AI 2004: Advances in Artificial Intelligence

A real-world labor market has complex worksite interactions between a worker and an employer. This paper investigates the behavior patterns of workers and employers with a job capacity and a job concentration empirically considering a strategic coalition in an agent-based computational labor market. Here, the strategic coalition can be formed autonomously among workers and/or among employers. For each experimental treatment, the behavior patterns of agents are varied with a job capacity and a job concentration depending on whether a coalition is allowed. Experimental results show that a strategic coalition makes workers and employers aggressive in worksite interactions against their partners.

Privacy preservation is crucial in the ubiquitous computing because a lot of privacy-sensitive information can be collected and distributed in the ubiquitous computing environment. The anonymity-based approach is one of well-known ones for privacy preservation in ubiquitous computing. It allows users to use pseudonyms when they join in a service area of the ubiquitous computing environment. This approach can protect users’ privacy by hiding the user’s real IDs, but it makes it difficult to provide ID-based services like buddy service, safety alert service, and so on. This paper proposes a multiagent architecture to make it possible to provide ID-based services in the anonymity-based privacy awareness systems. The proposed architecture employs so-called the

white page agent

which maintains the current pseudonyms of users and allows users to get the pseudonyms of other users from the agent. Even though the white page agent contains the pseudonyms of users, it is enforced not to disclose the association of real user IDs with pseudonyms by adopting encryption techniques. This paper presents in detail the proposed architecture, its constituent components and their roles and how they provide ID-based services in the anonymity-based system. As an example, it also presents how buddy service can take place in the proposed architecture.

In this paper, we present a Top-Down/Bottom-Up (TDBU) design approach for critical damage reporting in intelligent sensor networks. This approach is a minimal hierarchical decomposition of the problem, which seeks a balance between achievability and complexity. Our simulated environment models two-dimensional square cells as autonomous agents which sense their local environment, reporting critical damage as rapidly as possible to a report delivery site (portal) by using only the adjacent-cell communication links. The global goal is to design agent properties which will allow the multi-agent network to detect critical damage anywhere on the network and to communicate this information to a portal whose location is unknown to the agents. We apply a TDBU approach together with genetic algorithms (GA) to address the global goal. Simulations show that our system can successfully report critical damage much better than random methods.

Traditionally optimization of defence operations are based on the findings of human-based war gaming. However, this approach is extremely expensive and does not enable analysts to explore the problem space properly. Recent research shows that both computer simulations of multi-agent systems and evolutionary computation are valuable tools for optimizing defence operations. A potential maneuver strategy is generated by the evolutionary method then gets evaluated by calling the multi–agent simulation module to simulate the system behavior. The optimization problem in this context is known as a black box optimization problem, where the function that is being optimized is hidden and the only information we have access to is through the value(s) returned from the simulation for a given input set. However, to design efficient search algorithms, it is necessary to understand the properties of this search space; thus unfolding some characteristics of the black box. Without doing so, one cannot estimate how good the results are, neither can we design competent algorithms that are able to deal with the problem properly. In this paper, we provide a first attempt at understanding the characteristics and properties of the search space of complex adaptive combat systems.

An Autonomous Science Agent is currently flying onboard the Earth Observing One Spacecraft. This software enables the spacecraft to autonomously detect and respond to science events occurring on the Earth. The package includes software systems that perform science data analysis, deliberative planning, and run-time robust execution. Because of the deployment to a remote spacecraft, this Autonomous Science Agent has stringent constraints of autonomy and limited computing resources. We describe these constraints and how they are reflected in our agent architecture.

With the wonders of the Internet and the promises of the worldwide information infrastructure, a highly secure authentication system is desirable. Biometric has been deployed in this purpose as it is a unique identifier. However, it also suffers from inherent limitations and specific security threats such as biometric fabrication. To alleviate the liabilities of the biometric, a combination of token and biometric for user authentication and verification is introduced. All user data is kept in the token and human can get rid of the task of remembering passwords. The proposed framework is named as Bio- Discretization. Bio-Discretization is performed on the face image features, which is generated from Non-Negative Matrix Factorization (NMF) in the wavelet domain to produce a set of unique compact bitstring by iterated inner product between a set of pseudo random numbers and face images. Bio- Discretization possesses high data capture offset tolerance, with highly correlated bitstring for intraclass data. This approach is highly desirable in a secure environment and it outperforms the classic authentication scheme.

The cochlear ear implant has become a standard clinical intervention for the treatment of profound sensorineural hearing loss. After 20 years of research into implant design, there are still many unanswered clinical questions that could benefit from new analysis and modelling techniques. This research aims to develop techniques for extracting the cochlea from medical images to support clinical outcomes. We survey the challenges posed by some of these clinical questions and the problems of cochlea modeling. We present a novel algorithm for extracting tubular objects with non-circular cross-sections from medical images, including results from generated and clinical data. We also describe a cochlea model, driven by clinical knowledge and requirements, for representation and analysis. The 3-dimensional cochlea representation described herein is the first to explicitly integrate path and cross-sectional shape, specifically directed at addressing clinical outcomes. The tubular extraction algorithm described is one of very few approaches capable of handling non-circular cross-sections. The clinical results, taken from a human CT scan, show the first extracted centreline path and orthogonal cross-sections for the human cochlea.

In the last years, there has been a large growth in gene expression profiling technologies, which are expected to provide insight into cancer related cellular processes. Machine Learning algorithms, which are extensively applied in many areas of the real world, are not still popular in the Bioinformatics community. We report on the successful application of the combination of two supervised Machine Learning methods, Bayesian Networks and

k

Nearest Neighbours algorithms, to cancer class prediction problems in three DNA microarray datasets of huge dimensionality (

Colon, Leukemia

and

NCI-60

). The essential gene selection process in microarray domains is performed by a sequential search engine and after used for the Bayesian Network model learning. Once the genes are selected for the Bayesian Network paradigm, we combine this paradigm with the well known

K

NN algorithm in order to improve the classification accuracy.

Vector quantizer takes care of special image features like edges also and hence belongs to the class of quantizers known as second generation coders. This paper proposes a vector quantization using wavelet transform and enhanced SOM algorithm for medical image compression. We propose the enhanced self-organizing algorithm to improve the defects of SOM algorithm, which, at first, reflects the error between the winner node and the input vector to the weight adaptation by using the frequency of the winner node. Secondly, it adjusts the weight in proportion to the present weight change and the previous weight change as well. To reduce the blocking effect and improve the resolution, we construct vectors by using wavelet transform and apply the enhanced SOM algorithm to them. Our experimental results show that the proposed method energizes the compression ratio and decompression ratio.

This paper analyzes the variability of pulse waveforms by means of approximate entropy (

ApEn

) and classifies three group objects using support vector machines (SVM). The subjects were divided into three groups according to their cardiovascular conditions. Firstly, we employed

ApEn

to analyze three groups’ pulse morphology variability (PMV). The pulse waveform’s

ApEn

of a patient with cardiovascular disease tends to have a smaller value and its variation’s spectral contents differ greatly during different cardiovascular conditions. Then, we applied a SVM to discriminate cardiovascular disease patients from non-cardiovascular disease controls. The specificity and sensitivity for clinical diagnosis of cardiovascular system is 85% and 93% respectively. The proposed techniques in this paper, from a long-term PMV analysis viewpoint, can be applied to a further research on cardiovascular system.

The quality of the fingerprint images greatly affects the performance of the minutiae extraction. In order to improve the performance of the system, many researchers have been made efforts on the image enhancement algorithms. If the adaptive preprocessing according to the fingerprint image characteristics is applied in the image enhancement step, the system performance would be more robust. In this paper, we propose an adaptive preprocessing method, which extracts five features from the fingerprint images, analyzes image quality with Ward’s clustering algorithm, and enhances the images according to their characteristics. Experimental results indicate that the proposed method improves both the quality index and block directional difference significantly in a reasonable time.

The paper presents a content based image retrieval approach with adaptive and intelligent image classification through on-line model modification. It supports geographical image retrieval over digitized historical aerial photographs in a digital library. Since the historical aerial photographs are grayscaled and low-resolution images, image retrieval is achieved on the basis of texture feature extraction. Feature extraction methods for geographical image retrieval are Gabor spectral filtering, Laws’ energy filtering, and Wavelet transformation, which are all the most widely used in image classification and segmentation. Adaptive image classification supports effective content based image retrieval through composite classifier models dealing with multi-modal feature distribution. The image retrieval methods presented in the paper are evaluated over a test bed of 184 aerial photographs. The experimental results also show the performance of different feature extraction methods for each image retrieval method.

Support Vector Machines (SVMs) have been used successfully for many classification tasks. In this paper, we investigate applying SVMs to classification in the context of image processing. We chose to look at classifying whether pixels have been corrupted by impulsive noise, as this is one of the simpler classification tasks in image processing. We found that the straightforward application of SVMs to this problem led to a number of difficulties, such as long training times, performance that was sensitive to the balance of classes in the training data, and poor classification performance overall. We suggest remedies for some of these problems, including the use of image filters to suppress variation in the training data. This led us to develop a two-stage classification process which used SVMs in the second stage. This two-stage process was able to achieve substantially better results than those resulting from the straightforward application of SVMs.

This paper investigates that loose clothing such as wearing dresses and human body shapes create individual eigenspaces and, as a result, conventional appearance-based method cannot be effective for recognizing human body postures. We introduce a

dress effect

due to loose clothing and a

figure effect

due to various human body shapes in this particular study. This study particularly proposes an image pre-processing by ‘

Laplacian of Gaussian (LoG)

’ filter over input images and a ‘

mean posture matrix

’ for creating an eigenspace in order to overcome the preceding effects. We have tested the proposed approach on various dress environments and body shapes, and robustness of the method has been demonstrated.

This paper is about tracking people in real-time as they move through the non-overlapping fields of view of multiple video cameras. The paper builds upon existing methods for tracking moving objects in a single camera. The key extension is the use of a stochastic transition matrix to describe people’s observed patterns of motion both within and between fields of view. The parameters of the model for a particular environment are learnt simply by observing a person moving about in that environment. No knowledge of the environment or the configuration of the cameras is required.

In this paper, we propose a caption detection and removal technique for reuse of TV scenes using the Multilayer Perceptrons (MLPs) and Genetic algorithms (GAs). The technique first detects the captions in a TV scene using the MLP-based caption detector, and then removes the detected captions using the GA-based region remover. In our technique, the caption removal problem is modeled as an optimization problem, which in our case, is solved by a cost function with isophote constraint that is minimized using a GA. The technique creates an optimal connection of all pairs of isophote disconnected by caption regions. To connect the disconnected isophote, we estimate the value of the smoothness, given by the best chromosome of the GA, and project this value in the isophote direction. Experimental results show a great possibility for automatic removal of captions in TV advertisement scenes.

The particle filter has attracted considerable attention in visual tracking due to its relaxation of the linear and Gaussian restrictions in the state space model. It is thus more flexible than the Kalman filter. However, the conventional particle filter uses system transition as the proposal distribution, leading to poor sampling efficiency and poor performance in visual tracking. It is not a trivial task to design satisfactory proposal distributions for the particle filter. In this paper, we introduce an improved particle filtering framework into visual tracking, which combines the unscented Kalman filter and the auxiliary particle filter. The efficient unscented auxiliary particle filter (UAPF) uses the unscented transformation to predict one-step ahead likelihood and produces more reasonable proposal distributions, thus reducing the number of particles required and substantially improving the tracking performance. Experiments on real video sequences demonstrate that the UAPF is computationally efficient and outperforms the conventional particle filter and the auxiliary particle filter.

This paper demonstrates a novel subspace projection technique via Non-Negative Matrix Factorization (NMF) to represent human facial image in low frequency subband, which is able to realize through the wavelet transform. Wavelet transform (WT), is used to reduce the noise and produce a representation in the low frequency domain, and hence making the facial images insensitive to facial expression and small occlusion. After wavelet decomposition, NMF is performed to produce region or part-based representations of the images. Non-negativity is a useful constraint to generate expressiveness in the reconstruction of faces. The simulation results on Essex and ORL database show that the hybrid of NMF and the best wavelet filter will yield better verification rate and shorter training time. The optimum results of 98.5% and 95.5% are obtained from Essex and ORL Database, respectively. These results are compared with our baseline method, Principal Component Analysis (PCA).

Populations of biased, non-random sequences may cause standard alignment algorithms to yield false-positive matches and false-negative misses. A standard significance test based on the

shuffling

of sequences is a partial solution, applicable to populations that can be described by

simple

models. Masking-out low information content intervals throws information away. We describe a new and general method,

modelling-alignment

: Population models are incorporated into the alignment process, which can (and should) lead to changes in the rank-order of matches between a query sequence and a collection of sequences, compared to results from standard algorithms. The new method is general and places very few conditions on the nature of the models that can be used with it. We apply modelling-alignment to local alignment, global alignment, optimal alignment, and the relatedness problem.

This paper reports a novel technique to classify short duration articulated object motion in video data. The motion is represented by a spatio-temporal template (STT), a view based approach, which collapses temporal component into a static grey scale image in a way that no explicit sequence matching or temporal analysis is needed, and characterizes the motion from a very high dimensional space to a low dimensional space. These templates are modified to be invariant to translation and scale. Two dimensional, 3 level dyadic wavelet transform applied on these templates results in one low pass subimage and nine highpass directional subimages. Histograms of STTs and of the wavelet coefficients at different scales are compared to establish significance of available information for classification. To further reduce the feature space, histograms of STTs are represented by orthogonal Legendre moments, and the wavelet subbands are modelled by generalized Gaussian density (GGD) parameters – shape factor and standard deviation. The preliminary experiments show that directional information in wavelet subbands improves the histogram-based technique, and that use of moments combined with GGD parameters improves the performance efficiency in addition to significantly reducing complexity of comparing directly the histograms.

This paper presents a novel two-factor authenticator which hashes tokenized random data and moment based palmprint features to produce a set of private binary string, coined as Discrete-Hashing code. This novel technique requires two factors (random number + authorized biometrics) credentials in order to access the authentication system. Absence of either factor will just handicap the progress of authentication. Besides that, Discrete-Hashing also possesses high discriminatory power, with highly correlated bit strings for intra-class data. Experimental results show that this two-factor authenticator surpasses the classic biometric authenticator in terms of verification rate. Our proposed approach provides a clear separation between genuine and imposter population distributions. This implies that Discrete-Hashing technique allows achievement of zero False Accept Rate (FAR) without jeopardizing the False Reject Rate (FRR) performance, which is hardly possible to conventional biometric systems.

This paper addresses the accurate gaze detection method by tracking facial and eye movement at the same time. For that, we implemented our gaze detection system with a wide and a narrow view stereo camera. In order to make it easier to detect the facial and eye feature positions, the dual IR-LED illuminators are also used for our system. The performance of detecting facial features could be enhanced by Support Vector Machine and the eye gaze position on a monitor is computed by a multi-layered perceptron. Experimental results show that the RMS error of gaze detection is about 2.4 degrees (1.68 degrees on X axis and 1.71 degrees on Y axis at the Z distance of 50 cm).

Recent research in ontologies and descriptions logics has focused on compromising between expressiveness and reasoning ability, with many other issues being neglected. One major issue that has been neglected is how one should go about in modifying ontologies as inconsistency arises. The central concern of this problem is therefore to determine the most rational way of modifying ontologies, such that no extra knowledge would be retained in or retracted from the knowledge base. The purpose of this paper is to outline the complexities in this and to present some insights into the problem of ontology modification. Description logic (DL) is used in this paper as the underlying logic for the representation of ontology, and ontology modification is performed based on this logic.

In this paper we explore the issue of using some aspects of the Conceptual Graph Theory formalism to define functions on ontologies. We exploit the formal definitions of type hierarchy and projection to define operations on an ontology, and then illustrate these ideas by demonstrating a knowledge base of historical interactions that was implemented on an ontology defined in this way.

In this paper, we propose a dynamic allocation method of basis functions, an Allocation/Elimination Gaussian Softmax Basis Function Network (AE-GSBFN), that is used in reinforcement learning. AE-GSBFN is a kind of actor-critic method that uses basis functions. This method can treat continuous high-dimensional state spaces, because basis functions required only for learning are dynamically allocated, and if an allocated basis function is identified as redundant, the function is eliminated. This method overcomes the curse of dimensionality and avoids a fall into local minima through the allocation and elimination processes. To confirm the effectiveness of our method, we used a maze task to compare our method with an existing method, which has only an allocation process. Moreover, as learning of continuous high-dimensional state spaces, our method was applied to motion control of a humanoid robot. We demonstrate that the AE-GSBFN is capable of providing better performance than the existing method.

In many applications of machine learning a series of feature extraction approaches and a series of classifiers are explored in order to obtain a system with the highest accuracy possible. In the application we discussed here at least two feature extraction approaches have been explored (Fast Fourier Transform) and Discrete Wavelet Transform, and at least two approaches to build classifiers have also been explored (Artificial Neural Networks and Support Vector Machines). If one combination seems superior in terms of accuracy rate, shall we adopt it as the one to use or is there a combination of the approaches that results in some benefit? We show here how we have combined classifiers considering the misclassification cost to obtain a more informative classification for its application in the field.

Landmarking is a recent and promising meta-learning strategy, which defines meta-features that are themselves efficient learning algorithms. However, the choice of landmarkers is often made in an ad hoc manner. In this paper, we propose a new perspective and set of criteria for landmarkers. Based on the new criteria, we propose a landmarker generation algorithm, which generates a set of landmarkers that are each subsets of the algorithms being landmarked. Our experiments show that the landmarkers formed, when used with linear regression are able to estimate the accuracy of a set of candidate algorithms well, while only utilising a small fraction of the computational cost required to evaluate those candidate algorithms via ten-fold cross-validation.

The algorithm selection problem aims to select the best algorithm for an input problem instance according to some characteristics of the instance. This paper presents a learning-based inductive approach to build a predictive algorithm selection system from empirical algorithm performance data of the Most Probable Explanation(MPE) problem. The learned model can serve as an algorithm selection meta-reasoner for the real-time MPE problem. Experimental results show that the learned algorithm selection models can help integrate multiple MPE algorithms to gain a better overall performance of reasoning.

In the field of cluster analysis, objective function based clustering algorithm is one of widely applied methods so far. However, this type of algorithms need the priori knowledge about the cluster number and the type of clustering prototypes, and can only process data sets with the same type of prototypes. Moreover, these algorithms are very sensitive to the initialization and easy to get trap into local optima. To this end, this paper presents a novel clustering method with fuzzy network structure based on limited resource to realize the automation of cluster analysis without priori information. Since the new algorithm introduce fuzzy artificial recognition ball, operation efficiency is greatly improved. By analyzing the neurons of network with minimal spanning tree, one can easily get the cluster number and related classification information. The test results with various data sets illustrate that the novel algorithm achieves much more effective performance on cluster analyzing the large data set with mixed numeric values and categorical values.

To modeling and classify underwater sound, hidden Markov tree (HMT) model in wavelet domain is adopted. Taking advantage of the models, the simulation time sequence of ocean noise can be produced. An improved classification approach based on HMT model and fuzzy maximum and minimum neural net work (FMMNN) is brought forward, which integrates the wavelet coefficients HMT models with FMMNN. The performance of this approach is evaluated experimentally in classifying four types of ocean noises. With an accuracy of more than 86%, this HMT-based approach is found to outperform previously proposed classifiers. Experiments prove that the new method is effective.

Nowadays, automated Web document classification is considered as an important method to manage and process an enormous amount of Web documents in digital forms that are extensive and constantly increasing. Recently, document classification has been addressed with various classified techniques such as naïve Bayesian, TFIDF (Term Frequency Inverse Document Frequency), FCA (Formal Concept Analysis) and MCRDR (Multiple Classification Ripple Down Rules). We suggest the BayesTH-MCRDR algorithm for useful new Web document classification in this paper. We offer a composite algorithm that combines a naïve Bayesian algorithm using Threshold and the MCRDR algorithm. The prominent feature of the BayesTH-MCRDR algorithm is optimisation of the initial relationship between keywords before final assignment to a category in order to get higher document classification accuracy. We also present the system we have developed in order to demonstrate and compare a number of classification techniques.

This paper presents an improvement ant colony optimization algorithm for mining classification rule called ACO-Miner. The goal of ACO-Miner is to effectively provide intelligible classification rules which have higher predictive accuracy and simpler rule list based on Ant-Miner. Experiments on data sets from UCI data set repository were made to compare the performance of ACO-Miner with Ant-Miner. The results show that ACO-Miner performs better than Ant-Miner with respect to predictive accuracy and rule list mined simplicity.

This paper presents a single scan algorithm for clustering large datasets based on a two phase process which combines two well known clustering methods. The Cobweb algorithm is modified to produce a balanced tree with subclusters at the leaves, and then K-means is applied to the resulting subclusters. The resulting method, Scalable Cobweb, is then compared to a single pass K-means algorithm and standard K-means. The evaluation looks at error as measured by the sum of squared error and vulnerability to the order in which data points are processed.

How to minimize misclassification errors has been the main focus of Inductive learning techniques, such as CART and C4.5. However, misclassification error is not the only error in classification problem. Recently, researchers have begun to consider both test and misclassification costs. Previous works assume the test cost and the misclassification cost must be defined on the same cost scale. However, sometimes we may meet difficulty to define the multiple costs on the same cost scale. In this paper, we address the problem by building a cost-sensitive decision tree by involving two kinds of cost scales, that minimizes the one kind of cost and control the other in a given specific budget. Our work will be useful for many diagnostic tasks involving target cost minimization and resource consumption for obtaining missing information.

As discovering association rules in a very large database is time consuming, researchers have developed many algorithms to improve the efficiency. Sampling can significantly reduce the cost of mining, since the mining algorithms need to deal with only a small dataset compared to the original database. Especially, if data comes as a stream flowing at a faster rate than can be processed, sampling seems to be the only choice. How to sample the data and how big the sample size should be for a given error bound and confidence level are key issues for particular data mining tasks. In this paper, we derive the sufficient sample size based on central limit theorem for sampling large datasets with replacement. This approach requires smaller sample size than that based on the Chernoff bounds and is effective for association rules mining. The effectiveness of the method has been evaluated on both dense and sparse datasets.

Structural break is one of the important concerns in non-stationary time series prediction. The cumulative sum of square (CUSUMS) statistic proposed by Brown et al (1975) has been developed as a general method for detecting a structural break. To better utilize this method, this paper analyses the operating conditions of the centered version of CUSUMS using three variables: the percentage of variance change, the post-break data size and the pre-break data size. In traditional approach of the centered CUSUMS, all available data are used for the break detection. Our analysis reveals that one can improve the accuracy of the break detection by either reducing the post-break data size or increasing pre-break data size. Based on our analysis, we propose a modified test statistic. The evidence shows that the modified statistic significantly improves the chance of detecting the structural breaks.

While knowledge discovery in databases techniques require statistically complete data, real world data is incomplete, so it must be preprocessed using completeness approximation methods. The success of these methods is impacted by whether redundancy in large amounts of data overcomes incompleteness mechanisms. We investigate this impact by comparing rule sets induced from complete data with rule sets induced from incomplete data that is preprocessed using complete case analysis. To control the incomplete data construction, we apply the well-defined incompleteness mechanisms missing-at-random and missing-completely-at-random to complete data. Initial results indicate that a medium level of pattern redundancy fails to fully overcome incompleteness mechanisms, and that characterizing an appropriate redundancy threshold is non-trivial.

In the self organising map (SOM), applying different learning parameters to the same input will lead to different maps. The question of how to select the best map is important. A map is good if it is relatively accurate in representing the input and ordered. A measure or measures are needed to quantify the accuracy and the ‘order’ of maps. This paper investigates the learning parameters in standard 2- dimensional SOMs to find the learning parameters that lead to optimal arrangements. An example of choosing a map in a real world application is also provided.

In this paper, we describe KES, a system that integrates text categorisation and information extraction in order to extract key elements of information from particular types of documents, with these informational elements being presented in such a way as to provide a concise summary of the input document. We describe the overall architecture of the system and its components, with a particular focus on the problems involved in handling the names of companies and individuals in this domain.

This paper describes the use of taxonomic hierarchies of conceptclasses (dependent class values) for knowledge discovery. The approach allows evidence to accumulate for rules at different levels of generality and avoids the need for domain experts to predetermine which levels of concepts should be learned. In particular, higher-level rules can be learned automatically when the data doesn’t support more specific learning, and higher level rules can be used to predict a particular case when the data is not detailed enough for a more specific rule. The process introduces difficulties concerning how to heuristically select rules during the learning process, since accuracy alone is not adequate. This paper explains the algorithm for using concept-class taxonomies, as well as techniques for incorporating granularity (together with accuracy) in the heuristic selection process. Empirical results on three data sets are summarized to highlight the tradeoff between predictive accuracy and predictive granularity.

One problem many Web users encounter is to keep track of changes of distant Web sources.

Push services

, informing clients about data changes, are frequently not provided by Web servers. Therefore it is necessary to apply intelligent

pull strategies

, optimizing reload requests by observation of data sources. In this article an adaptive pull strategy is presented that optimizes reload requests with respect to the ‘age’ of data and lost data. The method is applicable if the remote change pattern may approximately be described by a piecewise deterministic behavior which is frequently the case if data sources are updated automatically. Emphasis is laid on an autonomous estimation where only a minimal number of parameters has to be provided manually.

Maximal frequent itemsets mining is one of the most fundamental problems in data mining. In this paper, we present CfpMfi, a new depth-first search algorithm based on CFP-tree for mining MFI. Based on the new data structure CFP-tree, which is a combination of FP-tree and MFI-tree, CfpMfi takes a variety pruning techniques and a novel item ordering policy to reduce the search space efficiently. Experimental comparison with previous work reveals that, on dense datasets, CfpMfi prunes the search space efficiently and is better than other MFI Mining algorithms on dense datasets, and uses less main memory than similar algorithm.

This paper presents empirical results for several versions of the multinomial naive Bayes classifier on four text categorization problems, and a way of improving it using locally weighted learning. More specifically, it compares standard multinomial naive Bayes to the recently proposed transformed weight-normalized complement naive Bayes classifier (TWCNB) [1], and shows that some of the modifications included in TWCNB may not be necessary to achieve optimum performance on some datasets. However, it does show that TFIDF conversion and document length normalization are important. It also shows that support vector machines can, in fact, sometimes very significantly outperform both methods. Finally, it shows how the performance of multinomial naive Bayes can be improved using locally weighted learning. However, the overall conclusion of our paper is that support vector machines are still the method of choice if the aim is to maximize accuracy.

This paper presents some empirical results showing that simple attribute scaling in the data preprocessing stage can improve the performance of linear binary classifiers. In particular, a class specific scaling method that utilises information about the class distribution of the training sample can significantly improve classification accuracy. This form of scaling can boost the performance of a simple centroid classifier to similar levels of accuracy as the more complex, and computationally expensive, support vector machine and regression classifiers. Further, when SVMs are used, scaled data produces better results, for smaller amounts of training data, and with smaller regularisation constant values, than unscaled data.

This paper proposes and evaluates a nearest-neighbor method to sub-stitute missing values in ordinal/continuous datasets. In a nutshell, the K-Means clustering algorithm is applied in the complete dataset (without missing values) before the imputation process by nearest-neighbors takes place. Then, the achieved cluster centroids are employed as training instances for the nearest-neighbor method. The proposed method is more efficient than the

traditional

nearest-neighbor method, and simulations performed in three benchmark data-sets also indicate that it provides suitable imputations, both in terms of prediction and classification tasks.

This paper introduces two new approaches to fit a linear regression model on interval-valued data. Each example of the learning set is described by a feature vector where each feature value is an interval. In the first proposed approach, it is fitted two independent linear regression models, respectively, on the mid-point and range of the interval values assumed by the variables on the learning set. In the second approach, is fitted a multivariate linear regression models on these mid-point and range. The prediction of the lower and upper bound of the interval value of the dependent variable is accomplished from its mid-point and range which are estimated from the fitted linear regression models applied to the mid-point and range of each interval values of the independent variables. The evaluation of the proposed prediction methods is based on the average behavior of the

root mean squared error

and the

determination coefficient

in the framework of a Monte Carlo experiment in comparison with the method proposed by Billard and Diday [2].

Association rule mining is a data mining technique that reveals interesting relationships in a database. Existing approaches employ different parameters to search for interesting rules. This fact and the large number of rules make it difficult to compare the output of confidence-based association rule miners. This paper explores the use of classification performance as a metric for evaluating their output. Previous work on forming classifiers from association rules has focussed on accurate classification, whereas we concentrate on using the properties of the resulting classifiers as a basis for comparing confidence-based association rule learners. Therefore, we present experimental results on 12 UCI datasets showing that the quality of small rule sets generated by Apriori can be improved by using the predictive Apriori algorithm. We also show that CBA, the standard method for classification using association rules, is generally inferior to standard rule learners concerning both running time and size of rule sets.

Analysis of queries posed to open-domain question-answering systems indicates that particular types of queries are dominant, e.g., queries about the identity of people, and about the location or time of events. We applied a rule-based mechanism and performed manual classification to classify queries into such commonly occurring types. We then experimented with different adjustments to our basic document retrieval process for each query type. The application of the best retrieval adjustment for each query type yielded improvements in retrieval performance. Finally, we applied a machine learning technique to automatically learn the manually classified query types, and applied the best retrieval adjustments obtained for the manual classification to the automatically learned query classes. The learning algorithm exhibited high accuracy, and the retrieval performance obtained for the learned classes was consistent with the performance obtained for the rule-based and manual classifications.

In this paper, we propose an automatic word-spacing method for a Korean text preprocessing system in resolving the problem of context-dependent word-spacing. The current method combines the stochastic-based method and partial parsing. First, the stochastic method splits an input sentence into a candidate-word sequence using word unigrams and syllable bigrams. Second, the system engages a partial parsing module based on the asymmetric relation between the candidate-words. The partial parsing module manages the governing relationship using words which are incorporated into the knowledge base as having a high-probability of spacing-error words. These elements serve as parsing trigger points based on their linguistic information, and they deter-mine the parsing direction as well as the parsing scope. Combining the stochastic- and linguistic-based methods, the current automatic word-spacing system becomes robust against the problem of context-dependant word-spacing. An average 8.98% amelioration of the total error rate is obtained for inner and external data.

This paper presents a novel approach to extracting shallow language patterns from text. The approach makes use of an attributed string matching technique which is based on two major but complementary factors:

lexical similarities and sentence structures

. The technique takes full advantage of a huge number of sentence patterns in a Treebank, while preserving robustness, with-out being bogged down into a complete linguistic analysis. The ideas described are implemented and an evaluation of 5,000 Chinese sentences is examined in order to justify its statistical significances.

The interpretation of complex discourse, such as arguments, is a difficult task that often requires validation, i.e., a system may need to present its interpretation of a user’s discourse for confirmation. In this paper, we consider the presentation of discourse interpretations in the context of a probabilistic argumentation system.We first describe our initial approach to the presentation of an interpretation of a user’s argument; this interpretation takes the form of a Bayesian subnet.We then report on the results of our preliminary evaluation with users, focusing on their criticisms of our system’s output. These criticisms motivate a content enhancement procedure that adds information to explain unexpected outcomes and removes superfluous content from an interpretation. The discourse generated by this procedure was found to be more acceptable than the discourse generated by our original method.

Speech recognition technique is expected to make a great impact on many user interface areas such as toys, mobile phones, PDAs, and home appliances. Those applications basically require robust speech recognition immune to environment and channel noises, but the dialogue pattern used in the interaction with the devices may be relatively simple, that is, an isolated-word type. The drawback of small-vocabulary isolated-word recognizer which is generally used in the applications is that, if target vocabulary needs to be changed, acoustic models should be retrained for high performance. However, if a phone model-based speech recognition is used with reliable unseen model prediction, we do not need to re-train acoustic models in getting higher performance. In this paper, we propose a few reliable methods for unseen model prediction in flexible vocabulary speech recognition. The first method gives optimal threshold values for stop criteria in decision tree growing, and the second uses an additional condition in the question selection in order to overcome the over-balancing phenomenon in the conventional method. The last proposes two-stage decision trees which in the first stage get more properly trained models and in the second stage build more reliable unseen models. Various vocabulary-independent situations were examined in order to clearly show the effectiveness of the proposed methods. In the experiments, the average word error rates of the proposed methods were reduced by 32.8%, 41.4%, and 44.1% compared to the conventional method, respectively. From the results, we can conclude that the proposed methods are very effective in the unseen model prediction for vocabulary- independent speech recognition.

In this paper, we propose a voice code verification method for an intelligent surveillance guard robot, wherein a robot prompts for a code (i.e. word or phrase) for verification. In the application scenario, the voice code can be changed every day for security reasoning and the targeting domain is unlimited. Thus, the voice code verification system not only requires the text-prompted and speaker independent verification, but also it should not require an extra trained model as an alternative hypothesis for log-likelihood ratio test because of memory limitation. To resolve these issues, we propose to exploit the sub-word based anti-models for log-likelihood normalization through reusing an acoustic model and competing with voice code model. The anti-model is automatically produced by using the statistical distance of phonemes against a voice code. In addition, a harmonics-based spectral subtraction algorithm is applied for a noisy robust system on an outdoor environment. The performance evaluation is done by using a common Korean database, PBW452DB, which consists of 63,280 utterances of 452 isolated words recorded in silent environment.

The

PolicyUpdater

system is a fully-implemented access control system that provides policy evaluations as well as dynamic policy updates. These functions are achieved by the use of a logic-based language

$\mathcal{L}$

to represent the underlying access control policies, constraints and update propositions. The system performs authorisation query evaluations and conditional policy updates by translating the language

$\mathcal{L}$

to a normal logic program in a form suitable for evaluation using the

Stable Model

semantics.

In this paper, we propose a new neighborhood structure based on the improvement graph for solving the Robust Graph Coloring Problem, an interesting extension of classical graph coloring. Different from the traditional neighborhood where the color of only one vertex is modified, the new neighborhood involves several vertices. In addition, the questions of how to select the modified vertices and how to modify them are modelled by an improvement graph and solved by a Dynamic Programming method. The experimental results clearly show that our new improvement graph based

k

-exchange cycle neighborhood improves the accuracy significantly, especially for large scale heuristic search.

Hex is a classic board game invented in the middle of the twentieth century by Piet Hein and rediscovered later by John Nash. The best Hex artificial players analyse the board positions by deducing complex virtual connections from elementary connections using the H-Search algorithm. In this paper, we extend the H-search with a new deduction rule. This new deduction rule is capable of discovering virtual connections that the H-search cannot prove. Thanks to this new deduction rule, the horizon of the artificial Hex players should move further away.

Controlling an autonomous camera in three-dimensional virtual environments is a difficult task which manifests itself in many interactive computer graphics applications, such as computer games. In this paper, we represent this problem as a constraint satisfaction problem which is often over-constrained. A range of complex requirements, such as frame coherence, occlusion and camera holes can be elegantly represented as constraints. We then apply both complete and incomplete search methods to find the optimal camera placement. An interactive computer games application was developed to experimentally evaluate these methods. Our experimental results and a discussion with related studies conclude that our approach is sophisticated both in modelling and solving the difficult task of 3D camera control.

We re-visit the problem of converting action specifications into system constraints by re-examining it in the very simple setting of STRIPS. This has the merit of making many thorny issues relatively transparent. The paper is in the form of an extended summary of ongoing work and many of the results are merely outlined. But sufficient details are included to indicate where we will be taking things further, and to encourage others to pursue the same objectives. These objectives are in some sense a kind of reverse-engineering, as the database community is evidently familiar with the idea of starting with constraints and then deriving action specifications from them. However in AI reactive systems, action specifications are often the primary entity, so techniques to unearth the implicit constraints can facilitate better designs.

Quantified Boolean formulas (QBFs) play an important role in artificial intelligence subjects, specially in planning, knowledge representation and reasoning [20]. In this paper we present ZQSAT (sibling of our FZQSAT [15]), which is an algorithm for evaluating quantified Boolean formulas. QBF is a language that extends propositional logic in such a way that many advanced forms of reasoning can be easily formulated and evaluated. ZQSAT is based on ZDD, which is a variant of BDD, and an adopted version of the DPLL algorithm. The program has been implemented in C using the CUDD package. The capability of ZDDs in storing sets of subsets efficiently enabled us to store the clauses of a QBF very compactly and led us to implement the search algorithm in such a way that we could store and reuse the results of all previously solved subformulas with few overheads. This idea along some other techniques, enabled ZQSAT to solve some standard QBF benchmark problems faster than the best existing QSAT solvers.

The paper proposes an approach for selecting partners in multi-agent negotiation with the use of possibilistic case-based reasoning. It predicts the possibility of successful negotiation with other agents based on their past negotiation behaviour and the derived qualitative expected utility for the current situation. The proposed approach allows the agents to select their most prospective negotiation partners based on a small sample of historical cases of previous negotiations even if they are different from the current situation. Partner selection models for both independent and correlated negotiation agents are detailed and demonstrated with simple scenarios.

Most propagation-based set constraint solvers approximate the set of possible sets that a variable can take by upper and lower bounds, and perform so-called set bounds propagation. However Lagoon and Stuckey have shown that using reduced ordered binary decision diagrams (ROBDDs) one can create a practical set domain propagator that keeps all information (possibly exponential in size) about the set of possible set values for a set variable. In this paper we first show that we can use the same ROBDD approach to build an efficient bounds propagator. The main advantage of this approach to set bounds propagation is that we need not laboriously determine set bounds propagations rules for each new constraint, they can be computed automatically. In addition we can eliminate intermediate variables, and build stronger set bounds propagators than with the usual approach. We then show how we can combine this with the set domain propagation approach of Lagoon and Stuckey to create a more efficient set domain propagation solver.

This paper describes methods to enable efficient use and administration of a CBR system for teledermatology in which the users are assumed to be non-computer literate. In particular, a user-friendly interface to enable a general practitioner to decide a diagnosis with the minimum number of questions asked is proposed. Specifically, we propose a technique to improve the usefulness of a decision tree approach in terms of general rather than specific questions. Additionally, we describe a new technique to minimise the number of questions presented to the user in the query process for training the CBR system. Importantly we apply FCA technique to enable the consultant dermatologist to validate new knowledge and supervised the incremental learning of the CBR system.

This paper presents a comparative study between a state-of-the-art clause weighting local search method for satisfiability testing and a variant modified to obtain longer-term memory from a global measure of clause perturbation. We present empirical evidence indicating that by learning which clauses are hardest to satisfy, the modified method can offer significant performance improvements for a well-known range of satisfiable problems. We conclude that our method’s ability to learn, and consequently to offer performance improvement, can be attributed to its ability to obtain information from a global measure of hardness, rather than from the contextual perspective exploited in previous works.

The work described in this paper presents a goal oriented navigation system in a behavior-based manner. The main contributions are, in first place the in-depth study of local navigation strategies and, on the other hand, the use of natural landmarks, namely corridors and emergency exit panels. Eliminating the centralized control of modules the system performs the task as a result of the combination of many relatively simple and light weight behaviors that run concurrently.

This paper proposes a novel approach for modeling partially connected feedforward neural networks (PCFNNs) by identifying input type which refers to whether an input is coupled or uncoupled with other inputs. The identification of input type is done by analyzing input sensitivity changes by varying the magnitude of input. In the PCFNNs, each input is linked to the neurons in the hidden layer in a different way according to its input type. Each uncoupled input does not share the neurons with other inputs in order to contribute to output in an independent manner. The simulation results show that PCFNNs outperform fully connected feedforward neural networks with simple network structure.

In this paper, we consider an anisotropic swarm model with an attraction/repulsion function and study its aggregation properties. It is shown that the swarm members will aggregate and eventually form a cohesive cluster of finite size around the swarm center. We also study the swarm cohesiveness when the motion of each agent is a combination of the inter-individual interactions and the interaction of the agent with external environment. Moreover, we extend our results to more general attraction/repulsion functions. The model in this paper is more general than isotropic swarms and our results provide further insight into the effect of the interaction pattern on individual motion in a swarm system.

We describe an Ant Colony Optimization (ACO) algorithm, ANT-MPE, for the most probable explanation problem in Bayesian network inference. After tuning its parameters settings, we compare ANT-MPE with four other sampling and local search-based approximate algorithms: Gibbs Sampling, Forward Sampling, Multistart Hillclimbing, and Tabu Search. Experimental results on both artificial and real networks show that in general ANT-MPE outperforms all other algorithms, but on networks with unskewed distributions local search algorithms are slightly better. The result reveals the nature of ACO as a combination of both sampling and local search. It helps us to understand ACO better, and, more important, it also suggests a possible way to improve ACO.

Traditional approaches to evolvable hardware (EHW), using a direct encoding, have not scaled well with increases in problem complexity. To overcome this there have been moves towards encoding a growth process, which however have not shown a great deal of success to date. In this paper we present the design of a morphogenetic EHW model that has taken the salient features of biological processes and structures to produce an evolutionary and growth model that consistently outperforms a traditional EHW approach using a direct encoding, and scales well to larger, more complex, problems.

Evolutionary computation techniques have been used widely to solve various optimization and learning problems. This paper describes the application of evolutionary computation techniques to a real world complex train schedule multiobjective problem. Three established algorithms (Genetic Algorithm GA, Particle Swarm Optimization PSO, and Differential Evolution DE) were proposed to solve the scheduling problem. Comparative studies were done on various performance indices. Simulation results are presented which demonstrates that DE is the best approach for this scheduling problem.

In fuzzy model, the consequent of fuzzy rule is often determined with degrees of belief or credibility because of vague information originating from evidence not strong enough and “lack of specificity”. In this paper, we present a fuzzy model incorporated with fuzzy Dempster-Shafer Theory. The consequent of fuzzy rule is not fuzzy propositions, but fuzzy Dempster-Shafer granules. The salient aspect of the work is that a very simplified analytic output of fuzzy model which is a special case of Sugeno-type fuzzy model is achieved when all fuzzy sets in fuzzy partition of the output space have the same power (the area under the membership function), and the determination of basic probability assignments associated with fuzzy Dempster-Shafer belief structure using fuzzy Naive Bayes. The construction method of fuzzy Naive Bayes and an learning strategy generating fuzzy rules from training data are proposed in this paper. A well-known example about time series prediction is tested, the prediction results show that our fuzzy modeling is very efficient and has strong expressive power to represent the complex system with uncertain situation.

The K-means Fast Learning Artificial Neural Network (KFLANN) is a small neural network bearing two types of parameters, the tolerance,

δ

and the vigilance,

μ

. In previous papers, it was shown that the KFLANN was capable of fast and accurate assimilation of data [12]. However, it was still an unsolved issue to determine the suitable values for

δ

and

μ

in [12]. This paper continues to follows-up by introducing Genetic Algorithms as a possible solution for searching through the parameter space to effectively and efficiently extract suitable values to

δ

and

μ

. It is also able to determine significant factors that help achieve accurate clustering. Experimental results are presented to illustrate the hybrid GA-KFLANN ability using available test data.

Based on the Antibody Clonal Selection Theory of immunology, the general steps of ICSA (Immune Clonal Selection Algorithm) are presented in this paper. The network framework of ICSA is put forward, and the dynamic characters of ICSA are analyzed based on the Lyapunov theory. Then, this paper gives a novel Artificial Immune System Algorithm, Pseudo- Grads Hybrid Immune Clonal Selection Network (GHICSN). The simulation results of some functions optimization indicate that GHICSN improves the performance of ICSA to some extent.

This paper proposes an enhanced RBF network that enhances learning algorithms between input layer and middle layer and between middle layer and output layer individually for improving the efficiency of learning. The proposed network applies ART2 network as the learning structure between input layer and middle layer. And the auto-tuning method of learning rate and momentum is proposed and applied to learning between middle layer and output layer, which arbitrates learning rate and momentum dynamically by using the fuzzy control system for the arbitration of the connected weight between middle layer and output layer. The experiment for the classification of number patterns extracted from the citizen registration card shows that compared with conventional networks such as delta-bar-delta algorithm and the ART2-based RBF network, the proposed method achieves the improvement of performance in terms of learning speed and convergence.

This paper demonstrates that the self-adaptive technique of Differential Evolution (DE) can be simply used for solving a multi-objective optimization problem where parameters are interdependent. The real-coded crossover and mutation rates within the NSGA-II have been replaced with a simple Differential Evolution scheme, and results are reported on a rotated problem which has presented difficulties using existing Multi-objective Genetic Algorithms. The Differential Evolution variant of the NSGA-II has demonstrated rotational invariance and superior performance over the NSGA-II on this problem.

Niching enables a

genetic algorithm

(GA) to maintain diversity in a population. It is particularly useful when the problem has multiple optima where the aim is to find all or as many as possible of these optima. When the fitness landscape of a problem changes overtime, the problem is called non–stationary, dynamic or time–variant problem. In these problems, niching can maintain useful solutions to respond quickly, reliably and accurately to a change in the environment. In this paper, we present a niching method that works on the problem substructures rather than the whole solution, therefore it has less space complexity than previously known niching mechanisms. We show that the method is responding accurately when environmental changes occur.

The aim of the current study is to assess the suitability of two Associative Memory (AM) models to character recognition problems. The two AM models under scrutiny are a One-Shot AM (OSAM) and an Exponential Correlation AM (ECAM). We compare these AMs on the resultant features of their architectures, including recurrence, learning and the generation of domains of attraction. We illustrate the impact of each of these features on the performance of each AM by varying the training set size, introducing noisy data and by globally transforming symbols. Our results show that each system is suited to different character recognition problems.

Loops are rarely used in genetic programming (GP), because they lead to massive computation due to the increase in the size of the search space. We have investigated the use of loops with restricted semantics for a problem in which there are natural repetitive elements, that of distinguishing two classes of images. Using our formulation, programs with loops were successfully evolved and performed much better than programs without loops. Our results suggest that loops can successfully used in genetic programming in situations where domain knowledge is available to provide some restrictions on loop semantics.

A negotiation agent exchanges proposals, supported by claims, with an opponent. Each proposal and claim exchanged reveals valuable information about the sender’s position. A negotiation may brake down if an agent believes that its opponent is not playing fairly. The agent aims to give the impression of fair play by responding with comparable information revelation whilst playing strategically to influence its opponent’s preferences with claims. It uses maximum entropy probabilistic reasoning to estimate unknown values inprobability distributions including the probability that its opponent will accept any deal.

We have built an agent service-based enterprise infrastructure: F-TRADE. With its online connectivity to huge real stock data in global markets, it can be used for online evaluation of trading strategies and data mining algorithms. The main functions in the F-TRADE include soft plug-and-play, and back-testing, optimization, integration and evaluation of algorithms. In this paper, we’ll focus on introducing the intelligent plug-and-play, which is a key system function in the F-TRADE. The basic idea for the soft plug-and-play is to build agent services which can support the online plug-in of agents, algorithms and data sources. Agent UML-based modeling, role model and agent services for the plug-and-play are discussed. With this design, algorithm providers, data source providers, and system module developers of the F-TRADE can expand system functions and resources by online plugging them into the F-TRADE.

Distributed blackboard is one of the popular agent communication architectures, where the blackboards location and communication topology are two important issues. However, there are few works about the issues. To solve this problem, this paper presents a new method, which applies

Multi-Medians Location Method

to compute the sub-blackboard locations, and applies

Steinner Tree Method

to compute the communication topology among sub-blackboards. The model can construct the distributed blackboard architecture effectively according to the current underlying network topology.

This paper proposes a multi-agent Q-learning algorithm called meta-game-Q learning that is developed from the meta-game equilibrium concept. Different from Nash equilibrium, meta-game equilibrium can achieve the optimal joint action game through deliberating its preference and predicting others’ policies in the general-sum game. A distributed negotiation algorithm is used to solve the meta-game equilibrium problem instead of using centralized linear programming algorithms. We use the repeated prisoner’s dilemma example to empirically demonstrate that the algorithm converges to meta-game equilibrium.

Robot navigation relies on a robust and real-time visual perception system to understand the surrounding environment. This paper describes a fast visual landmark search and recognition mechanism for real-time robotics applications. The mechanism models two stages of visual perception named preattentive and attentive stages. The pre-attentive stage provides a global guided search by identifying regions of interest, which is followed by the attentive stage for landmark recognition. The results show the mechanism validity and applicability to autonomous robot applications.

The paper presents a novel image (feature) interpolation method to reinforce the adaptive object recognition system. The system deals with texture images in a sequence according to changing perceptual conditions. When it recognizes several classes of objects under variable conditions, a fundamental problem is that two or more classes are overlapped on the feature space. This interpolation method is useful to resolve the overlapping issue.

A basic scheme for extracting digital image watermark is proposed using independent component analysis (ICA). The algorithm in terms of fastICA is discussed and used to separate the watermark from the mixed sources. The behavior of the proposed approach with several robustness tests of the image watermark is also carried out to demonstrate that ICA technique could provide a flexible and robust system for performing digital watermark detection and extraction. The preliminary experimental results show that the proposed watermarking method is effective and robust to some possible attacks.

Many current face recognition algorithms are based on face representations found by unsupervised statistical methods. One of the fundamental problems of face recognition is dimensionality reduction. Principal component analysis is a well-known linear method for reducing dimension. Recently, locally linear embedding (LLE) is proposed as an unsupervised procedure for mapping higher-dimensional data nonlinearly to a lower-dimensional space. This method, when combined with fisher linear discriminant models, is called extended LLE (ELLE) in this paper. Furthermore, the ELLE yields good classification results in the experiments. Also, we apply the Gabor wavelets as a pre-processing method which contributes a lot to the final results because it deals with the detailed signal of an image and is robust to light variation. Numerous experiments on ORL and AR face data sets have shown that our algorithm is more effective than the original LLE and is insensitive to light variation.

Automated image processing has great potential to aid in the classification of biological images. Many natural structures such as neurons exhibit fractal properties, and measures derived from fractal analysis are useful in differentiating neuron types. When fractal properties are not constant in all parts of the neuron, multifractal analysis may provide superior results. We applied three methods to elucidate the variation within 16 rat retinal ganglion cells: local connected fractal dimension (LCFD), mass-radius (MR) and maximum likelihood multifractal (MLM) analyses. The LCFD method suggested that some of the neurons studied are multifractal. The MR method was inconclusive due to the finite size of the cells. However, the MLM method was able to show the multifractal nature of all the samples, and to provide a superior multifractal spectrum. We conclude that the latter method warrants further attention as it may improve results in other application areas.

Edge detection is one of the fundamental operations incomputer vision with numerous approaches to it. In nowadays, many algorithms for edge detection have been proposed. However, most conventional techniques have assumed clear images or Gaussian noise images, thus their performance could decrease with the impulse noise. In this paper, we present an edge detection approach using Discrete Singular Convolution algorithm. The DSC algorithm efficiently detects edges not only original images but also noisy images which are added by Gaussian and impulse noise. Therefore, we evaluate that the performance of the DSC algorithm is compared with other algorithms such as the Canny, Bergholm, and Rothwell algorithm.

Nonparametric methods in statistics is introduced to analyze decision tables. First, the raw decision table is translated to a series of corresponding contingency tables between each condition attribute and the decision attribute, and then, dependence significance testing is finished to make sure if a condition attribute is correlated with the decision. Finally, we get the reduct of decision table at a special significance level, as well as all the first-order rules. Our experiments show that the nonparametric statistical method we proposed is feasible and efficiently.

This paper investigates how we can precisely define what process designers are ought achieve for what they have promised and more importantly in a way that satisfies human users. Toward these goals, an interaction model for processes and an Affect Monitoring Framework (AMF) are proposed based on our analysis on speech act theory and cognitive-based emotion models. The Affect Monitoring Framework is to detect and predict negative affects on users and to resolve caused or predicted causes of negative affects automatically.

We have proposed a new approach called

ontology services-driven integration of business intelligence

(BI) to designing an integrated BI platform. In such a BI platform, multiple ontological domains may get involved, such as domains for business, reporting, data warehouse, and multiple underlying enterprise information systems. In general, ontologies in the above multiple domains are heterogeneous. So, a key issue emerges in the process of building an integrated BI platform, that is, how to support ontology transformation and mapping between multiple ontological domains. In this paper, we present semantic aggregations of semantic relationships and ontologies in one or multiple domains, and the ontological transformation from one domain to another. Rules for the above semantic aggregation and transformation are described. This work is the foundation for supporting BI analyses crossing multiple domains.

How to assess the performance of machine learning algorithms is a problem of increasing interest and urgency as the data mining application of myriad algorithms grows. Rather than predictive accuracy, we propose the use of information-theoretic reward functions. The first such proposal was made by Kononenko and Bratko. Here we improve upon our alternative Bayesian metric, which provides a fair betting assessment of any machine learner. We include an empirical analysis of various Bayesian classification learners.

This paper gives an analysis of multi-class e-mail categorization performance, comparing a character

n

-gram document representation against a word-frequency based representation. Furthermore the impact of using available e-mail specific meta-information on classification performance is explored and the findings are presented.

Attributes reduction is a crucial problem in rough set application to data mining. In this paper, we introduce the Universal RED problem model, or

UniRED

, which transforms the discrete attributes reduction problems on Boolean space into continuous global optimization problems on real space. Based on this transformation, we develop a coordinate descent algorithm

RED2.1

for attributes reduction problems. In order to examine the efficiency of our algorithms, we conduct the comparison between our algorithm

RED2.1

and other reduction algorithms on some problems from UCI repository. The experimental results indicate the efficiency of our algorithm.

Recommender Systems seek to furnish personalized suggestions automatically based on user preferences. These preferences are usually expressed as a set of items either directly or indirectly given by the user (e.g., the set of products the user bought in a virtual store). In order to suggest new items, Recommender Systems generally use one of the following approaches:

Content Based Filtering, Collaborative Filtering or hybrid filtering methods

. In this paper we propose a strategy to improve the quality of recommendation in the first user contact with the system. Our approach includes a suitable plan to acquiring a user profile and a hybrid filtering method based on

Modal Symbolic Data

. Our proposed technique outperforms the

Modal Symbolic Content Based Filter

and the standard

kNN Collaborative Filter based on Pearson Correlation

.

Most databases employ the relational model for data storage. To use this data in a propositional learner, a propositionalization step has to take place. Similarly, the data has to be transformed to be amenable to a multi-instance learner. The Proper Toolbox contains an extended version of RELAGGS, the Multi-Instance Learning Kit MILK, and can also combine the multi-instance data with aggregated data from RELAGGS. RELAGGS was extended to handle arbitrarily nested relations and to work with both primary keys and indices. For MILK the relational model is flattened into a single table and this data is fed into a multi-instance learner. REMILK finally combines the aggregated data produced by RELAGGS and the multi-instance data, flattened for MILK, into a single table that is once again the input for a multi-instance learner. Several well-known datasets are used for experiments which highlight the strengths and weaknesses of the different approaches.

This paper proposes a new sigma point selection strategy to better capture the information of a probability distribution. By doing so, the non-local sampling problem inherent in the original unscented transformation (UT) is fundamentally eliminated. It is argued that the improved UT (IUT) outperforms the original UT at the cost of increased but comparable computation burden and will be useful in constructing a nonlinear filter.

A wrapper is a program which extracts data from a web site and reorganizes them in a database. Wrapper generation from web sites is a key technique in realizing such a metasearch system. We present a new method of automatic wrapper generation for metasearch using our efficient learning algorithm for

term trees

. Term trees are ordered tree structured patterns with structured variables, which represent structural features common to tree structured data such as HTML files.

The classification problem in data mining is to discover models from training data for classifying unknown instances. Associative classification builds the classifier rules using association rules and it is more accurate compared to previous methods. In this paper, a new method named CSFP that builds a classifier from strong frequent patterns without the need to generate association rules is presented. We address the rare item problem by using a partitioning method. Rules generated are stored using a compact data structure named

CP-Tree

and a series of pruning methods are employed to discard weak frequent patterns. Experimental results show that our classifier is more accurate than previous associative classification methods as well as other state-of-the-art non-associative classifiers.

Wavelet Domain Hidden Markov Model (WD HMM), in particular Hidden Markov Tree (HMT), has recently been proposed and applied to gray level image analysis. In this paper, color texture analysis using WD HMM is studied. In order to combine color and texture information to one single model, we extend WD HMM by grouping the wavelet coefficients from different color planes to one vector. The grouping way is chose according to a tradeoff between computation complexity and effectiveness. Besides, we propose Multivariate Gaussian Mixture Model (MGMM) to approximate the marginal distribution of wavelet coefficient vectors and to capture the interactions of different color planes. By employing our proposed approach, we can improve the performance of WD HMM on color texture classification. The experiment shows that our proposed WD HMM provides an 85% percentage of correct classifications (PCC) on 68 color images from an Oulu Texture Database and outperforms other methods.

We exploit the merits of C4.5 decision tree classifier with two stacking meta-learners: back-propagation multilayer perceptron neural network and naive-Bayes respectively. The performance of these two hybrid classification schemes have been empirically tested and compared with C4.5 decision tree using two US data sets (raw data set and new data set incorporated with domain knowledge) simultaneously to predict US bank failure. Significant improvements in prediction accuracy and training efficiency have been achieved in the schemes based on new data set. The empirical test results suggest that the proposed hybrid schemes perform marginally better in term of AUC criterion.

There are a great many metrics available for measuring the interestingness of rules. In this paper, we design a distinct approach for identifying association rules that maximizes the interestingness in an applied context. More specifically, the interestingness of association rules is defined as the dissimilarity between corresponding clusters. In addition, the interestingness assists in filtering out those rules that may be uninteresting in applications. Experiments show the effectiveness of our algorithm.

Classification is an important data mining problem. Emerging Patterns (EPs) are itemsets whose supports change significantly from one data class to another. Previous studies have shown that classifiers based on EPs are competitive to other state-of-the-art classification systems. In this paper, we propose a new type of Emerging Patterns, called Maximal Emerging Patterns (MaxEPs), which are the longest EPs satisfying certain constraints. MaxEPs can be used to condense the vast amount of information, resulting in a significantly smaller set of high quality patterns for classification. We also develop a new “overlapping” or “intersection” based mechanism to exploit the properties of MaxEPs. Our new classifier, Classification by Maximal Emerging Patterns (CMaxEP), combines the advantages of the Bayesian approach and EP-based classifiers. The experimental results on 36 benchmark datasets from the UCI machine learning repository demonstrate that our method has better overall classification accuracy in comparison to JEP-classifier, CBA, C5.0 and NB.

In this paper, we present a new approach to learning the classification of questions. Question classification received interest recently in the context of question answering systems for which categorizing a given question would be beneficial to allow improved processing of the document to identify an answer. Our approach relies on relative simple preprocessing of the question and uses standard decision tree learning. We also compared our results from decision tree learning with those obtained using Naïve Bayes. Both results compare favorably to several very recent studies using more sophisticated preprocessing and/or more sophisticated learning techniques. Furthermore, the fact that decision tree learning proved more successful than Naïve Bayes is significant in itself as decision tree learning is usually believed to be less suitable for NLP tasks.

This paper addresses the problem of mining exceptions from multidimensional databases. The goal of our proposed model is to find association rules that become weaker in some specific subsets of the database. The candidates for exceptions are generated combining previously discovered multidimensional association rules with a set of significant attributes specified by the user. The exceptions are mined only if the candidates do not achieve an expected support. We describe a method to estimate these expectations and propose an algorithm that finds exceptions. Experimental results are also presented.

We propose a multivariate decision tree inference scheme by using the minimum message length (MML) principle (Wallace and Boulton, 1968; Wallace and Dowe, 1999). The scheme uses MML coding as an objective (goodness-of-fit) function on model selection and searches with a simple evolution strategy. We test our multivariate tree inference scheme on UCI machine learning repository data sets and compare with the decision tree programs C4.5 and C5. The preliminary results show that on average and on most data-sets, MML oblique trees clearly perform better than both C4.5 and C5 on both “right”/“wrong” accuracy and probabilistic prediction – and with smaller trees, i.e., less leaf nodes.

There are three main methods for handling continuous variables in naive Bayes classifiers, namely, the normal method (parametric approach), the kernel method (non parametric approach) and discretization. In this article, we perform a methodologically sound comparison of the three methods, which shows large mutual differences of each of the methods and no single method being universally better. This suggests that a method for selecting one of the three approaches to continuous variables could improve overall performance of the naive Bayes classifier. We present three methods that can be implemented efficiently

v

-fold cross validation for the normal, kernel and discretization method. Empirical evidence suggests that selection using 10 fold cross validation (especially when repeated 10 times) can largely and significantly improve over all performance of naive Bayes classifiers and consistently outperform any of the three popular methods for dealing with continuous variables on their own. This is remarkable, since selection among more classifiers does not consistently result in better accuracy.

With the explosive growth of the Internet, e-mails are regarded as one of the most important methods to send e-mails as a substitute for traditional communications. As e-mail has become a major mean of communication in the Internet age, exponentially growing spam mails have been raised as a main problem. As a result of this problem, researchers have suggested many methodologies to solve it. Especially, Bayesian classifier-based systems show high performances to filter spam mail and many commercial products available. However, they have several problems. First, it has a cold start problem, that is, training phase has to be done before execution of the system. The system must be trained about spam and non-spam mail. Second, its cost for filtering spam mail is higher than rule-based systems. Last problem, we focus on, is that the filtering performance is decreased when E-mail has only a few terms which represent its contents. To solve this problem, we suggest spam mail filtering system using concept indexing and Semantic Enrichment. For the performance evaluation, we compare our experimental results with those of Bayesian classifier which is widely used in spam mail filtering. The experimental result shows that the proposed system has improved performance in comparison with Bayesian classifier respectively.

Most documented Bayesian network (BN) applications have been built through knowledge elicitation from domain experts (DEs). The difficulties involved have led to growing interest in machine learning of BNs from data. There is a further need for combining what can be learned from the data with what can be elicited from DEs. In this paper, we propose a detailed methodology for this combination, specifically for the parameters of a BN.

Rough set theory (RST) was introduced into radar emitter signal (RES) recognition. A novel approach was proposed to discretize continuous interval valued features and attribute reduction method was used to select the best feature subset from original feature set. Also, rough neural network (NN) classifier was designed. Experimental results show that the proposed hybrid approach based on RST and NN achieves very high recognition rate and good efficiency. It is proved to be a valid and practical approach.

In Kernel based Nonlinear Subspace (KNS) methods, the subspace dimensions have a strong influence on the performance ofthe subspace classifier. In this paper, we propose a new method of systematically and efficiently selecting optimal, or near-optimal subspace dimensions for KNS classifiers using a search strategy and a heuristic function termed as the

Overlapping

criterion. The task of selecting optimal subspace dimensions is equivalent to finding the best ones from a given problem-domain solution space. We thus employ the Overlapping criterion of the subspaces as a heuristic function, by which the search space can be pruned to find the best solution to reduce the computation significantly. Our experimental results demonstrate that the proposed mechanism selects the dimensions efficiently without sacrificing the classification accuracy. The results especially demonstrate that the computational advantage for

large

data sets is significant.

We present here an original work that uses machine learning techniques to combine time series forecasts. In this proposal, a machine learning technique uses features of the series at hand to define adequate weights for the individual forecasting methods being combined. The combined forecasts are the weighted average of the forecasts provided by the individual methods. In order to evaluate this solution, we implemented a prototype that uses a MLP network to combine two widespread methods. The experiments performed revealed significantly accurate forecasts.

It is indubitable that we can obtain numerous discovered patterns using a Web mining model. However, there are many meaningless patterns, and also some discovered patterns might include uncertainties as well. Therefore, the difficult problem is how to utilize and maintain the discovered patterns for the effectiveness of using the Web data. This paper presents a Web data mining and reasoning model for this problem. The objective of mining is automatic ontology extraction; whereas, the objective of reasoning is the utilization and maintenance of discovered knowledge on the ontology. The model also deals with pattern evolution.

In this paper, we address the problem of disambiguating a sequence of semantically overloaded icons. We formulate the problem as a constraint satisfaction problem and discuss the knowledge representation required to facilitate checking of constraints. Our algorithm helps in reducing the size of vocabulary and hence makes these interfaces more usable for the disabled population.

In this paper, we developed an intelligent grading system, which scores descriptive examination papers automatically, based on Probabilistic Latent Semantic Analysis (PLSA). For grading, we estimated semantic similarity between a student paper and a model paper. PLSA is able to represent complex semantic structures of given contexts, like text passages, and are used for building linguistic semantic knowledge which could be used in estimating contextual semantic similarity. In this paper, we marked the real examination papers and we can acquire about 74% accuracy of a manual grading, 7% higher than that from the Simple Vector Space Model.

The conversational agent understands and provides users with proper information based on natural language. Conventional agents based on pattern matching have much restriction to manage various types of real dialogues and to improve the answering performance. For the effective construction of conversational agents, we propose a domain-adaptive conversational agent that infers the user’s intention with two-stage inference and incrementally improves the answering performance through a learning dialogue. We can confirm the usefulness of the proposed method with examples and usability tests.

We present a new feature extraction method for robust speech recognition in the presence of additive white Gaussian noise. The proposed method is made up of two stages in cascade. The first stage is denoising process based on the wavelet domain hidden Markov tree model, and the second one is reduction of the influence of the residual noise in the filter bank analysis. To evaluate the performance of the proposed method, recognition experiments were carried out for noisy speech with signal-to-noise ratio from 25 dB to 0 dB. Experiment results demonstrate the superiority of the proposed method to the conventional ones.

For a free-word order language such as Korean, case marking remains a central topic in generative grammar analyses for several reasons. Case plays a central role in argument licensing, in the signalling of grammatical functions, and has the potential to mark properties of information structure. In addition, case marking presents a theoretical test area for understanding the properties of the syntax-morphology interface. This is why it is no exaggeration to say that parsing Korean sentences starts from work on the case system of the language. This paper reports the project that develops a Korean Resource Grammar (KRG, Kim and Yang 2004), built upon the constrain-based mechanisms of feature unification and multiple inheritance type hierarchies as an extension of HPSG (Head-driven Phrase Structure Grammar), and shows that the results of its implementation in the Linguistic Knowledge Building System (cf. Copestake 2002) prove its empirical and theoretical efficiency in parsing case-related phenomena.

The

Belief-Desire-Intention

(BDI) model of agency is an architecture based on Bratman’s theory of practical reasoning. Hierarchical Task Network (HTN) decomposition on the other hand is a

planning

technique which has its roots in classical planning systems such as STRIPS. Despite being used for different purposes, HTN and BDI systems appear to have a lot of similarities in the problem solving approaches they adopt. This paper presents these similarities. A systematic method for mapping between the two systems is developed, and experimental results for different kinds of environments are presented.

This paper presents a formal method based on the high-level semantics of processes to reason about continuous change. With a case study we show how the semantics of processes can be integrated with the situation calculus. Our aim is to overcome some limitations of the earlier works and to realize the automated reasoning about continuous change.

We present a time and energy optimal controller for a two-wheeled differentially driven robot. We call a

mission

the task of bringing the robot from an initial state to a desired final state (a state is the aggregate vector of the position and velocity vectors). The proposed controller is time optimal in the sense that it can determine the minimum amount of time required to perform a mission. The controller is energy optimal in the sense that given a time constraint of

n

seconds, the controller can determine what is the most energy efficient sequence of accelerations to complete the mission in

n

seconds.

Guarino and Welty have developed the notion of identity condition (IC) as a subsumption constraint of ontological analysis, that is, IC must be inherited or carried among sorts. In practical cases, a sort is often regarded to carry more than one identity condition via inheritance. Thus, we provided the idea of multiple ICs in [8]. Here, we extend our idea in order-sorted logic because this logic is one of the most promising ways to treat a sort ontology. For this purpose, we reconsider the definition of identity and rigidity in terms of order-sorted language and possible world semantics. Then, we propose an inheritance mechanism of identity conditions in order to solve subsumption inconsistency among sorts. We present a practical example of knowledge management to illustrate the advantages of our formalism.

Fuzzy system modeling approximates highly nonlinear systems by means of fuzzy if-then rules. In the literature, different approaches are proposed for mining fuzzy if-then rules from historical data. These approaches usually utilize fuzzy clustering in

structure identification

phase. In this research, we are going to analyze three possible approaches from the literature and try to compare their performances in a medical diagnosis classification problem, namely Aachen Aphasia Test. Given the fact that the comparison is conducted on a single data set; the conclusions are by no means inclusive. However, we believe that the results might provide some valuable insights.

A component neural networks parallel training algorithm PLA is proposed, which encourages component neural network to learn from expected goal and the others, so all component neural networks are trained simultaneously and interactively. In the stage of combining component neural networks, we provide a parallel weight optimal approach GASEN-e by expanding GASEN proposed by Zhou et al, which assign weight for every component neural network and bias for their ensemble. Experiment results show that a neural networks ensemble system is efficient constructed by PLA and GASEN-e.

The use of microcontroller in neural network realizations is cheaper than those specific neural chips. In this study, an intelligent gas concentration estimation system is described. A neural network (NN) structure with tapped time delays was used for the concentration estimation of CCI

4

gas from the trend of transient sensor responses. After training of the NN, the updated weights and biases were applied to the embedded neural network implemented on the 8051 microcontroller. The microcontroller based gas concentration estimation system performs NN based concentration estimation, the data acquisition and user interface tasks. This system can estimate the gas concentrations of CCI

4

with an average error of 1.5 % before the sensor response time. The results show that the appropriateness of the system is observed.

In this paper we introduce an Ant Colony Optimisation (ACO) algorithm to find solutions for the well-known Knight’s Tour problem. The algorithm utilizes the implicit parallelism of ACO’s to simultaneously search for tours starting from all positions on the chessboard. We compare the new algorithm to a recently reported genetic algorithm, and to a depth-first backtracking search using Warnsdorff’s heuristic. The new algorithm is superior in terms of search bias and also in terms of the rate of finding solutions.

Based on the Antibody Clonal Selection Theory of immunology, we put forward a novel artificial immune system algorithm, Immune Clonal Selection Algorithm for Multiuser Detection in DS-CDMA Systems. The performance of the new detector, named by ICSMUD, is evaluated via computer simulations. When compared with Optimal Multiuser detection, ICSMUD can reduce the computational complexity significantly. When compared with detectors based on Standard Genetic Algorithm and A Novel Genetic Algorithm, ICSMUD has the best performance in eliminating multiple-access interference and “near-far” resistance and performs quite well even when the number of active users and the packet length are considerably large.

Immune clone selection algorithm is a new intelligent algorithm which can effectively overcome the prematurity and slow convergence speed of traditional evolution algorithm because of the clonal selection strategy and clonal mutation strategy. We apply the immune clonal selection algorithm to the process of modeling normal behavior. We compare our algorithm with the algorithm which applies the genetic algorithm to intrusion detection and applies the negative selection algorithm of the artificial immune system to intrusion detection in the dataset kddcup99. The experiment results have shown that the rule set obtained by our algorithm can detect unknown attack behavior effectively and have higher detection rate and lower false positive rate.

Salinity is a growing problem that affects millions of hectares of agricultural land. Due to the devastating effects of dryland salinity, land owners, the government and catchment groups require cost effective information in the form of salinity maps. This paper investigates the use of a backpropagation neural network to map dryland salinity in the Wimmera region of Victoria, Australia. Data used in this research includes radiometric readings from airborne geophysical measurements and satellite imagery. The results achieved were very promising and indicate the potential for further research in this area.

A new solution to real time signal detection in the noise is presented in this paper. The proposed approach uses the modified RBF neural network (RBFNN) for the purposes of enhancing the ability of signal detection with low signal-to-noise radio (SNR). The characteristics and the advantages of the normalized RBFNN are discussed. As an application, the extraction of singletrial evoked potentials (EP) is investigated. The performance of the presented method is also addressed and compared with adaptive and common RBFNN methods. Several results are included to show the applicability and the effectiveness of the new model.

An important issue in genetic algorithms is the relationship between the difficulty of a problem and the choice of encoding. Two questions remain unanswered: is their a statistically demonstrable relationship between the difficulty of a problem and the choice of encoding, and, if so, what it the actual mechanism by which this occurs?

In this paper we use components of a rigorous statistical methodology to demonstrate that the choice of encoding has a real effect upon the difficulty of a problem. Computer animation is then used to illustrate the actual mechanism by which this occurs.

The paper proposes the

semipublic encrypting scheme

for visual cryptography using the

Q’tron neural-network

(

Q’tron NN

) model. This encrypting scheme hides only the true secret from the public. That is, the pictorial meaning appearing in a

public share

describes the public information in a document while leaving its confidential part undisclosed. A piece of confidential information is retrievable if and only if a right

user share

is available. The method to construct the Q’tron NN to fulfill the aforementioned scheme will be investigated. An application that uses the scheme for

key distribution

in a public area will be demonstrated.

Negative selection algorithm is used to generate detector for change detection, anomaly detection. But it can not be adapted to the change of self data because the match threshold must be set at first. In this paper, inspired from T-cells maturation, a novel algorithm composed of positive and negative selection is proposed to generate T-detector. Genetic algorithm is used to evolve the detectors with lower match threshold. The proposed algorithm is tested by simulation experiment for anomaly detection and compared with negative selection algorithm. The results show that the proposed algorithm is more effective than negative selection algorithm. Match threshold is selfadapted and False Positive is controlled by parameter S.