.

Whereas conventional GRID service is static, wireless GRID supports mobility, and it should maintain geographic position to support efficient resource sharing and routing. When the devices are highly mobile, there will be much traffic to exchange the geographic position information of each mobile node, and this makes adverse effect on efficient battery usage and network congestion. To minimize the network traffic between mobile users, we can use dead reckoning (DR) algorithm for each mobile nodes, where each node uses the algorithm to estimates its own movement (also other node’s movement), and when the estimation error is over threshold, the node sends the UPDATE (including position, velocity, etc) packet to other devices. As the estimation accuracy is increased, each node can minimize the number of UPDATE packet transmission. To improve the prediction accuracy of DR algorithm, we propose Kalman filter based DR approach, and we also propose the adaptive Kalman gain control to minimize the number of UPDATE packet to distant device. To experiment our scheme, we implement a popular network game (BZFlag) with our scheme added on each mobile node, and the results show that we can achieve better prediction accuracy and reduction of network traffic by 12 percents.

In this paper, we propose that, in order to improve customer satisfaction, we need to incorporate communication modes (e.g., speech act) in the current standards of web services specifications. We show that with the communication modes, we can estimate various affects on service consumers during their interactions with web services. With this information, a web-service management system can automatically prevent and compensate potential negative affects, and even take advantage of positive affects.

Collaborative Filtering (CF) techniques are important in the e-business era as vital components of many recommender systems, for they facilitate the generation of high-quality recommendations by leveraging the similar preferences of community users. However, there is still a major problem preventing CF algorithms from achieving better effectiveness, the sparsity of training data. Lots of ratings in the training matrix are not collected. Few current CF methods try to do data smoothing before predicting the ratings of an active user. In this work, we have validated the effectiveness of data smoothing for memory-based and hybrid collaborative filtering algorithms. Our experiments show that all these algorithms achieve a higher accuracy after proper smoothing. The average mean absolute error improvements of the three CF algorithms, Item Based, k Nearest Neighbor and Personality Diagnosis, are 6.32%, 8.85% and 38.0% respectively. Moreover, we have compared different smoothing methods to show which works best for each of the algorithms.

Web spider is a widely used approach to obtain information for search engines. As the size of the Web grows, it becomes a natural choice to parallelize the spider’s crawling process. However, parallel execution often causes redundant web pages to occupy vast storing space. How to solve this problem becomes a significant issue for the design of next generation web spiders. In this paper, we employ the method from multi-agent coordination to design a parallel spider model and implement it on the multi-agent platform MAGE. Through the control of central facilitator agent, spiders can coordinate each other to avoid redundant pages in the web page search process. Experiment results demonstrate that it is very effective to improve the collection efficiency and can eliminate redundant pages with a tiny efficiency cost.

This paper presents an economical inspired intelligent approach for modeling learners in learning systems. Decision making in complex systems like e-learning systems requires processing of large amounts of heterogeneous data and information from dispread sources. Moreover, most of the decision parameters are incomplete and uncertain. Lacking of a complete model of learner is the prominent problem of current learning systems. In this paper, a market based method for describing Learner’s preferences to the learning system is provided. The proposed approach strives for applying a Dempster-Shafer decision making over a society of self motivated agents. It tries to present a final learner agent with a high degree of similarity to the user for the purpose that it can act as a model of learner through the system. An implicit learning is also implemented by the idea of Stocks in real markets which can improve decision making efficiently.

The personalized recommendation system is required to save efforts in searching the items in ubiquitous commerce, it is very important for a recommendation system to predict accurately by analyzing user’s preferences. A recommendation system utilizes in general an information filtering technique called collaborative filtering, which is based on the ratings matrix of other users who have similar preference. This paper proposes the user preference through Bayesian categorization for recommendation to overcome the sparsity problem and the first-rater problem of collaborative filtering. In addition, to determine the similarity between the users belonging to a particular class and new users, we assign different statistical values to the items that the users evaluated using Naive Bayesian classifier. We evaluated the proposed method on the EachMovie datasets of user ratings and it was found to significantly outperform the previously proposed method.

Stochastic local search techniques have been successful in solving propositional satisfiability (SAT) problems encoded in conjunctive normal form (CNF). Recently complete solvers have shown that there are advantages to tackling propositional satisfiability problems in a more expressive natural representation, since the conversion to CNF can lose problem structure and introduce significantly more variables to encode the problem. In this work we develop a non-CNF SAT solver based on stochastic local search techniques. Crucially the system must be able to represent how true a proposition is and how false it is, as opposed to the usual stochastic methods which represent simply truth or degree of falsity (penalty). Our preliminary experiments show that on certain benchmarks the non-CNF local search solver can outperform highly optimized CNF local search solvers as well as existing CNF and non-CNF complete solvers.

Most techniques for probabilistic reasoning focus on reasoning about conditional probability constraints. However, human experts are accustomed to representing uncertain knowledge in the form of expectation rather than probability distribution directly in many cases. It is necessary to provide a logic for encoding hybrid probabilistic knowledge bases that contain expectation knowledge as well as the purely probabilistic knowledge in the form of conditional probability. This paper constructs a nonmonotonic logic for reasoning about hybrid probabilistic knowledge bases. We extend the propositional logic for reasoning about expectation to encoding hybrid probabilistic knowledge by introducing the conditional expectation constraint formula. Then we provide an approach to nonmonotonic reasoning about hybrid probabilistic knowledge bases. Finally,we compare this logic with related works.

Continuous time Bayesian network is a new kind of dynamic graphical models developed in recent year, which describe structured stochastic processes with finitely many states that evolve over continuous time. The parameters for each variable in the model represent a finite state continuous time Markov process, whose transition model is a function of its parents. This paper presents an algorithm for updating parameters from an existing CTBN model with a set of data samples. It is a unified framework for online parameter estimation and batch parameter updating where a pre-accumulated set of samples is used. We analyze different conditions of the algorithm, and show its performance in experiments.

In [11,13] we showed how to combine propositional BDI logics using Gabbay’s

fibring

methodology. In this paper we extend the above mentioned works by providing a tableau-based decision procedure for the combined/fibred logics. We show how to uniformly construct a tableau calculus for the combined logic using Governatori’s labelled tableau system

KEM

.

This paper proposes the canonical representation of multiplication operation on trapezoidal fuzzy numbers using the L

− − 1

-R

− − 1

Inverse Function Arithmetic Representation method. Finally, the canonical representation proposed in this paper is applied to solve a fuzzy multiple criteria decision making problem of selection of plant location.

We propose to use the generalized visibility graph (Vgraph) to represent search space for finding a natural-looking path in computer games. The generalized Vgraph is the extension of the visibility graph to the generalized polygonal world which is produced as a result of dilating polygonal obstacles. That is, the generalized visibility graph is constructed on the expanded boundaries of obstacles so that the path keeps an amount of clearance from obstacles. We also introduce an algorithm that can efficiently incorporate a start and a goal location to the map represented in a generalized Vgraph for quick path-finding. The A* algorithm with Euclidean distance is used for quick path-finding. The proposed approach is compared to other major search space representations analytically and empirically. The results show that the map can be generated efficiently by using the generalized Vgraph and the paths found look natural.

The extended goals in non-deterministic domains are often expressed in temporal logic, particularly in

CTL

and

E

A

G

LE

. No work has given a formal comparison between

E

A

G

LE

and

CTL

on semantics, though it is said that the capability of representing the “intentional” aspects of goals and the possibility of dealing with failure are the main new features of

E

A

G

LE

w.r.t.

CTL

.

According to the formal semantics for

E

A

G

LE

and

CTL

, we prove that all the

E

A

G

LE

formulas in which only

LV

1

operators (i.e. the operators representing the “intentional” aspects of goals) appear and some

E

A

G

LE

formulas including

LV

2

operators (i.e. the operators dealing with failure and qualitative preferences) can be replaced by some

CTL

formulas without any change on semantics. Finally, we also find some basic and important goals in non-deterministic domains that exceed the expressive ability of

E

A

G

LE

.

This paper proposes a method for collision avoidance of a mobile manipulator. The method deals with the problem of driving a mobile manipulator from a starting configuration to a goal configuration avoiding obstacles. It modifies planned trajectory at every sampling time using elastic force and potential field force. It puts way poses throughout a planned trajectory, and the trajectory modification is accomplished by adjusting the way poses. The way poses are adjusted subject to the elastic force and the potential field force. The procedure repeats until the robot reaches its goal configuration through the way poses. This method results in smooth and adaptive trajectory in an environment with moving as well as stationary obstacles.

Developing software agents has been complicated by the problem of how knowledge should be represented and used. Many researchers have identified that agents need not require the use of complex representations, but in many cases suffice to use “the world” as their representation. However, the problem of introspection, both by the agents themselves and by (human) domain experts, requires a knowledge representation with a higher level of abstraction that is more ‘understandable’. Learning and adaptation in agents has traditionally required knowledge to be represented at an arbitrary, low-level of abstraction. We seek to create an agent that has the capability of learning as well as utilising knowledge represented at a higher level of abstraction.

We firstly explore a reactive learner (

Falcon

) and reactive plan execution engine based on BDI (JACK) through experiments and analysis. We then describe an architecture we have developed that

combines

the BDI framework to the low-level reinforcement learner and present promising results from experiments using our minefield navigation domain.

Knowledge-based model construction approach has been applied to many problems. Previous research doesn’t provide an approach to construct models to deal with time-critical dynamic decision problems. This paper presents a Knowledge-based Time-critical Dynamic Decision Modeling system (KTDDM) for time-critical dynamic decision-making. The system architecture and functional modules are described. A new knowledge representation framework is provided to support the whole model construction process. This paper applies the KTDDM system prototype to construct a decision model for the time-critical dynamic medical problem on cardiac arrest to demonstrate the effectiveness of this approach.

Protein kinases, a family of enzymes, have been viewed as an important signaling intermediary by living organisms for regulating critical biological processes such as memory, hormone response and cell growth. The unbalanced kinases are known to cause cancer and other diseases. With the increasing efforts to collect, store and disseminate information about the entire kinase family, it not only leads to valuable data set to understand cell regulation but also poses a big challenge to extract valuable knowledge about metabolic pathway from the data. Data mining techniques that have been widely used to find frequent patterns in large datasets can be extended and adapted to kinase data as well. This paper proposes a framework for mining frequent itemsets from the collected kinase dataset. An experiment using AMPK regulation data demonstrates that our approaches are useful and efficient in analyzing kinase regulation data.

This paper presents an automatic Bayesian network construction algorithm where association analysis is employed to guide the construction of network structure. The proposed method is studied in context of data imputation together with a previously proposed technique for automatic Bayesian network construction, Backpropagation neural networks, and two traditional data imputation techniques. The results show that the proposed method performs better or at least as well as does the best of other methods in 84.62% of the cases.

One of the main issues in the rule/pattern mining is of measuring the interestingness of a pattern. The interestingness has been evaluated previously in literature using several approaches for association as well as for sequential mining. These approaches generally view a sequence as another form of association for computations and understanding. But, by doing so, a sequence might not be fully understood for its statistical significance such as dependence and applicability. This paper proposes a new framework to study sequences’ interestingness. It suggests two kinds of Markov processes, namely Bayesian networks, to represent the sequential patterns. The patterns are studied for statistical dependencies in order to rank the sequential patterns interestingness. This procedure is very shown when the domain knowledge is not easily accessible.

Existing approaches for multi-dimensional frequent patterns mining rely on the construction of data cube. Since the space of a data cube grows explosively as dimensionality or cardinality grows, it is too costly to materialize a full data cube, esp. when dimensionality or cardinality is large. In this paper, an efficient method is proposed to mine multi-dimensional frequent patterns without data cube construction. The main contributions include: (1) formally proposing the concept of multi-dimensional frequent pattern and its pruning strategy based on Extended Apriori Property, (2) proposing a novel structure called Multi-dimensional Index Tree (MDIT) and a MDIT-based multi-dimensional frequent patterns mining method (MDIT-Mining), and (3) conducting extensive experiments which show that the space consuming of MDIT is more than

4

orders of multitudes smaller than that of data cube along with the increasing of dimensionality or cardinality at most cases.

There still exist two key problems required to be solved in the classification rule extraction, i.e. how to select attributes and discretize continuous attributes effectively. The lack of efficient heuristic information is the fundamental reason that affects the performance of currently used approaches. In this paper, a new measure for determining the importance level of the attributes based on the trained SVM is proposed, which is suitable for both continuous attributes and discrete attributes. Based on this new measure, a new approach for rule extraction from trained SVM and classification problems with continuous attributes is proposed. The performance of the new approach is demonstrated by several computing cases. The experimental results prove that the approach proposed can improve the validity of the extracted rules remarkably compared with other rule extracting approaches, especially for the complicated classification problems.

Feature selection for the individuals of bagging is studied in this paper. Ensemble learning like bagging can effectively improve the performance of single learning machines, and so can feature selection, but few has studied whether feature selection could improve bagging of single learning machines. Therefore, two typical feature selection approaches namely the embedded feature selection model with the prediction risk criteria and the filter model with the mutual information criteria are used for the bagging of support vector machines respectively. Experiments performed on the UCI data sets show the effectiveness of feature selection for the bagging of support vector machines.

A novel method for recognition two kinds of lung sounds is presented. The proposed scheme is based on the analysis of a wavelet packet decomposition(WPD). Normal and abnormal lung sounds data were sampled from various subjects. Each signal is segmented to inspiration and expiration. From their high dimension WPD coefficients, we build the compact and meaningful energy feature vectors, then use them as the input vectors of the artificial neural network(ANN) to classify the lung sound types. Extensive experimental results show that this feature extraction method has convincing recognition efficiency although not yet good enough for clinical use.

Quality and reliability of wireless communication is an actual issue for design of modern high-efficiency information systems in the wide area of human activities. In the paper, the problem of wireless communication reliability and methods of its evaluation are studied. The quality of communication at actual spot is estimated with the method proposed by the authors. It is based on the usage of a prediction mathematical model presenting the time series for receiving signal level data. Different model classes are considered for the data description including neural network models. Special model training procedure based on the Aggregative Learning Method (ALM) is applied along with expert approach for the data classification. The validity and the efficiency of the proposed approach have been tested through its application for different cases including “open-air” and “in-building” environments. Cellular phone communication network of DoCoMo Inc. is used as a test bed for the proposed method.Classification abilities of the method are shown reliable for estimation of the communication quality. Characterized with high computational efficiency and simple decision making procedure, the derived method can be useful for design of simple and reliable real-time systems for communication quality monitoring.

A prediction scheme for the MPEG video traffic in ATM networks using a BiLinear Recurrent Neural Network (BLRNN) is proposed in this paper. Since the BLRNN is based on the bilinear polynomial, it has been successfully used in modeling highly nonlinear systems with time-series characteristics, and the BLRNN can be a natural choice in predicting the MPEG video traffic with a bursty nature in the ATM networks. The proposed BLRNN-based predictor is applied to MPEG-1 and MPEG-4 video traffic data. The performance of the proposed BLRNN-based predictor is evaluated and compared with the conventional MultiLayer Perceptron Type Neural Network (MLPNN)-based predictor. When compared with the MLPNN-based predictor, the proposed BLRNN-based predictor shows 27%-51% improvement in terms of the Relative Mean Square Error (RMSE) criterion.

The objective of this paper is to develop a dynamic neural network scheme for fault detection and isolation (FDI) in the reaction wheels of a satellite. The goal is to decide whether a bus voltage fault, a current loss fault or a temperature fault has occurred in one of the three reaction wheels and further to localize which wheel is faulty. In order to achieve these objectives, three dynamic neural networks are introduced to model the dynamics of the wheels on all three axes independently. Due to the dynamic property of the wheel, the architecture utilized is the Elman recurrent network with backpropagation learning algorithm. The effectiveness of this neural network-based FDI scheme is investigated and a comparative study is conducted with the performance of a generalized observer-based scheme. The simulation results have demonstrated the advantages of the proposed neural network-based method.

We retrospect Chen’s chaotic neural network and then propose a new chaotic neural network model whose activation function is composed of Gauss and Sigmoid function. And the time evolution figures of the largest Lyapunov exponents of chaotic single neural units are plotted. Based on the new model, the model with different parameters is applied to combinational optimization problems. 10-city traveling salesman problem (TSP) is given to make a comparison between Chen’s and the new model with different parameters. Finally on the simulation results we conclude that the novel chaotic neural network model we proposed is more effective.

In this paper, a short-term load forecasting model using wavelet-based neural network architecture termed a Multiscale BiLinear Recurrent Neural Network (M-BLRNN) is proposed. The M-BLRNN is a combination of several BiLinear Recurrent Neural Network (BLRNN) models. Each BLRNN predicts a signal at a certain resolution level obtained by the wavelet transform. The experiments and results on the load data from the North-American Electric Utility (NAEU) show that the M-BLRNN outperforms both a traditional MultiLayer Perceptron Type Neural Network (MLPNN) and the BLRNN in terms of the Mean Absolute Percentage Error (MAPE).

Recurrent Neural Networks (RNNs) are one in which self-loops and backward weight connections between neurons are allowed. As a result of these network characteristics, recurrent networks can address temporal behaviors which not possible in feedforward neural networks, such as their behavior in the limit reaches a steady state (fixed point), an oscillation (limit cycle), and an a periodic instability (choas). Since RNNs have been increasingly applied to many dynamic system applications, there have been extensive efforts to develop a variety of architectures and training algorithms concerning on the enhancement of dynamic system characteristics. This work focuses on comparison of the selected and proposed training algorithms for RNNs. To evaluate the performance of the algorithms in the daily stock price forecasting in terms of efficiency index and computational time. A simple analysis on the complexity of RNNs was also carried out. It is noted that when comparing the speed of the algorithm, two components to be taken into account : the computation complexity and the space complexity.

To improve equipment throughput and device yield, a malfunction in plasma equipment should be accurately diagnosed. An ex-situ diagnosis model was constructed by applying neural network to scanning electron microscope (SEM) image of plasma-etched patterns. The experimental data were collected from a plasma etching of tungsten thin films. Faults in plasma were generated by simulating a variation in process parameters. Feature vectors were obtained by applying direct and wavelet techniques to SEM images. The wavelet techniques generated three feature vectors composed of detailed components. The diagnosis models constructed were evaluated in terms of the recognition and diagnosis accuracies. The direct technique yielded much smaller recognition accuracy with respect to the wavelet technique. The improvement was about 82%. For the diagnosis accuracy, the improvement was about 30%. All these results demonstrate that the direct method is a more effective feature extraction in constructing a SEM-based neural network diagnosis model.

The task of QoS routing is to find a route in the network, which has sufficient resources to satisfy the constraints on delay, delay-jitter, cost, etc. The delay-cost-constrained routing problem is NP-complete. In this paper we present a new method of multi-constrained routing based on Hopfield neural network (HNN) that solves the optimum routing problem for supporting QoS in Ad hoc networks, which is called MCADHP. The idea of MCADHP is to reduce the NP-complete problem to a simpler one, which can be solved in polynomial time in order that the neural network could be robust to the changed of the network topology. Under this assumption the general principles involved in the design of the proposed neural network and the method regarding the relationships of different coefficients in the energy function are discussed. The performance of the proposed neural model is studied by both theoretical analysis and computer simulations. Simulation results indicate that our proposed scheme is very effective and outperforms previous schemes.

Based on the feature map principle, Sparse Kernel Ridge Regression (SKRR) model is proposed. SKRR obtains the sparseness by backward deletion feature selection procedure that recursively removes the feature with the smallest leave-one-out score until the stop criterion is satisfied. Besides good generalization performance, the most compelling property of SKRR is rather sparse, and moreover, the kernel function needs not to be positive definite. Experiments on synthetic and benchmark data sets validate the feasibility and validity of SKRR.

Shevade et al.[1] are successful in extending some improved ideas to Smola and Scholkopf’s SMO algorithm[2] for solving regression problems, simply named SMOreg. In this paper, we use SMOreg in exactly the same way as linear regression(LR) is used in locally weighted linear regression[5](LWLR): a local SMOreg is fit to a subset of the training instances that is in the neighborhood of the test instance whose target function value is to be predicted. The training instances in this neighborhood are weighted, with less weight being assigned to instances that are further from the test instance. A regression prediction is then obtained from SMOreg taking the attribute values of the test instance as input. We called our improved algorithm locally weighted SMOreg, simply LWSMOreg. We conduct extensive empirical comparison for the related algorithms in two groups in terms of relative mean absolute error, using the whole 36 regression data sets obtained from various sources and recommended by Weka[3]. In the first group, we compare SMOreg[1] with NB[4](naive Bayes), KNNDW[5](k-nearest-neighbor with distance weighting), and LR. In the second group, we compare LWSMOreg with SMOreg, LR, and LWLR. Our experimental results show that SMOreg performs well in regression and LWSMOreg significantly outperforms all the other algorithms used to compare.

In recent years, palmprint identification has been developed for security purpose. In this paper, a novel scheme of palmprint identification is proposed. We apply 2-dimensional 2_band (Discrete Wavelet Transform) and 3_band wavelet decomposition to get the low subband images, and then use them as identification feature vectors. We choose support vector machines as classifier. The experimental results demonstrate that it is a simple and accurate identification strategy and the correct recognition rate is high up to 100%.

This paper proposes a novel classifier combination system that can be used by classification systems under dynamically varying environments. The proposed method adopts the concept of context-awareness and the similarity between classes, and the system working environments are learned (clustered) and identified as environmental contexts. The proposed method fitness correlation table is used to explore the most effective classifier combination for each identified context. We use t-test for classifier selection and fusion decision and proposed context modeling and t-test. The group of selected classifiers is combined based on t-test decision model for reliable fusion. The knowledge of individual context and its associated chromosomes representing the optimal classifier combination is stored in the context knowledge base. Once the context knowledge is accumulated the system can react to dynamic environment in real time.

In this paper, we extend the original non-negative matrix factorization (NMF) to kernel NMF (KNMF). The advantages of KNMF over NMF are: 1) it could extract more useful features hidden in the original data through some kernel-induced nonlinear mappings; 2) it can deal with data where only relationships (similarities or dissimilarities) between objects are known; 3) it can process data with negative values by using some specific kernel functions (e.g. Gaussian). Thus, KNMF is more general than NMF. To further improve the performance of KNMF, we also propose the SpKNMF, which performs KNMF on sub-patterns of the original data. The effectiveness of the proposed algorithms is validated by extensive experiments on UCI datasets and the FERET face database.

This paper describes an approach to the use of citation links to improve the scientific paper classification performance. In this approach, we develop two refinement functions, a linear label refinement (LLR) and a probabilistic label refinement (PLR), to model the citation link structures of the scientific papers for refining the class labels of the documents obtained by the content-based Naive Bayes classification method. The approach with the two new refinement models is examined and compared with the content-based Naive Bayes method on a standard paper classification data set with increasing training set sizes. The results suggest that both refinement models can significantly improve the system performance over the content-based method for all the training set sizes and that PLR is better than LLR when the training examples are sufficient.

Feature subset selection and data reduction is a fundamental and most explored area in machine learning and data mining. Rough set theory has been witnessed great success in attribute reduction. A series of reduction algorithms were constructed for all kinds of applications based on rough set models. There is usually more than one reduct for some real world data sets. It is not very clear which one or which subset of the reducts should be selected for learning. Neither experimental comparison nor theoretic analysis was reported so far. In this paper, we will review the proposed attribute reduction algorithms and reduction selection strategies. Then a series of numeric experiments are presented. The results show that, statistically speaking, the classification systems trained with the reduct with the least features get the best generalization power in terms of single classifiers. Furthermore, Good performance is observed from combining the classifiers constructed with multiple reducts compared with Bagging and random subspace ensembles.

Applying one-class classification to the minorities in an imbalance data has been shown to have the potential to achieve better performance than conventional learning schemes. Parameter optimization is a significant issue when the one-class classifier is sensitive to the parameters. For one-class learning scheme with the kernel function as one-class SVM and SVDD, besides the parameters involved in the kernel, the rejection rate is another one-class specific parameter. In this paper, we proposed an improved framework in which the minority target class is used first for learning in the classification stage; then both minority and majority class are employed for estimating the generalization performance. This performance is set as the optimization criteria. Experiments on UCI and Reuters text data show that both of the parameter optimized one-class classifiers outperform the other standard one-class learning schemes.

This paper proposes a clustering-based nonlinear dimensionality reduction approach. It utilizes the clustering approaches to form the clustering structure by which the distance between any two data points are rescaled to make data points from different clusters separated more easily. This rescaled distance matrix is then provided to improve the nonlinear dimensionality reduction approaches such as Isomap to achieve the better performance. Furthermore, the proposed approach also decreases the time complexity on the large data sets, as it provides good neighborhood structure that can speed up the subsequent dimensionality reducing process. Unlike the supervised approaches, this approach does not take the labelled data set as prerequisite, so that it is unsupervised. This makes it applicable to the broader domains. The conducted experiments by classification on benchmark data sets have validated the proposed approach.

Kernel PCA, like other kernel-based techniques, is suffered from memory requirement and computational problems as well as from a tedious training procedure. This work shows that the objective function of Kernel PCA,

i.e.

the reconstruction error can be upper bounded by the distortion of K-means algorithm in the feature space. From this relation, we propose a simplification of Kernel PCA’s training procedure by Kernel K-means algorithm. The application of preimage reconstruction algorithm allows further simplification and leads to a more computational economic solution.

Most traditional relevance feedback systems simply choose the top ranked Web pages as the source of providing the weights of candidate query expansion terms. However, the contents of such top-ranked Web pages is often composed of heterogeneous sub-topics which can be and should be recognized and distinguished. However, current approaches treat retrieved Web pages as one unit and often fail to extract good quality candidate query expansion terms.

In this paper, our basic idea is that the Web pages properly clustered into a sub-topic cluster can be used as a better source than whole given Web pages, to provide more topically coherent relevance feedback for that specific sub-topic. Thus, we propose

Clustering-Based Relevance Feedback for Web Pages

, which utilizes three methods to cluster retrieved Web pages into several subtopic-clusters. These three methods cooperate to construct good quality clusters by respectively supporting Web page Segmentation, Term Selection,

k

Seed Centroid Selection. Here, the automatically selected terms indicate the relevance feedback to construct all sub-topic clusters and assign the given Web pages to proper clusters. Each subset of the selected terms, which occurs in the Web pages assigned into a sub-topic cluster, indicates the relevance feedback to expand a query over that sub-topic cluster. Our experimental results showed that the clustering performances based on two traditional term-weighting methods (i.e., an unsupervised method and a supervised method) can be significantly improved with our methods.

The task of finding semantically related words from a text corpus has applications in – to name a few – lexicon induction, word sense disambiguation and information retrieval. The text data in real world, say from the World Wide Web, need not be grammatical. Hence methods relying on parsing or part-of-speech tagging will not perform well in these applications. Further even if the text is grammatically correct, for large corpora, these methods may not scale well. The task of building semantically related sets of words from a corpus of documents and allied problems have been studied extensively in the literature. Most of these techniques rely on the usage of part-of-speech or parse information. In this paper, we explore a less expensive method for finding semantically related words from a corpus without parsing or part-of-speech tagging to address the above problems. This work focuses on building sets of semantically related words from a corpus of documents using traditional data clustering techniques. We examine some key results and possible applications of this work.

“

Listening to several things at once

” is a people’s dream and one goal of AI and robot audition, because people can listen to at most two things at once according to psychophysical observations. Current noise reduction techniques cannot help to achieve this goal because they assume quasi-stationary noises, not interfering speech signals. Since robots are used in various environments, robot audition systems require minimum

a priori

information about their acoustic environments and speakers. We evaluate a missing feature theory approach that interfaces between sound source separation (SSS) and automatic speech recognition. The essential part is the estimate of reliability of each feature of separated sounds. We tested two kinds of robot audition systems that use SSS: independent component analysis (ICA) with two microphones, and geometric source separation (GSS) with eight microphones. For each SSS, automatic missing feature mask generation is developed. The recognition accuracy of two simultaneous speech improved to an average of 67.8 and 88.0% for ICA and GSS, respectively.

As a special form of unknown words, Chinese abbreviations represent significant problems for Chinese text processing. The goal of this study is to automatically find the definition for a Chinese abbreviation in the context where both the abbreviation and its definition occur, enforcing the constraint of one sense per discourse for an abbreviation. First, the candidate abbreviation-definition pairs are collected, and then a SVM approach using context information is employed to classify candidate abbreviation-definition pairs so that the pairs can be identified. The performance of the approach is evaluated on a manually annotated test corpus, and is also compared with two other machine learning approaches: Maximum Entropy and Decision Tree. Experimental results show that our approach reaches a good performance.

This paper proposes a method for Korean clause boundary recognition. Clause boundary identification can be regarded as a three-class classification task, and it can be converted into a two-phase binary classification task. Then it is natural to apply SVMs to clause boundary recognition, since SVMs are basically binary classifiers. Specifically we first recognize the ending points of clauses, and then identify the starting points by considering the typological characteristics of Korean. In addition, since there is not a standard Korean corpus containing clause boundary information, we prepare a Korean clause identification dataset. In the evaluation, support vector machines yield the improvement of performance over memory-based learning or decision trees.

Feature-Line Method deems that a line between two points in the same class of space represents the space feature better than a single point. However, it brings faults in the classification results in terms of distance only. Here coefficient was put forward to eliminate the influence of the off-group point to classification, which was also combined with the central distance of class, then formed the improved algorithm, which is used in two different capacity document repositories. The results of experiment show that the improved algorithm support large document repositories very well, and it can be used in large-scale text classification and text retrieval.

Sports video has been widely studied due to its tremendous commercial potentials. Despite encouraging results from various specific sports games, it is almost impossible to extend a summarization system for a new sports game due to the lack of sports video modeling. In this paper, we automatically generate multi-level summarizations for sports video based on Basic Semantic Unit (BSU), which is specially presented for sports video and can be extended to a new sports game. The results of the preliminary experiments indicate that our work provides a generic, novel and effective solution for automatic multi-level sports video summarizations.

We present a novel Web Pages Summarizer

ContextSummarizer

that groups the given Web pages into ‘sense-clusters’ respecting a user’s topical interests. ContextSummarizer constructs then an extractive summary for each sense-cluster. A user’s topical interest is described by the user who selects and refines some of the word senses disambiguated within the content contexts of the given Web pages. The semantic similarity measures between the contents of Web pages/segments/sentences and the user-selected word senses were used to choose the most topically relevant sentences as the extractive summaries referring to a user’s topical interest. ContextSummarizer addresses the semantic-alignment problem between the content of a Web page, the user’s topical interest, and the extractive summary of the Web page. Our case studies and experimental results showed that our query-topic focused extractive summaries returns more topically relevant sentences for an extractive summary than those produced by existing summarization systems.

Statistical model is an effective method for character motion modeling. In this paper, a variable duration motion texture is proposed to represent complex human motion that is statistically similar to the original captured motion data. The motion texture is defined as a three-level structure with moton abstracts, motons and their distribution. The motion texture is modeled by a Semi-SLDS (Semi- Switching Linear Dynamic System), which provides an intuitive framework for describing the continuous but nonlinear dynamics of human motion. To explicitly incorporate duration modeling capability, the Semi-SLDS is adopted to improve SLDS by replacing the Markov switching layer with semi-Markov model. In addition, the proposed approach is proved flexible and effective by several motion applications, namely motion synthesis, motion recognition and motion compression.

Motion blending allows the generation of new motions by interpolation or transition between motion capture sequences, which is widely accepted as a standard technique in computer animation. But traditional blending approaches let the user choose manually the transition time and duration. This paper presents a new motion blending method for smoothly blending between two motion capture clips and automatically selecting the transition time and duration.To evaluate the effectiveness of the improved method, we have done extensive experiments. The experiment results show that the novel motion blending method is effective in smoothly blending between two motion sequences.

Old films usually have typical damages from dirt, scratch, and scribbling. These damages make an image degradation of vertical line scratches or blotches in frames. This paper proposes an efficient line scratch detection technique and efficient inpainting method based on MSE (mean square error) to fill the identified line scratch areas. Previous line scratch detection algorithms can only detect full column line scratches; however, we found that partial line scratch should also be identified for better film restoration. We identify line scratches using block-by-block inspection; thereby we can detect partial line scratches. After identifying the line scratches, we use a modified inpainting scheme, which uses MSE measure to compute gradient vector of the inpainting regions. In our experiment with old Korean films, we show that our scheme gives better video quality with much reduced computational complexity.

Recently, a great number of people can share the same digital contents, because it is possible to copy and to transmit of the digital contents easy and fast. These properties of the digital contents are causes of that reduce the will to creation of makers and that hamper industrial development. Therefore recent studies focus on the protection of digital contents. However it is not efficient that traditional encryption algorithms apply to the digital image/video contents, because of the long encryption time. To solve this problem, recent studies use the partial encryption algorithm that encrypts some parts of the image or the video frame. However there are still problems which features do not have the semantic information, because previous studies extract the features for reducing the encryption time. In this paper, we proposed the partial encryption method using the face region as the feature because the face has the semantic information and is the most important part in the digital content, especially the video contents. As shown by experimental results, the proposed method can reduce the encryption time and can improve the protection strength using the traditional encryption algorithms for the digital contents.

In this paper, we propose a novel relevance feedback approach using adaptive clustering based on region representation. Performance of content based image retrieval system is usually very low because of the semantic gap between the low level feature representation and the user’s high level concept in a query image. Semantically relevant images may exhibit very different visual characteristics, and may be scattered in several clusters. Our main goal is finding semantically related clusters to reduce this semantic gap. Our method consists of region based clustering process and cluster-merging process. All segmented regions of relevant images are grouped into semantically related clusters, and clusters are merged by estimating the number of the clusters. We form representatives of clusters as the optimal query. A region based image similarity measure is used to calculate the distance between the multipoint optimal query and an image in the database. Experiments have demonstrated that the proposed approach is effective in improving the performance of image similarity retrieval system.

In this study, we explore the roles of learning and evolution in a non-cooperative autonomous system through a spatial IPD (Iterated Prisoner’s Dilemma) game. First, we propose a new agent model playing the IPD game; the game has a gene of the coded parameters of reinforcement learning. The agents evolve and learn during the course of the game. Second, we report an empirical study. In our simulation, we observe that the spatial structure affects learning and evolution. Learning is not effective for achieving mutual cooperation except under certain special conditions. The learning process depends on the spatial structure.

Based on the clonal selection theory, a novel artificial immune systems algorithm, immune clonal selection evolutionary strategy for constrained optimization (ICSCES), is put forward. The new algorithm uses the stochastic ranking constraint-handling technique, realizs local search using clonal proliferation and clonal selection, and global search using clonal deletion. The experimental results on ten benchmark problems show, compared with the (

μ

,

λ

) evolutionary strategies adopting stochastic ranking technique and dynamic penalty function method, ICSCES has the ability of significantly improving the search performance both in convergence speed and precision.

Creativity has long been recognized as vital to organizational success. IT (Information Technology) may play a supporting role to help organizations and groups in support of creativity. However idea generation inside the personal creative process is still a black box. This motivates us to propose an intelligent system for supporting personal creativity. This study proposes a human-machine interactive mechanism based on IGA (interactive genetic algorithm) for evolution process to help individuals’ creativity. Chance discovery has been applied as the strategy in evolution process. It provides more efficient evolution. The intelligent system could bring individuals stimulus to improve personal creativity.

This paper describes a creativity support system for assisting human creative activities through patents. It integrates problem definition, reminding, creating, evaluating, and visualizing modules to automatically generate creative ideas. The system utilizes natural language processing techniques to map the patents to vectors by which related patent objects are automatically reminded around the given problem. These reminded objects are divided into text fragments, from which new idea can be generated through an interactive genetic algorithm. The system differs from existing creativity support systems in that it automatically reminds patents and generates creative ideas. This has been empirically validated by the conducted experiments.

There has emerged a surge of research activity on multiobjective optimization using evolutionary computation in recent years and a number of well performing algorithms have been published. The quick and highly efficient multiobjective evolutionary algorithm based on dominating tree has been criticized mainly for its restricted elite archive and absence of density estimation. This paper improves the algorithm in these two aspects. The nearest distance between the node and other nodes in the same sibling chain is used as its density estimation; the population growing and declining strategies are proposed to avoid the retreating and shrinking phenomenon caused by the restricted elite archive. The simulation results show that the improved algorithm is able to maintain a better spread of solutions and converge better in the obtained nondominated front compared with NSGA-II, SPEA2 and the original algorithm for most test functions.

The business environment is fully characterized with uncertainties. In order to minimize risk and maximize future returns proper portfolio model must be designed. In this paper the application of fuzzy theory to portfolio selection is presented. Fuzzy logic is utilized in the estimation of expected return and risk. Using fuzzy logic, managers can extract useful information and estimate expected return by using not only statistical data, but also economical and financial behaviours of the companies and business strategies. In the formulated fuzzy portfolio model, fuzzy set theory gives chance of possibility trade-off between risk and return. This is obtained by assigning satisfaction degree between criteria and constraints and defining tolerance for the constraints in order to obtain goal value in objective risk function. Using the formulated fuzzy portfolio model, a Genetic Algorithm (GA) is applied to find optimal values of risky securities. The obtained results satisfy the efficiency of the proposed method.

Pairs mining targets to mine pairs relationship between entities such as between stocks and markets in financial data mining. It has emerged as a kind of promising data mining applications. Due to practical complexities in the real-world pairs mining such as mining high dimensional data and considering user preference, it is challenging to mine pairs of interest to traders in business situations. This paper presents fuzzy genetic algorithms to deal with these issues. We introduce a fuzzy genetic algorithm framework to mine pairs relationship, and propose strategies for the fuzzy aggregation and ranking of identified pairs to generate final optimum pairs for decision making. The proposed approaches are illustrated through mining stock pairs and stocktrading rule pairs in stock market. The performance shows that the proposed approach is promising for mining pairs helpful for real trading decision making.

Feature selection plays an important role in pattern classification. In this paper, a hybrid genetic algorithm (HGA) is adopted to find a subset of the most relevant features. The approach utilizes an improved estimation of the conditional mutual information as an independent measure for feature ranking in the local search operations. It takes account of not only the relevance of the candidate feature to the output classes but also the redundancy between the candidate feature and the already-selected features. Thus, the ability of the HGA to search for the optimal subset of features has been greatly enhanced. Experimental results on a range of benchmark datasets demonstrate that the proposed method can usually find the excellent subset of features on which high classification accuracy is achieved.

In this paper, we proposed evolutionary filter and classifier ensemble. We designed the face recognition system that is consists of training module and testing module. The training module is evolution step, that process make filter and classifier combination. In testing step, we identified face recognition using knowledge from made training step. The filters are applied preprocessing step. Captured images are varying illuminant images so we proposed evolutionary preprocessing filter combinations and classifier ensemble for data context. The Proposed classifier selection for efficient object recognition based on evolutionary computation and data context knowledge called context based evolutionary system. In proposed method, we distinguish the data characteristics of input image and filter selects a classifier system accordingly using evolutionary algorithm. The proposed method is high more than single method.

Based on the previous introduced Quantum-behaved Particle Swarm Optimization (QPSO), in this paper, a revised QPSO with novel iterative equation is proposed. While the iterative equation in the QPSO is educed from exponential distribution, the novel one derives from the distribution function of the sum of two random variables with exponential and normal distribution, respectively. The Revised QPSO also maintains the mean best position of the swarm as in the previous QPSO to make the swarm more efficient in global search. The experiment results on benchmark functions show that Revised QPSO has stronger global search ability than QPSO and PSO.

This paper proposes a new selection mechanism in evolutionary algorithm for fuzzy systems that can be applied to robot learning of shooting ability in robot soccer. In generic evolutionary algorithms, evaluation and selection are performed on the chromosome level, where a selected chromosome may include non-effective or bad genes. This may lead to an increase in the uncertainty of the solutions. To solve this problem, we propose a rule-scoring method for gene level selection, which grades genes at the same position in the chromosomes. This method is applied to a fuzzy path planner for the shooting of a soccer robot, where each fuzzy rule is encoded as a gene. Simulation and experimental results show the effectiveness and the applicability of the proposed method.

Due to the uncertainty of the processing time in the practical production, no idle flow shop scheduling problem with fuzzy processing time is introduced. The objective is to find a sequence that minimizes the mean makespan and the spread of the makespan by using a method for ranking fuzzy numbers. The particle swarm optimization (PSO) is a population-based optimization technique that has been applied to a wide range of problems, but there is little reported in respect of application to scheduling problems because of its unsuitability for them. In the paper, PSO is redefined and modified by introducing genetic operations such as crossover and mutation to update the particles, which is called GPSO and successfully employed to solve the formulated problem. Several benchmarks with fuzzy processing time are used to test GPSO. Through the comparative simulation results with genetic algorithm, the feasibility and effectiveness of the proposed method are demonstrated.

The widespread use of the web technology presents an opportunity for business to use the Internet as a tool for electronic customer relationship management (e-CRM). Despite agreement that e-CRM has direct or indirect impact on customer satisfaction, the significance and the determinants of e-CRM in influencing customer satisfaction have not been well researched. This study empirically develops a temporal model explaining the relationship between e-CRM and online customer satisfaction. A theoretical framework that consists of e-CRM initiatives: system quality, information quality and service quality; intrinsic success: responsiveness and efficiency; and objective: customer satisfaction is further expanded. An electronic questionnaire is used to collect sample data. Then this model is tested by means of the statistical analysis method of Structure Equation Model (SEM). These findings should be of great interest to both researchers and practitioners. Discussion and implication are presented in the end.

This paper considers nonlinearly mixed integer reliability design problems in which both the number of redundancy components and the corresponding reliability of each component in each subsystem are to be decided simultaneously so as to maximize the system reliability. The reliability design problems have been studied in the literature for decades, usually using mathematical programming or heuristic/metaheuristic optimization approaches. The difficulties encountered for both methodologies are to maintain feasibility with respect to three nonlinear constraints, namely, cost and weight constraints, and constraints on the products of volume and weight. A penalty-guided particle swarm optimization approach is presented for solving the mixed integer reliability design problems. It can efficiently and effectively search over promising feasible and infeasible regions to find the feasible optimal or near optimal solution. Numerical examples indicate that the proposed approach performs better than other approaches for four reliability-redundant allocation design problems considered in this paper.

In the urban/territorial planning process, the quality of the evaluation procedure is crucial. It is necessary to select and implement innovative tools able to handle the huge amount of available data concerning territorial systems in order to extract useful information from them to enhance the quality of evaluation procedure for urban/territorial planning. This paper selects some tools derived from Artificial Intelligence, and incorporated GIS through the elaboration of various types of available data, to extract and build knowledge directly from experimental data and also to represent the extracted knowledge very effectively and communicatively, in the form of sets of spatial transformation rules. It describes the structure of the data mining tools which are most suitable for applications in the field of urban planning, aimed at discovering the transformation rules driving the evolution of cities in special of metropolitan in analysis.

Vibration control is a pivotal study subject for vehicle suspension system. In this paper, the physical model of semi-active quarter-coach suspension was established by the base of the theories of Buckingham’s Pi theorem. According to the characteristics of the semi-active suspension of diesel-truck, a hybrid intelligent control algorithm-Fuzzy cerebellar model articulation control combined the Fuzzy logic control with cerebellar model articulation control techniques was presented and used to perform online control of semi-active suspension for the first time, novel weight-update laws were derived that guarantee the stability of closed-loop system, both information retrieval and learning rules were described by algebraic equations in matrix form. The results of experiment of closed -loop excited by three typical vibration signals showed that the control strategy proposed here can obviously reduce the value of mean square root of vertical acceleration for semi-active suspension system, compared with the traditional control strategy.

In this paper, we propose semantic Bayesian networks that infer the user’s intention based on Bayesian networks and their semantic information. Since conversation often contains ambiguous expressions, managing the context or the uncertainty is necessary to support flexible conversational agents. The proposed method drives the mixed-initiative interaction (MII) that prompts for missing concepts and clarifies for spurious concepts to understand the user’s intention correctly. We have applied it to an information retrieval service for Web sites so as to verify the usefulness.

The Traveling Salesman Problem (TSP) is a very hard optimization problem in the field of operations research. It has been shown to be NP-hard, and is an often-used benchmark for new optimization techniques.This paper will to bring up a three-tier multi-agent approach for solving the TSP. This proposed approach supports the distributed solving to the TSP. It divides into three-tier (layer), the first tier is ant colony optimization agent and its function is generating the new solution continuously; the second-tier is genetic algorithm agent, its function is optimizing the current solutions group; and the third tier is fast local searching agent and its function is optimizing the best solution from the beginning of the trial. Ultimately, the experimental results have shown that the proposed hybrid approach has good performance with respect to the quality of solution and the speed of computation.

An agent in a multi-agent system (MAS) has to select appropriate agents to assign tasks. Unfortunately no agent in an open environment can identify the states of all agents, so this selection must be done according to local information about the other known agents; however this information is limited and may contain uncertainty. In this paper we investigate how overall performance of MAS is affected by learning parameters for adaptive strategies to select partner agent for collaboration. We show experimental results using simulation and discuss why overall performance of MAS varies.

To enhance the performance of evolutionary algorithms, different parallel computation models have been proposed, and they have been implemented on parallel computers to speed up the computation. Instead of using expensive parallel computing facilities, in this paper we propose to implement parallel evolutionary computation models on easily available networked PCs, and present a multi-agent framework to support parallelism. To evaluate the proposed approach, different kinds of experiments have been conducted to assess the developed system and the preliminary results show the efficiency of our approach.

In this study, we introduce a neurogenetic approach to the design of fuzzy controllers. The design procedure exploits the technology of Computational Intelligence (CI) focusing on the use of genetic algorithms and neurofuzzy networks (NFN). The crux of the design concerns the selection and determination of optimal values of the scaling factors of the fuzzy controllers, which are essential to the entire optimization process. First, the tuning of the scaling factors of the fuzzy controller is carried out, and then the development of a nonlinear mapping for the scaling factors is realized by using GA- based NFN.

In this paper, we propose a architecture of Genetic Algorithms (GAs)-based Polynomial Neural Networks(PNN), discuss a comprehensive design methodology and carry out a series of numeric experiments. GA-based design procedure at each stage (layer) of PNN leads to the selection of preferred nodes (or PNs) with optimal parameters (such as the number of input variables, input variables, and the order of the polynomial) available within PNN. A detailed design procedure is discussed in detail. To evaluate the performance of the GA-based PNN, the model is experimented with by using Medical Imaging System (MIS) data for application to Multi-variable software process. A comparative analysis shows that the proposed GA-based PNN is model with higher accuracy as well as more superb predictive capability than other intelligent models presented previously.

The goal of this paper is to provide a visual structure to view the temporal and thematic content of English news documents. This paper combines both information extraction techniques and visualization techniques to visualize the documents content. Topical information present in the documents are extracted by applying Booksteins model and by calculating the various measures associated with the model. Temporal information is extracted as date information from the documents. Later the extracted information is converted into a signal and both discrete and continuous wavelet transfor-mations are applied over the signal. The content of the document is visualized at varying levels of detail by multi resolution analysis techniques available in wavelets. This gives a wavelet based visualization model which enables us to view the spread of topics across a document, the portions of the document that are most involved in topic description, and the contribution of documents in the corpus from the temporal perspective.

In this paper we present LP-TPOP, a domain-independent temporal partial-order planning algorithm that handles CBI-style action model. By the utilization of ground actions variable binding constraints can be eliminated from partial plans, then all the constraints needed to be handled are temporal constraints, which are maintained in a revised simple temporal constraint network (r-STN) structure. The exact scheduling of plan returned by planning phase is calculated by a linear programming module. By exploiting incremental algorithms the efficiency of LP-TPOP can be further improved. LP-TPOP is proved to be sound and complete.

In multi-agent systems, there is often the need for an agent to cooperate with others so as to ensure that a given task is achieved timely and cost effectively. Currently multi-agent systems maximize this through mechanisms such as coalition formation, trust and risk assessments, etc. In this paper, we incorporate the concept of insurance with trust and risk mechanisms in multi-agent systems. The novelty of this proposal is that it ensures continuous sharing of resources while encouraging expected utility to be maximized in a dynamic environment. Our experimental results confirm the feasibility of our approach.

In this paper, we present a new algorithm of swarm intelligence, namely, Cat Swarm Optimization (CSO). CSO is generated by observing the behaviors of cats, and composed of two sub-models, i.e., tracing mode and seeking mode, which model upon the behaviors of cats. Experimental results using six test functions demonstrate that CSO has much better performance than Particle Swarm Optimization (PSO).

In this paper, we propose an improved fuzzy discrete Particle Swarm Optimization method (IFD-PSO), and apply this method to TSP. We use fuzzy matrix space to represent the corresponding TSP solution, and bring forward the transformation method of fuzzy matrix space. Heuristic information is employed to improve the convergence speed. The experiment results show that IFD-PSO has a better performance and achieves satisfactory effect.

This paper proposed a new event correlation technique to enhance the heuristic of the increment hypothesis updating (IHU) algorithm. This approach estimates the likelihood of each fault in the faults set and removes these faults with less likelihood. By this approach we also can determine whether an event is spurious or not. Simulation shows that this approach can get a high accuracy and fast speed of correlation even if the network has a high event loss and spuriousness.

In this paper, we describe our approach to achieve accurate localization using RFID for Cellular Phone Robot (CPR). We solely rely on RFID mechanism and complement coordinate errors accumulated during the wheel-based CPR navigation. We especially focus on how to distribute RFID tags (tag pattern) and how many to place (tag granularity) on the floor to accomplish efficient navigations. We define the error in navigation and use it to compare the effectiveness of various RFID floor settings through a simulation. Identified tag patterns and granularities would be useful to many USN applications where the adoption of RFID technology is appropriate.

This paper presents the design and prototype of a mesoscale (13 cm long) six-legged walking robot whose locomotion is actuated by a piezoelectric actuator named LIPCA, which consists of multiple layers of glass/epoxy and carbon/epoxy that encapsulate a unimorph piezoceramic actuator. Inspired by the walking kinematics of cockroaches, our robot uses the alternating tripod gait (the front and rear legs on the same side move together with the middle leg on the other side for the locomotion), and has six legs that are designed to mimic the function of those of cockroaches. All the experiments with the prototype show a possibility of a small, light, and agile walking robot that is actuated by LIPCA without using any conventional electromagnetic actuator.

Modern manufacturing systems have to cope with dynamic changes and uncertainties such as machine break down, hot orders and other kinds of disturbances. Holonic manufacturing systems (HMS) provide a flexible and decentralized manufacturing environment to accommodate changes dynamically. In this paper, A new class of Time Petri Nets(TPN), Buffer-nets, for defining a Scheduling Holon is proposed, which enhances the modeling techniques for manufacturing systems with features that are considered difficult to model. The proposed novel GA algorithm performs the population alternation according to the features of the evolution of the populations in natural. Simulation results show that the proposed GA is more efficient than standard GAs. The proposed HPGA synthesizes the merits in both PSO and GA. The simulation results of the example show that the methods to scheduling holon are effective for fulfilling the scheduling problem.

Dysarthria is a collection of motor speech disorder. A severity of dysarthria is traditionally evaluated by human expertise or a group of listener. This paper proposes a new indicator called

speech consistency score (SCS)

. By considering the relation of speech similarity-dissimilarity, SCS can be applied to evaluate the severity of dysarthric speaker. Aside from being used as a tool for speech assessment, SCS can be used to predict the possible outcome of speech recognition as well. A number of experiments are made to compare predicted recognition rates, generated by SCS, with the recognition rates of two well-known recognition systems, HMM and ANN. The result shows that the

root mean square error

between the prediction rates and recognition rates are less than 7.0% (R

2

= 0.74) and 2.5% (R

2

= 0.96) for HMM and ANN, respectively. Moreover, to utilized the use of SCS in general case, the test on unknown recognition set showed the error of 11 % (R

2

= 0.48) for HMM.

This paper presents the results of recent modification in the

Constructive Adaptive User Interface

(CAUI) that induces a model of emotional impressions towards certain musical piece structures and improvises a piece based on the model. The CAUI previously employed a ready-made melody generating module with its internal workings abstracted from the CAUI. Utilizing such black-box modules, however, may impede further effort to enhance the CAUI. To address this problem, a replacement module that automatically creates tunes tailored to the listener’s impressions has been incorporated. Current results indicate that the CAUI may induce relevant relations that can support the adaptive improvisation of impression-causing tunes of a musical piece.

In this paper, we propose an adaptive inventory control model for a supply chain consisting of one supplier and multiple retailers with nonstationary customer demands. The objective of the adaptive inventory control model is to minimize inventory related cost. The inventory control parameter is safety lead time. Unlike most extant inventory control approaches, modeling the uncertainty of customer demand as a statistical distribution is not a prerequisite in this model. Instead, using a reinforcement learning technique called action-reward based learning, the control parameter is designed to adaptively change as customer demand pattern changes. A simulation based experiment was performed to compare the performance of the adaptive inventory control model.

We propose

C

ontext-

A

ware

P

R

oduct

B

undling

A

rchitecture (CARBA). It is necessary for the various products to be easily and immediately integrated, according to customers’ changed requirements. In order to integrate information from various resources such as airline, hotel reservation, and so on, a semantic web service supporting an ontology based travel information system is required. CARBA is basically implemented as a semantic web service, with several components for reconfiguring a bundle of traveling products, and is guaranteeing traveler’s mobility in ubiquitous computing environments.

The Semiring Constraint Satisfaction Problem (SCSP) framework is a popular approach for the representation of partial constraint satisfaction problems. In this framework preferences (semiring values) can be associated with tuples of values of the variable domains. Bistarelli et al. [1] define an abstract solution to a SCSP which consists of the best set of solution tuples for the variables in the problem. Sometimes this abstract solution may not be good enough, and in this case we want to change the constraints so that we solve a problem that is slightly different from the original problem but has an acceptable solution. In [2] we propose a relaxation of a SCSP where we define a measure of distance (a semiring value from a second semiring) between the original SCSP and a relaxed SCSP. In this paper we show how the two semirings can be combined into a single semiring. This combined semiring structure will allow us to use existing tools for SCSPs to solve Combined Semiring Relaxations of SCSPs. At this stage our work is preliminary and needs further investigation to develop into a useful algorithm.

In analysis of the market, detecting not only differences in consumer groups or changes but also their causal factors observed in consumer behavior is expected because it enables the marketer to take marketing actions. Although rule-discovery approaches can efficiently identify differences in groups or changes, it is still difficult to explain the causes of them. In this paper we propose an algorithm to detect causal differences in two bayesian networks by search and probability inference. We perform some experimental studies to analyze consumer behavior in purchasing personal computer.

The Generalized Minimum Spanning Tree (GMST) problem requires spanning exactly one node from every cluster in an undirected graph. GMST problems are encountered in telecommunications network planning. A Tabu Search (TS) for the GMST problem is presented in this article. In our computational tests on 194 TSPLIB instances, TS found 152 optimal solutions. For those 42 unsolved instances, our algorithm has improved some previously best known solutions. Lower bounds of some unknown problems are improved by our heuristic relaxation algorithm.

This paper describes an approach to the investigation of brood size in the brood recombination crossover method in genetic programming for object recognition problems. The approach is examined and compared with the standard crossover operator on three object classification problems of increasing difficulty. The results suggest that the brood recombination method outperforms the standard crossover operator for all the problems in terms of the classification accuracy. As the brood size increases, the system effective performance can be improved. When it exceeds a certain point, however, the effective performance will not be improved and the system will become less efficient.

There are two main objectives for this paper : (1) we will propose a more general class of consecutive-component systems which generalizes both the typical consecutive-

k

-out-of-

n

:F systems and two-dimensional consecutive-

k

- out-of-

n

:F systems, (2) we will propose an immune algorithm to investigate the optimal maintenance policy for the proposed consecutive-component systems. Numerical results are reported and compared with those of implicit enumeration.

An application of Immune Genetic Algorithm (IGA) suitable for the optimal design to intelligent AC contactors is presented in this paper. Besides the ability of stochastic global searching of Simple Genetic Algorithm (SGA), the IGA draws into the mechanisms exist in biological immune system such as immune memory, immune regulation, antibody diversity and others. The simulation results show that IGA overcomes the disadvantages of premature convergence of SGA, and improve the global searching efficiency and capability. This algorithm has been successfully used in the optimal design to the intelligent AC contactors.

Geometric constraint problem is equivalent to the problem of solving a set of nonlinear equations substantially. In this paper we propose a new optimization algorithm—organizational evolutionary algorithm (OEA) and apply it into the geometric constraint solving. In OEA the colony is composed of the organizations. Three organizational evolutionary operators–split operator, merging operator and coordinating operator can lead the colony to evolve. These three kinds of operators have different functions in the algorithm. Split operator limits the scale of the organization, and makes sure a part of organization come into next generation directly, which maintains the variety of the generation. Merging operator makes use of the leader’s information fully and acts as a local searching function. Cooperating operator increases the degree of adaptability between the two organizations by the interactions. The experiment shows that OEA has good capability in the geometric constraint solving.

In this paper, we incorporate the insurance concept in buying and selling model for agents to trade in the open multi-agent marketplace. During buying, agents purchase insurance as a method to search for potential sellers and select their partners based on the information provided by insurance agents. During selling, agents purchase insurance as a method to protect themselves against potential risk. The insurance concept greatly simplifies the trading procedure in the open marketplace. The novelty of this proposal is that it ensures a dynamic trading environment while agents continue to seek maximum utility and being fully protected by insurance. Our experimental results confirm the feasibility of our approach.

In this paper, we argue that the insertion of domain knowledge into ensemble of diverse evolutionary checkers can produce improved strategies and reduce evolution time by restricting search space. The evolutionary approach for game is different from the traditional one that exploits knowledge of the opening, middle, and endgame stages, so that it is not sometimes efficient to evolve simple heuristic that is found easily by humans because it is based purely on a bottom-up style of construction. In this paper, we have proposed the systematic insertion of opening knowledge and an endgame database into the framework of evolutionary checkers. Also, common knowledge, the combination of diverse strategies is better than the single best one, is inserted into the middle stage and is implemented using crowding algorithm and a strategy combination scheme. Experimental results show that the proposed method is promising for generating better strategies.

In computer games, the level design and balance of character attributes are the key features of interesting games. Level designers adjust the attributes of the game characters and opponent behavior to create appropriate levels of difficult, and avoid player frustration. Generally, opponent behavior is defined by a static script, however, this results in repetitive levels and environments, making in difficult to maintain the player’s interest. Accordingly, this paper proposes a dynamic scripting method that can sustain the degree of interest intended by the level designer by adjusting the opponent behaviors while playing the game. The player’s countermeasure pattern for dynamic level design is modeled using a Gaussian Mixture Model (GMM). The proposed method is applied to a shooter game, and the experimental results maintain the degree of interest intended by the level designer.

Customer relationship management (CRM) is one of the biggest concerns in a telecom company. Application of data mining can support telecom CRM effectively, and a systematical architecture of data mining has been needed to support every aspects of telecom CRM roundly. To solve this problem, the systematical application architecture of data mining in telecom CRM is established and components of five modules in the architecture are specified in this paper. Data mining algorithms based-on swarm intelligence improved by our own have been adopted in these modules. SIMiner, a self-development data mining software system based on swarm intelligence, is applied in this architecture. Finally, an application example is given to illuminate that telecom companies can make marketing strategies roundly and effectively with the support of the application architecture.

The high incidence of brain disease, especially brain tumor, has increased significantly in recent years. It is becoming more and more concernful to discover knowledge through mining medical brain image to aid doctors’ diagnosis. In this paper, we introduce a notion of image sequence similarity patterns (ISSP) for medical image database. These patterns are significant in medical images because it is the similarity of objects each of which has an image sequence that is meaningful. We design the new algorithms with the guidance of the domain knowledge to discover ISSP for similarity retrieval. Our experiments demonstrate that the results of similarity retrieval are satisfying.

NB(naive Bayes) is a probabilistic classification model, which is based on the attribute independence assumption. However, in many real-world data mining applications, this assumption is often violated. Responding to this fact, researchers have made a substantial amount of effort to improve NB’s accuracy by weakening its attribute independence assumption. For a recent example, Webb et al.[1] propose a model called Averaged One-Dependence Estimators, simply AODE, which weakens the attribute independence assumption by averaging all models from a restricted class of one-dependence classifiers. Motivated by their work, we believe that assigning different weights to these one-dependence classifiers can result in significant improvement. Based on this belief, we present an improved algorithm called Weightily Averaged One-Dependence Estimators, simply WAODE. We experimentally tested our algorithm in Weka system[2], using the whole 36 UCI data sets[3] selected by Weka[2], and compared it to NB, SBC[4], TAN[5], NBTree[6], and AODE[1]. The experimental results show that WAODE significantly outperforms all the other algorithms used to compare.

This paper proposes SV-kNNC, a new algorithm for k-Nearest Neighbor (kNN). This algorithm consists of three steps. First, Support Vector Machines (SVMs) are applied to select some important training data. Then, k-mean clustering is used to assign the weight to each training instance. Finally, unseen examples are classified by kNN. Fourteen datasets from the UCI repository were used to evaluate the performance of this algorithm. SV-kNNC is compared with conventional kNN and kNN with two instance reduction techniques: CNN and ENN. The results show that our algorithm provides the best performance, both predictive accuracy and classification time.

In this study, support vector machine (SVM) is used as a metamodeling technique to design a business risk identification system. First of all, a bagging sampling technique is used to generate different training sets. Based on the different training sets, different SVM models with different parameters, i.e., base models, are then trained to formulate different classifiers. Finally, a SVM-based metamodel (i.e., metaclassifier) can be produced by learning from all base models. For illustration the proposed metamodel is applied to a real-world business insolvency risk classification problem.

Locally linear embedding approach (LLE) is one of most efficient nonlinear dimensionality reduction approaches with good representational capacity for a broader range of manifolds and high computational efficiency. However, LLE and its variants are based on the assumption that the whole data manifold is evenly distributed so that they fail to nicely deal with most real problems that are unevenly distributed. This paper first proposes an approach to judge whether the manifold is even or not, and then logically divides the unevenly distributed manifold into many evenly distributed sub-manifolds, where the neighourhood size for each sub-manifold is automatically determined based on its structure. It is proved, by visualization and classification experiments on benchmark data sets, that our approach is competitive.

This paper utilizes

structural

properties of those alphanumeric characters, which have variable writing units. Writing units reveal number, shape, size, order of stroke, and speed in writing. It uses a string of pen tip’s positions and tangent angles of every consecutive point as a feature vector sequence of a stroke. We constructed a prototype recognizer that uses the “Dynamic Time Warping” (DTW) algorithm to align handwritten strokes with stored stroke templates and determine their similarity. Separate system is trained for original and preprocessed writing samples and achieved recognition rates of 85.87% and 88.59% respectively. This introduces novel real time handwriting recognition on Nepalese alphanumeric characters, which are independent of number of strokes, as well as their order.

A new calibration model for plasma diagnosis was constructed by combining radio frequency impedance match data, wavelet, and neural network. A total of 30 fault symptoms were simulated with the variations in the four process parameters. Both discrete wavelet transformation (DWT) and continuous wavelet transformation (CWT) were utilized to filter the sensor information. Three types of diagnosis models (raw-, DWT-, and CWT-based models) were constructed. The comparisons revealed that the improvement in the prediction performance of DWT and CWT data models over the raw data model were about 42% and 30%, respectively.

Many existing plagiarism detection systems fail in detecting plagiarism when there are an abundant garbage in the copied programs. This is because they do not use the structural information efficiently. In this paper, we propose a novel plagiarism detection system which uses parse tree kernels. By incorporating parse tree kernels into the system, it efficiently handles the structural information within source programs. A comparison with existing systems such as SID and JPlag shows that the proposed system can detect plagiarism more accurately due to its ability of handling structural information.

A natural question in

Information Bottleneck

method is how many “groups” are appropriate. The dependency on prior knowledge restricts the applications of many

Information Bottleneck

algorithms. In this paper we aim to remove this dependency by formulating the parameter choosing as a model selection problem, and solve it using the minimum message length principle. Empirical results in the documentation clustering scenario indicates that the proposed method works well for the determination of the optimal parameter value for information bottleneck method.

To complete missing values, a solution is to use attribute correlations within data. However, it is difficult to identify such relations within data containing missing values. Accordingly, we develop a kernel-based missing data imputation method in this paper. This approach aims at making optimal statistical parameters: mean, distribution function after missing-data are imputed. We refer this approach to

p

arameter

op

timization method

(POP algorithm, a random regression imputation). We experimentally evaluate our approach, and demonstrate that our POP algorithm is much better than deterministic regression imputation in efficiency of generating an inference on the above two parameters. The results also show our algorithm is computationally efficient, robust and stable for the missing data imputation.

In recent years, the support vector machine (SVM) has been extensively applied to deal with various data classification problems. However, it has also been observed that, for some datasets, the classification accuracy delivered by the SVM is very sensitive to how the cost parameter and the kernel parameters are set. As a result, the user may need to conduct extensive cross validation in order to figure out the optimal parameter setting. How to expedite the model selection process of the SVM has attracted a high degree of attention in the machine learning research community in recent years. This paper proposes an advanced data reduction algorithm aimed at expediting the model selection process of the SVM. Experimental results reveal that the proposed mechanism is able to deliver a speedup of over 70 times without causing meaningful side effects and compares favorably with the alternative approaches.

A new methodology on the algorithm of sequential minimal optimization (SMO) for power system load was presented. In order to solve the problem that support vector machines (SVM) can not deal with large scale data, this paper introduces the modified algorithm of SMO to increase operational speed by use of a single threshold value. Adopting the actual data from the distribution network of a certain domestic city, and the load is forecasted by use of support vector regression (SVR) which is based on the modified SMO algorithm and proper kernel function. The forecasted results are compared with those SVR employing quadratic programming (QP) optimization algorithm and BP artificial neural method, and it is shown that the presented forecasting method is more accurate and efficient.

This paper proposes an effective system to detect adult image. We take this task as a two-class pattern classification problem, The system first applies histogram color model to detect the skin regions, then extracts color, texture, shape features from skin regions, after that, the features is fed to a SVM to determine whether the input image is benign or not. Experimental results show that the proposed method can achieve a satisfactory classification performance with high speed, which is suitable for real-world applications.

Kernel Matching Pursuit Machine (KMPM) is a relatively new learning algorithm utilizing Mercer kernels to produce non-linear version of conventional supervised and unsupervised learning algorithm. But the commonly used Mercer kernels can’t expand a set of complete bases in the feature space (subspace of the square and integrable space). Hence the decision-function found by the machine can’t approximate arbitrary objective function in feature space as precise as possible. Multiresolution analysis (MRA) shows promise for both nonstationary signal approximation and pattern recognition, so we combine KMPM with multiresolution analysis technique to improve the performance of the machine, and put forward a MRA shift-invariant kernel, which is a Mercer admissive kernel by theoretical analysis. An MRA kernel matching pursuit machine (MKMPM) is constructed in this paper by Shannon MRA shift-invariant kernel. It is shown that MKMPM is much more effective in the problems of regression and pattern recognition by a large number of comparable experiments.

This work aims to explore the use of gene expression data in discriminating heterogeneous cancers. We introduce hybrid learning methodology that integrates genetic algorithms (GA) and artificial neural networks (ANN) to find optimal subsets of genes for tissue/cancer classification. This method was tested on two published microarray datasets: (1) NCI60 cancer cell lines and (2) the GCM dataset. Experimental results on classifying both datasets show that our GA/ANN method not only outperformed many reported prediction approaches, but also reduced the number of predictive genes needed in classification analysis.

Based on energy distribution analysis of FRIT coefficients, a novel feature extraction method of low computation complexity in FRIT domain was proposed for texture classification in this paper. A ‘one-against-one’ multi-class SVM with RBF kernel was adopted as classifier. Experiments carried out on abundant texture databases with varying sizes demonstrated its validity.

There are limitations for the existing methods to model multivariate time series because that defining the input components is highly difficult. The main purpose of this paper is to expand the principal components analysis (PCA) method to extract the joint information of multiple variables. First, both the linear correlations and the nonlinear correlations are detected to initialize an embedding delay window, which contains enough information for prediction. Then, the PCA method is expanded to extract the joint information of multiple variables in a complex system. Finally, neural network makes predictions on the basis of approximating both the functional relationship between different variables and the map between current state and future state.

Prolog is a well-known logic programming language. The information on the correct partial answers (cpa) and correct call patterns (ccp) of goals is useful for static analysis of Prolog programs. Decorated tree (DT) semantics is a goal-independent denotational semantics for Prolog that has been shown to be promising in Prolog program analysis. We extend the work in [7] and propose a two-step method for achieving cpa or ccp semantics from DT semantics. This paper is mainly concerned with the design of semantic domains.

A design of the fingerprint capture system based on chip FPS200 is introduced. The capture system uses the MCU interface with DSP. In this paper, a new method based on the intrinsic directional features of fingerprint is proposed to evaluate the quality of images acquired by FPS200 sensor. Using directional map, the available area of image and clearness is checked to calculate the overall image quality score that can be used to quantitatively determine the quality of the fingerprint image. If the quality of image is unsatisfied, prompt is presented: whether the finger is wet or dry, then DSP adjusts the parameters of FPS200 sensor automatically until image of good quality has been acquired. Experimental results show that the performance of AFIS has indeed improved.

We present a method for representing, tracking and human following by fusing distributed multiple vision systems in ISpace, with application to pedestrian tracking in a crowd. And the article presents the integration of color distributions into particle filtering. Particle filters provide a robust tracking framework under ambiguity conditions. We propose to track the moving objects by generating hypotheses not in the image plan but on the top-view reconstruction of the scene. Comparative results on real video sequences show the advantage of our method for multi-object tracking. Simulations are carried out to evaluate the proposed performance. Also, the method is applied to the intelligent environment and its performance is verified by the experiments.

In this paper, we propose a new method for measuring the closeness of multiple sets of rules that are combined using Dempster’s rule of combination to improve classification performance. The closeness provides an insight into combining multiple sets of rules in classification − in what circumambience the performance of combinations of some sets of rules using Dempster’s rule is better than that of others. Experiments have been carried out over the 20-newsgroups benchmark data collection, and the empirical results show that when the closeness between two sets of rules is higher than that of others, the performance of its combination using Dempster’s rule is better than the others.

This paper proposes a multiple SVMs enabled sales forecasting support system (SFSS). The SFSS has a two-stage system architecture. In the first stage, agglomerative hierarchical clustering(AHC) is used to partition the goods into several patterns based on similarity measure. In the second stage, multiple SVMs that best fit partitioned patterns are constructed by finding the appropriate kernel function and the optimal free parameters of SVMs. The experiment shows that this integrated system achieves significant improvement in forecasting performance compared with single SVMs models.

Metal cutting operations constitute a large percentage of the manufacturing activity. Cutting tool condition monitoring is certainly the important monitoring requirement of unintended machining operations. A multi-purpose intelligent tool condition monitoring technique for metal cutting process will be introduced in this paper. The knowledge based intelligent pattern recognition algorithm is mainly composed of a fuzzy feature filter and algebraic neurofuzzy networks. It can carry out the fusion of multi-sensor information to enable the proposed intelligent architecture to recognize the tool condition successfully.

A simplified Fuzzy-PID controller is designed for 2-axes antenna stabilization and tracking system. Next, the performance of the controller is further verified by computer simulations with Matlab, a 2-Axes antenna, and a small unmanned test helicopter. With the advantages of embedded controller and simplified fuzzy control theory, high performance techniques for antenna tracking control are designed. A comparison between the performance of the conventional PID controller and the Fuzzy-PID controller, which is designed by the same PID control gains, is made as a way to verify the performance of the designed antenna servo control system. The proposed Fuzzy-PID controller has better superior performance than the conventional PID controller with respect to all cases of simulations and experiments.

The neuron based nonlinear PID control method for plant with uncertainties is proposed in this paper. In this control system, the neuron based nonlinear PID controller is constructed by selecting the neuron inputs, and the control action of the nonlinear PID controller are determined by modifying the neuron weights on-line. The neuron based nonlinear PID controller produces the control signal in model-free way. With an example of the basis weight control of a paper machine, the experiments are made. The simulation results show that the neuron model-free controller has good control performance, fast transient response and strong robustness.

This paper proposes a color-shape based method (CSBM) based on color, area, and perimeter intercepted lengths of segmented objects in an image. It characterizes the shape of an object by the intercepted lengths obtained by intercepting the object perimeter by eight lines with different orientations passing through the object center. The experimental results show that CSBM provides a better performance than fuzzy color histogram (FCH) and conventional color histogram (CCH). Besides, it is insensitive to translation, rotation, distortion, scaling, and hue variations, but impressionable to contrast and noise variations.

As an effort to detect the mood of a blog, regardless of the length and writing style, we propose a hybrid approach to detecting blog text’s mood, which incorporates commonsense knowledge obtained from the general public (ConceptNet) and the Affective Norms English Words (ANEW) list. Our approach picks up blog text’s unique features and compute simple statistics such as term frequency, n-gram, and point-wise mutual information (PMI) for the SVM classification method. In addition, to catch mood transitions in a given blog text, we developed a paragraph-level segmentation based on a mood flow analysis using a revised version of the GuessMood operation of ConceptNet and an ANEW-based affective sensing module. For evaluation, a mood corpus comprised of real blog texts has been built semi-automatically. Our experiments using the corpus show meaningful results for 4 mood types: happy, sad, angry, and fear.

In this paper, we propose a noble classification algorithm for content-based audio signal retrieval. The algorithm uses the Gradient-Based Fuzzy C-Means with a Mercer Kernel (GBFCM(MK)) to perform clustering of Gaussian Probability Density Function (GPDF) data of a Gaussian Mixture Model (GMM). The GBFCM(MK) algorithm incorporates a kernel method into the GBFCM to implicitly perform nonlinear mapping of the input data into a high-dimensional feature space. Experiments and results for several audio data sets have shown that the GBFCM(MK)-based classification algorithm has accuracy improvements of 3.14%-7.49% over classification algorithms employing the traditional k-means and the Fuzzy C-Mean (FCM), respectively.

Hyperlink structure of the Web provides valuable information for ranking query results and has been used in some famous search engines. The development of search engines such as personalized and topic-sensitive search intensifies the need of quick rank algorithms. In this paper, a link based rank called ExpRank is proposed. It can converge quickly and reserve the fundamental features of PageRank. Experimental results comparing with PageRank on real dataset are also discussed.

This paper presents a biologically-inspired method of perceiving the TROIs(: Texture Region Of Interest) from various texture images. Our approach is motivated by a computational model of neuron cells found in the primary visual cortex. An unsupervised learning schemes of SOM(: Self-Organizing Map) is used for the block-based image clustering, plus 2D spatial filters referring to the response properties of neuron cells is used for extracting the spatial features from an original image and segmenting any TROI from the clustered image. To evaluate the effectiveness of the proposed method, various texture images were built, and the quality of the extracted TROI was measured according to the discrepancies. Our experimental results demonstrated a very successful performance.

To comprehend English-written texts successfully, readers have to construct a referential map of textual information. Referential device is one of the important means for helping readers comprehend a text. The purpose of this study is to develop a computer-assisted environment to enhance EFL college students’ comprehension. Four modules, natural language processing (NLP), user interface, recording, and feedback module are included. Among these four modules, the feedback module compares students’ initial maps with the expert’s. The results of comparison will inform students what referents are incorrect and offer them appropriate scaffoldings. The recording module records all of students’ behavioral data. From the data, the teacher can identify the difficulties students encounter and different performance levels among students with various reading proficiencies.

Enterprise data warehousing technology aims at providing integrated, consolidated and historical data for users to analyze businesses and make decisions. In order to obtain the correct results, the high data quality is required. In this paper, we analyze the quality problems of enterprise data warehouse and present an object-oriented framework for data quality management. In this framework, an object-oriented data quality model (OODQM) is built. The data quality requirements, the participators, the data quality checking object, and the possible data quality problems, form the core components of OODQM. The method we provide is a goal-driven method. Once the data quality goal is built, we manage data quality by the interaction of those components of OODQM.

Rebuilding Minimalist Grammars (MG) into Categorial Minimalist Grammars (CMG) as extension of MG towards partially commutative linear logic (pCLL) is of significance. But the bijective syntax-semantics interface established in CMG sometimes fails to obtain some semantic proof trees which involves more than one quantifier. Satisfactory solution has not been found yet. Aimed at this problem, we keep the type

psoa

of HPSG (Head-Driven Phrase Structure Grammar) that the glue semantics approach removes and extend HPSG towards a HPSG Deductive System (HDS) as a fragment of pCLL the way MG are rebuilt. We handle the Subcategorization Principle, the Trace Principle, and the Semantic Principle of HPSG with HDS; establish the correspondence between syntax and semantics; obtain with this correspondence the problematic semantic proof tree, and thus solve the above-mentioned problem.

Multi-document summarization has become a key technology in natural language processing. This paper proposes a strategy for Chinese multi-document summarization based on clustering and sentence extraction. As for clustering, we propose two heuristics to automatically detect the proper number of clusters: the first one makes full use of the summary length fixed by the user; the second is a stability method, which has been applied to other unsupervised learning problems. We also discuss a global searching method for sentence selection from the clusters. To evaluate our summarization strategy, an extrinsic evaluation method based on classification task is adopted. Experimental results on news document set show that the new strategy can significantly enhance the performance of Chinese multi-document summarization.

The multi-document summarizer using genetic algorithm-based sentence extraction (SBGA) regards summarization process as an optimization problem where the optimal summary is chosen among a set of summaries formed by the conjunction of the original articles sentences. To solve the NP hard optimization problem, SBGA adopts genetic algorithm, which can choose the optimal summary on global aspect. To improve the accuracy of term frequency, SBGA employs a novel method TFS, which takes word sense into account while calculating term frequency. The experiments on DUC04 data show that our strategy is effective and the ROUGE-1 score is only 0.55% lower than the best participant in DUC04.

Dictionary-based biological concept extraction is still the state-of-the-art approach to large-scale biomedical literature annotation and indexing. The

exact dictionary lookup

is a very simple approach, but always achieves low extraction recall because a biological term often has many variants while a dictionary is impossible to collect all of them. We propose a generic extraction approach, referred to as

approximate dictionary lookup

, to cope with term variations and implement it as an extraction system called

MaxMatcher

. The basic idea of this approach is to capture the significant words instead of all words to a particular concept. The new approach dramatically improves the extraction recall while maintaining the precision. In a comparative study on GENIA corpus, the recall of the new approach reaches a 57% recall while the

exact dictionary lookup

only achieves a 26% recall.

Word Sense Disambiguation(WSD) is one of the traditionally most difficult problems in natural language processing and has broad theoretical and practical implications. One of the main difficulties for WSD systems is the lack of relevant knowledge–commonly known as the knowledge acquisition bottleneck problem. We present in this paper a novel method that utilizes dynamic Web data obtained through Web search engines to effectively enrich the semantic knowledge for WSD systems. We demonstrated through a word sense disambiguation system the large quantity and good quality of the extracted knowledge.

Application of natural language understanding to requirements gathering remains a field that has only limited explorations so far. This paper presents an approach to extract the object oriented elements of the required system. This approach starts with assigning the parts of speech tags to each word in the given input document. Further, to resolve the ambiguity posed by the pronouns, the pronoun resolutions are performed before normalizing the text. Finally the elements of the object-oriented system namely the classes, the attributes, methods and relationships between the classes, sequence of actions, the use-cases and actors are identified by mapping the ‘parts of speech- tagged’ words onto the Object Oriented Modeling Language elements using mapping rules which is the key to a successful implementation of user requirements.

Traditional statistical approaches for identifying multi-word terms have to handle a large amount of noisy data and are extremely time consuming. This paper introduces a

multi-word term extraction system

for extracting multi-word terms from a set of documents based on the co-related

text-segments

existing in these documents. The system uses a short predefined stoplist as an initial input to segment a set of documents into text-segments, calculates the segment-weights of all text-segments, and then applies the short text-segments to segment the longer text-segments based on the weight values recursively until all text-segments cannot be further divided. The resultant text-segments can thus be identified as terms based on a specified threshold. The initial experimental result on a set of traditional Chinese documents shows that this system can achieve a minimum of 76.39% of recall rate and a minimum of 91.05% of precision rate on retrieving multiple occurrences terms, which include 18.30% of new identified terms.

A new segmentation algorithm for synthetic aperture radar (SAR) image is proposed using multiscale self-growing probabilistic decision-based neural network (MSPDNN). The proposed algorithm is able to find the natural number of category in SAR image based on the Bayesian information criterion (BIC). The learning process starts from a single SAR image at proper scale randomly initialized in the feature space, and grows adaptively during the learning process until most appropriate number of category are found. Experimental results of the proposed algorithm are proposed and compared with that of previous algorithms.

In this paper, we present a real-time face detection method based on hybrid neural networks. We propose a modified version of fuzzy min-max (FMM) neural network for feature analysis and face classification. A relevance factor between features and pattern classes is defined to analyze the saliency of features. The measure can be utilized for the feature selection to construct an adaptive skin-color filter. The feature extraction module employs a convolutional neural network (CNN) with a Gabor transform layer to extract successively larger features in a hierarchical set of layers. In this paper we first describe the behavior of the proposed FMM model, and then introduce the feature analysis technique for skin-color filter and pattern classifier.

In this paper, a new GA-based constructive algorithm is proposed for wavelet neural networks. Wavelets will be added to the WNNs from low resolution level to high resolution level. At each resolution, the translation parameters of a new wavelet is trained using GA, and output weights is obtained using least square techniques. The proposed algorithm is suitable to the high dimensional problems.

TSP is a problem finding out the shortest distance out of possible courses where one starts a certain city and turns back to a starting city, visiting every city only once among

N

cities. This paper proposes the new method using both population initialization and sequential transformation method at the same time and then proves the improvement of capability by comparing them with existing methods.

Usually, the pseudo-inverse of the Jacobian matrix needs to be calculated in the conventional laws for the Single Gimbal Control Moment Gyroscopes (SGCMGs). However, the steering law can not work when the Jacobian matrix is singular and its pseudo-inverse is indefinite. To avoid the conditions stated above, a new steering law is designed using radial basis function(RBF) neural networks. This algorithm can output the desired gimbal angles directly according to the momentum command. And also, this algorithm can deal with the singular conditions since the pseudo-inverse of the Jacobian matrix is not needed. Simulation results demonstrate the effectiveness of the steering law.

The purpose of this study is to identify the hierarchical radial basis function neural networks and select important input features for each sub-RBF neural network automatically. Based on the pre-defined instruction/operator sets, a hierarchical RBF neural network is created and evolved by using Extended Compact Genetic Programming (ECGP), and the parameters are optimized by Differential Evolution (DE) algorithm. Empirical results on benchmark system identification problems indicate that the proposed method is efficient.

A novel method to fingerprint classification, in which the naïve Bayes classifier (NB) and OVA SVMs are integrated, is presented. In order to solve the tie problem of combing OVA SVMs, we propose a subsumption architecture dynamically organized by the probability of classes. NB calculates the probability using singularities and pseudo codes, while OVA SVMs are trained on FingerCode. The proposed method not only tolerates ambiguous fingerprint images by combining different fingerprint features, but produces a classification accuracy of 90.8% for 5-class classification on the NIST 4 database, that is higher than conventional methods.

We propose a reconfigurable neural network structure which has capability to process supervised or unsupervised learning algorithm computation. The proposed structure is based on modular structure which can configure artificial neural network architecture flexibly. Main processing unit of the proposed structure is designed to obtain flexibility of its internal structure by specific instructions. Therefore it is possible to configure MLP (Multi-Layer Perceptron) with back-propagation for alphabet recognition and parallel SOM for impulse noise detection problem. The performance comparison with the matlab simulation shows its value in the aspects of reliability.

Car plate recognition is an important problem in many traffic related applications. In this paper, we focus on car plate localization—the first step of car plate recognition. We propose a hybrid method based on Pulse Coupled Neural Network (PCNN) and wavelet analysis. First of all, we use PCNN to enhance the image. Then, regions of interest (ROIs) will be get through wavelet analysis. After that, PCNN enhancement is applied again in ROIs, followed by a training and classification process for final labelling ROIs as car plate regions or not. Experiment results show that the precision can get 96%, which is higher than other localization methods on the same image database.

Owing to the highly complicated nature and the escalating cost involved in construction claims, it is highly desirable for the parties to a dispute to know with some certainty how the case would be resolved if it were taken to court. The use of artificial neural networks can be a cost-effective technique to help to predict the outcome of construction claims, on the basis of characteristics of cases and the corresponding past court decisions. This paper presents the application of a split-step particle swarm optimization (PSO) model for training perceptrons to predict the outcome of construction claims in Hong Kong. The advantages of global search capability of PSO algorithm in the first step and local fast convergence of Levenberg-Marquardt algorithm in the second step are combined together. The results demonstrate that, when compared with the benchmark backward propagation algorithm and the conventional PSO algorithm, it attains a higher accuracy in a much shorter time.

In this paper, a robust image segmentation method is proposed. The relationship between pixel intensities and distance between pixels are introduced to the traditional neighbourhood potential function To perform an unsupervised segmentation, the Bayes Information Criterion (BIC) is used to determine the class number, the K-means is employed to initialise the classification and calculate the mean values and variances of the classes. The segmentation is transformed to maximize a posteriori(MAP) procedure. Then, the Iterative Conditional Model (ICM) is employed to solve the MAP problem. In the experiments, the proposed method is compared with other segmentation techniques, for noisy image segmentation applying on synthetic and real images. The experiment results shows that the proposed algorithm is the better choice.

To achieve natural human computer interface, a new gaze detection method is proposed, which allows user’s natural head and eye movement with one camera system and four IR-LED illuminators. This paper has following 4 advancements compared to previous works. First, all procedures for detecting gaze position are operated automatically. Second, although we use the seethrough glasses attached with eye detecting camera, the change of facial position cannot affect the gaze detection accuracy. Third, we use elliptical hough transform and geometric transform in order to detect accurate pupil region. Fourth, to solve the problem of ambiguous coin face-on of pupil shape, we use the EKF (Extended Kalman Filter) and can track continuous eye movement.

Most of the image-based traffic monitoring system (ITMS) adopts auto iris lens to control the amount of incoming light to camera. Auto iris mechanism measures total light energy in camera’s field of view (FOV) and controls iris opening mechanically and inversely proportional to the light energy perceived. Thus, under counterlight, it causes the reduction of incoming light to produce dark scene where brighter one is desirable. To overcome this difficulty, some camera provides a function to define a region of interest (ROI) in FOV and measures light energy only in ROI. Thus, if we leave out counterlight area from ROI, the iris may properly be controlled. However, in ITMS, it frequently happens that large vehicle with white or black roof passes under camera, covers most of the FOV, and results in undesirable iris change. In this paper, we suggest a new iris control mechanism, called user-controlled iris (UCI), in which iris control depends only on background brightness. Since UCI is not sensitive to counterlight or foreground object’s brightness, it can maintain the optimal environment for vehicle detection for ITMS.

Invariant object recognition (IOR) has been one of the most active research areas in computer vision. However, there is no technique able to achieve the best performance in all possible domains. Out of many techniques, convolutional network (CN) is proven to be a good candidate in this area. Given large numbers of training samples of objects under various variation aspects such as lighting, pose, background, etc., convolutional network can learn to extract invariant features by itself. This comes with the price of lengthy training time. Hence, we propose a circular pairwise classification technique to shorten the training time. We compared the recognition accuracy and training time complexity between our approach and a benchmark generic object recognizer LeNet7 which is a monolithic convolutional network.

This paper presents an efficient approach to achieve fast and accurate face detection in still gray level images. The structure of eye region is used as a robust cue to find possible eye pairs. Candidates of eye pair at different scales are discovered by finding regions which roughly match with the binary eye pair template. To obtain real ones, all the eye pair candidates are then verified by using SVM. Faces are finally located according to the eyes position. The proposed method is robust to deal with illumination changes, moderate rotations, glasses wearing and partial face occlusions. The proposed method is evaluated on the BioID face database. Comparative experimental results demonstrate its effectiveness.

In this paper, we present a new and fast algorithm of fuzzy segmentation for MR image, which is corrupted by the intensity inhomogeneity. The algorithm is formulated by modifying the FFCM algorithm to incorporate a gain field, which compensate for such inhomogeneities. In each iteration, we allow the gain field transforming to a gain field image and filter it using an iterative low-pass filter, and then revert the gain field image to gain field term again for the next iteration. We also use c-means algorithm initializing the centroids to further accelerate our algorithm. Our method reduces lots of executive time and will obtain a high-quality result. The efficiency of the algorithm is demonstrated on different magnetic resonance images.

Presently, distributed source coding (DSC) and distributed video coding (DVC) are given high attention in sensor network and other multimedia transmission. That is due to their contribution to easy encoding and robust multimedia communication. In this paper, we propose a new DVC framework where two LVQ (lattice vector quantization) and a rate-variable LDPC (lowdensity parity-check) are exploited. In both LVQ and LDPC, we use the interpolated from decoded frames as side information to help coding. Because of LVQ’s exploitation to dependence of source and the better error-correcting capacity of LDPC, our system achieves more than 1 dB improvement in PSNR than the referenced. Meanwhile, the property of low-complexity encoding is still preserved.

In this paper, an edge-based matching algorithm is proposed that combines the generalized Hough transform (GHT) and the Chamfer matching to complement weakness of either method. First, the GHT is used to find approximate object positions and orientations, and then these parameters are refined using the Chamfer matching with distance interpolation. Matching accuracy is further enhanced by using a subpixel algorithm. The algorithm was successfully tested on many images containing various electronic components.