Intelligent Data Engineering and Automated Learning – IDEAL 2008

Exploitation-oriented Learning

(XoL) is a novel approach to goal-directed learning from interaction. Though

reinforcement learning

is much more focus on the learning and can gurantee the optimality in

Markov Decision Processes

(MDPs) environments, XoL aims to learn

a rational policy

, whose expected reward per an action is larger than zero, very quickly. We know PS-r* that is one of the XoL methods. It can learn

an useful rational policy

that is not inferior to a random walk in

Partially Observed Markov Decision Processes

(POMDPs) environments where the number of types of a reward is one. However, PS-r* requires

O

(

MN

2

) memories where

N

and

M

are the numbers of types of a sensory input and an action.In this paper, we propose PS-r

#

that can learn an useful rational policy in the POMDPs environments by

O

(

MN

) memories. We confirm the effectiveness of PS-r

#

in numerical examples.

In this paper, we apply a new method to forecast short-term traffic flows. It is kernel regression based on a Mahalanobis metric whose parameters are estimated by gradient descent methods. Based on the analysis for eigenvalues of learned metric matrices, we further propose a method for evaluating the effectiveness of the learned metrics. Experiments on real data of urban vehicular traffic flows are performed. Comparisons with traditional kernel regression with the Euclidean metric under two criterions show that the new method is more effective for short-term traffic flow forecasting.

Distance measures are an important means to find the difference of data. In this paper, we develop a type of hybrid weighted distance measures which are based on the weighted distance measures and the ordered weighted averaging operator, and aslo point out some of their special cases. Then, we apply the developed measures to pattern recognition.

For traditional human face based biometrics, usually one task (face recognition) is learned at one time. This single task learning (STL) approach may neglect potential rich information resources hidden in other related tasks, while multitask learning (MTL) can make full use of the latent information. MTL is an inductive transfer method which improves generalization by using the domain information contained in the training signals of related tasks as an inductive bias. In this paper, backpropagation (BP) network based MTL approach is proposed for face recognition. The feasibility of this approach is demonstrated through two different face recognition experiments, which show that MTL based on BP neural networks is more effective than the traditional STL approach, and that MTL is also a practical approach for face recognition.

Synchronous finite state machines are very important for digital sequential systems. Among other important aspects, they represent a powerful way for synchronising hardware components so that these components may cooperate adequately in the fulfilment of the main objective. In this paper, we propose to use an evolutionary methodology inspired from quantum computation to yield a concise and efficient evolvable hardware that implements the state machine control logic.

This paper presents a new strategy in designing artificial neural networks. We call this strategy as

adaptive merging and growing strategy

(AMGS). Unlike most previous strategies on designing ANNs, AMGS puts emphasis on autonomous functioning in the design process. The new strategy reduces or increases an ANN size during training based on the learning ability of hidden neurons and the training progress of the ANN, respectively. It merges correlated hidden neurons to reduce the network size, while it splits existing hidden neuron to increase the network size. AMGS has been tested on designing ANNs for five benchmark classification problems, including Australian credit card assessment, diabetes, heart, iris, and thyroid problems. The experimental results show that the proposed strategy can design compact ANNs with good generalization ability.

Many research efforts have been done on the automatic classification of heart sound signals to support clinicians in heart sound diagnosis. Recently, hidden Markov models (HMMs) have been used quite successfully in the automatic classification of the heart sound signal. However, in the classification using HMMs, there are so many heart sound signal types that it is not reasonable to assign a new class to each of them. In this paper, rather than constructing an HMM for each signal type, we propose to build an HMM for a set of acoustically-similar signal types. To define the classes, we use the KL (Kullback-Leibler) distance between different signal types to determine if they should belong to the same class. From the classification experiments on the heart sound data consisting of 25 different types of signals, the proposed method proved to be quite efficient in determining the optimal set of classes.

In this paper a semi-fragile watermark scheme is proposed based on the Logistic chaotic sequence and singular value decomposition (SVD). Chaotic sequence is very sensitive to the initial value and has zero-value of cross-correlation. With these properties of chaotic sequence, the uniqueness of the watermark can be efficiently obtained. Some singular values show robustness under lossy compression and can be adopted as a good carrier of watermark. Hence SVD quantization is also employed in our algorithm. In the proposed approach, algorithm efficiency and the influence on host image quality will be discussed theoretically and experimentally. The experimental results show the high robustness against lossy compression and the place being maliciously tampered can be accurately detected and located on the protected digital images.

The electrical network should ensure that an adequate supply is available to meet the estimated load of the consumers in both the near and more distant future. This must of course, be done at minimum possible cost consistent with satisfactory reliability and quality of the supply. In order to avoid excessive voltage drop and minimise loss, it may be economical to install apparatus to balance or partially balance the loads. It is believed that the technology to achieve an automatic load balancing lends itself readily for the implementation of different types of algorithms for automatically rearranging the connection of consumers on the low voltage side of a feeder for optimal performance. In this paper the authors present a Support Vector Machines (SVM) implementation. The loads are first normalised and then sorted before applying the SVM to do the balancing.

In this paper, a new method for license plate detection based on AdaBoost is proposed. In the proposed method, auto-correlation feature, which is ignored by previous learning-based method, is introduced to feature pool. Since that there are two types of Chinese license plate, one type is deeper-background-lighter-character and the other is lighter-background-deeper-character, training a detector cannot convergent. To avoid this problem, two detectors are designed in the proposed method. Experimental results show the superiority of proposed method.

A union-based rule-antecedent fuzzy neural networks (URFNN), which can guarantee a parsimonious knowledge base with reduced number of rules, is proposed. The URFNN allows union operation of input fuzzy sets in the antecedents to cover bigger input domain compared with the complete structure rule which consists of AND combination of all input variables in its premise. To construct the URFNN, we consider the union-based logic processor (ULP) which consists of OR and AND fuzzy neurons. The fuzzy neurons exhibit learning abilities as they come with a collection of adjustable connection weights. In the development stage, genetic algorithm (GA) constructs a Boolean skeleton of URFNN, while gradient-based learning refines the binary connections of GA-optimized URFNN for further improvement of the performance index. A cart-pole system is considered to verify the effectiveness of the proposed method.

In this paper, we propose a new face detection model, which is developed by combining the conventional AdaBoost algorithm for human face detection with a biologically motivated face-color preferable selective attention. The biologically motivated face-color preferable selective attention model localizes face candidate regions in a natural scene, and then the Adaboost based face detection process only works for those localized face candidate areas to check whether the areas contain a human face. The proposed model not only improves the face detection performance by avoiding miss-localization of faces induced by complex background such as face-like non-face area, but can enhances a face detection speed by reducing region of interests through the face-color preferable selective attention model. The experimental results show that the proposed model shows plausible performance for localizing faces in real time.

In this paper, we propose a top-down object biased attention model which is based on human visual attention mechanism integrating feature based bottom-up attention and goal based top-down attention. The proposed model can guide attention to focus on a given target colored object over other objects or feature based salient areas by considering the object color biased attention mechanism. We proposed a growing fuzzy topology ART that plays important roles for object color biased attention, one of which is to incrementally learn and memorize features of arbitrary objects and the other one is to generate top-down bias signal by competing memorized features of a given target object with features of an arbitrary object. Experimental results show that the proposed model performs well in successfully focusing on given target objects, as well as incrementally perceiving arbitrary objects in natural scenes.

Many eye gaze systems use special infrared (IR) illuminator and choose IR-sensitive CCD camera to estimate eye gaze. The IR based system has the limitation of inaccurate gaze detection in ambient natural light and the number of IR illuminator and their particular location has also effect on gaze detection. In this paper, we present a eye gaze detection method based on computer screen illumination as light emitting source and choose high speed camera for image acquisition. In order to capture the periodic flicker patterns of monitor screen the camera is operated on frame rate greater than twice of the screen refresh rate. The screen illumination produced a mark on the corneal surface of the subject’s eye as screen-glint. The screen reflection information has two fold advantages. First, we can utilize the screen reflection as screen-glint, which is very useful to determine where eye is gazing. Secondly the screen-glint information utilize to localized eye in face image. The direction of the user’s eye gaze can be determined through polynomial calibration function from the relative position of the center of iris and screen-glint in both eyes. The results showed that our propose configuration could be used for gaze detection method and this will lead to increased gaze detection role for the next generation of human computer interfaces.

This work presents a novel GA-Taguchi-based feature selection method. Genetic algorithms are utilized with randomness for “global search” of the entire search space of the intractable search problem. Various genetic operations, including crossover, mutation, selection and replacement are performed to assist the search procedure in escaping from sub-optimal solutions. In each iteration in the proposed nature-inspired method, the Taguchi methods are employed for “local search” of the entire search space and thus can help explore better feature subsets for next iteration. The two-level orthogonal array is utilized for a well-organized and balanced comparison of two levels for features—a feature is or is not selected for pattern classification—and interactions among features. The signal-to-noise ratio (SNR) is then used to determine the robustness of the features. As a result, feature subset evaluation efforts can be significantly reduced and a superior feature subset with high classification performance can be obtained. Experiments are performed on different application domains to demonstrate the performance of the proposed nature-inspired method. The proposed hybrid GA-Taguchi-based approach, with wrapper nature, yields superior performance and improves classification accuracy in pattern classification.

We proposed a novel algorithm of supervised feature selection and adaptation for enhancing the classification accuracy of unsupervised Nonnegative Matrix Factorization (NMF) feature extraction algorithm. At first the algorithm extracts feature vectors for a given high dimensional data then reduce the feature dimension using mutual information based relevant feature selection and finally adapt the selected NMF features using the proposed Non-negative Supervised Feature Adaptation (NSFA) learning algorithm. The supervised feature selection and adaptation improve the classification performance which is fully confirmed by simulations with text-document classification problem. Moreover, the non-negativity constraint, of this algorithm, provides biologically plausible and meaningful feature.

In this paper, a neural network-based algorithm is proposed to explore the order of the measured data points in surface fitting. In computer-aided design, the ordered points serves as the input to fit smooth surfaces so that a reverse engineering (i.e. RE) system can be established for 3D sculptured surface design. The geometry feature recognition capability of back-propagation neural networks is explored in this paper. Scan or measuring number and 3D coordinates are used as the inputs of the proposed neural networks to determine the curve to which each data point belongs and the order number of data point in the same curve. In the segmentation process, the neural network output is segment number; while the segment number and sequence number in the same curve are the outputs when sequencing the points in the same curve. After evaluating a large number of trials with various neural network architectures, two optimal models are selected for segmentation and sequence. The proposed model can easily adapt for new data from another sequence for surface fitting. In comparison to Lin et al.’s (1998) method, the proposed algorithm neither needs to calculate the angle formed by each point and its two previous ones nor causes any chaotic phenomenon.

Nonnegative matrix factorization (NMF) is a popular method for multivariate analysis of nonnegative data, the goal of which is decompose a data matrix into a product of two factor matrices with all entries in factor matrices restricted to be nonnegative. NMF was shown to be useful in a task of clustering (especially document clustering). In this paper we present an algorithm for orthogonal nonnegative matrix factorization, where an orthogonality constraint is imposed on the nonnegative decomposition of a term-document matrix. We develop multiplicative updates directly from true gradient on Stiefel manifold, whereas existing algorithms consider additive orthogonality constraints. Experiments on several different document data sets show our orthogonal NMF algorithms perform better in a task of clustering, compared to the standard NMF and an existing orthogonal NMF.

In this paper, we introduce a new concept of

enhancement

and

relaxation

to discover features in input patterns in competitive learning. We have introduced mutual information to realize competitive processes. Because mutual information is an average over all input patterns and competitive units, it cannot be used to detect detailed feature extraction. To examine in more detail how a network is organized, we introduce the enhancement and relaxation of competitive units through some elements in a network. With this procedure, we can estimate how the elements are organized with more detail. We applied the method to a simple artificial data and the famous Iris problem to show how well the method can extract the main features in input patterns. Experimental results showed that the method could more explicitly extract the main features in input patterns than the conventional techniques of the SOM.

This paper proposed a hybrid functional link artificial neural network (HFLANN) embedded with an optimization of input features for solving the problem of classification in data mining. The aim of the proposed approach is to choose an optimal subset of input features using genetic algorithm by eliminating features with little or no predictive information and increase the comprehensibility of resulting HFLANN. Using the functionally expanded selected features, HFLANN overcomes the non-linearity nature of problems, which is commonly encountered in single layer neural networks. An extensive simulation studies has been carried out to illustrate the effectiveness of this method over to its rival functional link artificial neural network (FLANN) and radial basis function (RBF) neural network.

Ensemble learning is one of the main directions in machine learning and data mining, which allows learners to achieve higher training accuracy and better generalization ability. In this paper, with an aim at improving generalization performance, a novel approach to construct an ensemble of neural networks is proposed. The main contributions of the approach are its diversity measure for selecting diverse individual neural networks and weighted fusion technique for assigning proper weights to the selected individuals. Experimental results demonstrate that the proposed approach is effective.

The generalization ability of a machine learning algorithm varies on the specified values to the model-hyperparameters and the degree of noise in the learning dataset. If the dataset has a sufficient amount of labeled data points, the optimal value for the hyperparameter can be found via validation by using a subset of the given dataset. However, for semi-supervised learning–one of the most recent learning algorithms–this is not as available as in conventional supervised learning. In semi-supervised learning, it is assumed that the dataset is given with only a few labeled data points. Therefore, holding out some of labeled data points for validation is not easy. The lack of labeled data points, furthermore, makes it difficult to estimate the degree of noise in the dataset. To circumvent the addressed difficulties, we propose to employ ensemble learning and graph sharpening. The former replaces the hyperparameter selection procedure to an ensemble network of the committee members trained with various values of hyperparameter. The latter, on the other hand, improves the performance of algorithms by removing unhelpful information flow by noise. The experimental results present that the proposed method can improve performance on a publicly available bench-marking problems.

The Self-Organizing Map (SOM), which projects a (high-dimensional) data manifold onto a lower-dimensional (usually 2-d) rigid lattice, is a commonly used manifold learning algorithm. However, a postprocessing – that is often done by interactive visualization schemes – is necessary to reveal the knowledge of the SOM. Thanks to the SOM property of producing (ideally) a topology preserving mapping, existing visualization schemes are often designed to show the similarities local to the lattice without considering the data topology. This can produce inadequate tools to investigate the detailed data structure and to what extent the topology is preserved during the SOM learning. A recent graph based SOM visualization, CONNvis [1], which exploits the underutilized knowledge of data topology, can be a suitable tool for such investigation. This paper discusses that CONNvis can represent the data topology on the SOM lattice despite the rigid grid structure, and hence can show the topology preservation of the SOM and the extent of topology violations.

This paper proposes a kind of novel kernel functions obtained from the reproducing kernels of Hilbert spaces associated with special inner product. SVM with the proposed kernel functions only need less support vectors to construct two-class hyperplane than the SVM with Gaussian kernel functions, so the proposed kernel functions have the better generalization. Finally, SVM with reproducing and Gaussian kernels are respectively applied to two benchmark examples: the well-known Wisconsin breast cancer data and artificial dataset.

Temporal regularity of pattern appearance can be regarded as an important criterion for measuring the interestingness in several applications like market basket analysis, web administration, gene data analysis, network monitoring, and stock market. Even though there have been some efforts to discover

periodic

patterns in time-series and sequential data, none of the existing works is appropriate for discovering the patterns that occur regularly in a transactional database. Therefore, in this paper, we introduce a novel concept of mining

regular

patterns from transactional databases and propose an efficient data structure, called Regular Pattern tree (RP-tree in short), that enables a pattern growth-based mining technique to generate the complete set of

regular

patterns in a database for a user-given

regularity

threshold. Our comprehensive experimental study shows that RP-tree is both time and memory efficient in finding

regular

pattern.

To help people obtain the most important information daily in the shortest time, a novel framework is presented for simultaneous key entities extraction and significant events mining from daily web news. The technique is mainly based on modeling entities and news documents as weighted undirected bipartite graph, which consists of three steps. First, key entities are extracted by scoring all candidate entities on a specific day and tracking their trends within a specific time window. Second, a weighted undirected bipartite graph is built based on entities and related news documents, then mutual reinforcement is imposed on the bipartite graph to rank both of them. Third, clustering on news articles generates daily significant events. Experimental study shows effectiveness of this approach.

This paper evaluates the performance of intelligent models using decision trees, rough sets, and neural networks for predicting smokers’ quitting behaviour. 18 models are developed based on 6 data sets created from the International Tobacco Control Four Country Survey. 13 attributes about smokers’ beliefs about quitting (BQ) and 13 attributes about smokers’ beliefs about smoking (BS) are used as inputs. The output attribute is the smokers’ status of making a quit attempt (MQA) or planning to quit (PTQ). The neural network models outperform both decision tree models and rough set models in terms of prediction ability. Models using both BQ and BS attributes as inputs perform better than models using only BQ or BS attributes. The BS attributes contribute more to MQA, whereas the BQ attributes have more impact on PTQ. Models for predicting PTQ outperform models for predicting MQA. Determinant attributes that affect smokers’ quitting behaviour are identified.

The depth information in the face represents personal features in detail. In particular, the surface curvatures extracted from the face contain the most important personal facial information. These surface curvature and eigenface, which reduce the data dimensions with less degradation of original information, are collaborated into the proposed 3D face recognition algorithm. The principal components represent the local facial characteristics without loss for the information. Recognition for the eigenface referred from the maximum and minimum curvatures is performed. To classify the faces, the max plus algebra based neural networks (morphological neural networks) optimized by hybrid genetic algorithm are considered. Experimental results on a 46 person data set of 3D images demonstrate the effectiveness of the proposed method.

Mobile devices can now handle a great deal of information thanks to the convergence of diverse functionalities. Mobile environments have already shown great potential in terms of providing customized services to users because they can record meaningful and private information continually for long periods of time. Until now, most of this information has been generally ignored because of the limitations of mobile devices in terms of power, memory capacity and speed. In this paper, we propose a novel method that efficiently infers semantic information and overcome the problems. This method uses an effective probabilistic Bayesian network model for analyzing various kinds of log data in mobile environments, which were modularized in this paper to decrease complexity. We also discuss how to discover and update the Bayesian inference model by using the proposed BN learning method with training data. The proposed methods were evaluated with artificial mobile log data generated and collected in the real world.

Skin analysis has potential uses in many fields, including computer assisted diagnosis for dermatology, topical drug efficacy testing for the pharmaceutical industry, and quantitative product comparison for cosmetics. In medicine, skin pores are the openings of hair follicles, oil glands, and sweat glands. There are many skin problems associated with skin pores, such as blackheads which are not dirt and cannot be washed away, enlarged pores which are due to over activity of the sebaceous glands in the skin. In computer-aided skin analysis, skin pores are helpful features for skin image registration, skin texture modeling, and skin statement evaluation. In this paper we mainly focus on image-based skin pores detection problem and propose an integrated solution based on fuzzy c-mean algorithm. In our work, research images include images taking by digital camera with long focus lens and images taking by microscope. A global luminance proportion method will be used for skin image preprocessing because of reflection and interreflection of light on the skin surface. We provide experiments to demonstrate the effective and efficiency of our solution.

A feature based relation classification approach is presented, in which probabilistic and semantic relatedness features between patterns and relation types are employed with other linguistic information. The importance of each feature set is evaluated with Chi-square estimator, and the experiments show that, the relatedness features have big impact on the relation classification performance. A series experiments are also performed to evaluate the different machine learning approaches on relation classification, among which Bayesian outperformed other approaches including Support Vector Machine (SVM).

In this paper a novel technique useful to guarantee privacy of sensitive data with specific focus on numeric databases is presented. It is noticed that analysts and decision makers are interested in summary values of the data rather than the actual values. The proposed method considers that the maximum information lies in association of attributes rather than their actual proper values. Therefore it is aimed to perturb attribute associations in a controlled way, by shifting the data values of specific columns by rotating fields. The number of rotations is determined via using a support function for association rule handling and an algorithm that computes the best-choice rotation dynamically. Final summary statistics such as average, standard deviation of the numeric data are preserved by making bin average replacements for the actual values. The methods are tested on selected datasets and results are reported.

Existing weighted frequent pattern (WFP) mining algorithms assume that each item has fixed weight. But in our real world scenarios the weight (price or significance) of an item can vary with time. Reflecting such change of weight of an item is very necessary in several mining applications such as retail market data analysis and web click stream analysis. In this paper, we introduce a novel concept of adaptive weight for each item and propose an algorithm AWFPM (adaptive weighted frequent pattern mining). Our algorithm can handle the situation where the weight (price or significance) of an item may vary with time. Extensive performance analyses show that our algorithm is very efficient and scalable for WFP mining using adaptive weights.

Manifold learning methods model high-dimensional data through low-dimensional manifolds embedded in the observed data space. This simplification implies that their are prone to trustworthiness and continuity errors. Generative Topographic Mapping (GTM) is one such manifold learning method for multivariate data clustering and visualization, defined within a probabilistic framework. In the original formulation, GTM is optimized by minimization of an error that is a function of Euclidean distances, making it vulnerable to the aforementioned errors, especially for datasets of convoluted geometry. Here, we modify GTM to penalize divergences between the Euclidean distances from the data points to the model prototypes and the corresponding geodesic distances along the manifold. Several experiments with artificial data show that this strategy improves the continuity and trustworthiness of the data representation generated by the model.

The Fractional Auto-Regressive Integrated Moving Average (FARIMA) model is often used to model and predict network traffic demand which exhibits both long-range and short-range dependence. However, finding the best model to fit a given set of observations and achieving good performance is still an open problem. We present a strategy, namely Aggregating Algorithm, which uses several FARIMA models and then aggregates their outputs to achieve a guaranteed (in a sense) performance. Our feasibility study experiments on the public datasets demonstrate that using the Aggregating Algorithm with FARIMA models is a useful tool in predicting network traffic demand.

Several algorithms for induction of decision trees have been developed to solve problems with large datasets, however some of them have spatial and/or runtime problems using the whole training sample for building the tree and others do not take into account the whole training set. In this paper, we introduce a new algorithm for inducing decision trees for large numerical datasets, called IIMDT, which builds the tree in an incremental way and therefore it is not necesary to keep in main memory the whole training set. A comparison between IIMDT and ICE, an algorithm for inducing decision trees for large datasets, is shown.

By 2006, the number of mobile subscribers in Africa outnumbered that of fixed line subscribers with nearly 200 million mobile subscribers across the continent [1][2]. By the end of 2007, it was estimated that the number of mobile subscribers would exceed 278 million subscribers [2]. Mobile Telephony has been viewed as a critical enabling technology that is capable of boosting local economies across Africa due to the ease of roll out of wireless technologies in comparison to fixed line networks. With the boom in wireless networks across Africa, a growing demand to effectively predict the rate of growth in demand for capacity in various sectors of the network has risen with cellular network operators. This paper looks at using

Spectral Analysis

techniques for the extraction of features from cellular network traffic data that could be linked to subscriber behavior. This could then in turn be used to determine capacity requirements within the network.

Recently, the necessity of personalized document summarization reflecting user interest from search results is increased. This paper proposes a personalized document summarization method using non-negative semantic feature (NSF) and non-negative semantic variable (NSV) to extract sentences relevant to a user interesting. The proposed method uses NSV to summarize generic summary so that it can extract sentences covering the major topics of the document with respect to user interesting. Besides, it can improve the quality of personalized summaries because the inherent semantics of the documents are well reflected by using NSF and the sentences most relevant to the given query are extracted efficiently by using NSV. The experimental results demonstrate that the proposed method achieves better performance the other methods.

Large-scale collaboration is a key success factor in today scientific experiments, usually involving a variety of digital resources, while Cooperative Information Systems (CISs) represent a feasible solution for sharing distributed information sources and activities. On this premise, the aim of this paper is to provide a paradigm for modeling scientific experiments as distributed processes that a group of scientists may go through on a network of cooperative e-nodes interacting with one another in order to offer or to ask for services. By discussing a bioinformatics case study, the paper details how the problem solving strategy related to a scientific experiment can be expressed by a workflow of single cooperating activities whose implementation is carried out on a prototypical service-based scientific environment.

In the natural brain, memory retaining, reconfiguration and retrieval are very important functions for maintaining the fresh memory. Especially reconfiguration takes an important role for arranging more relevant information to the personal aspect and purpose closely. The closely rearranged information is easy to be activated for processing. In this perspective, the strategy for reconfiguration of Personal Knowledge Network was proposed. Personal Knowledge network is reconstructed by Type matching selection and knowledge reconfiguration algorithm. The proposed system was applied to the Virtual Memory and tested with the sample data.

In the application of brain-computer interfaces (BCIs), energy features are both physiologically well-founded and empirically effective to describe electroencephalogram (EEG) signals for classifying brain activities. Recently, a linear method named extreme energy ratio (EER) for energy feature extraction of EEG signals in terms of spatial filtering was proposed. This paper gives a nonlinear extension of the linear EER method. Specifically, we use the kernel trick to derive a kernelized version of the original EER feature extractor. The solutions for optimizing the criterion in kernel EER are provided for future use.

This paper gives an outline of a non-invasive brain machine interface (BMI) implemented for controlling a motorized wheelchair online. Subjects were trained by using an effective feedback training method, and they could then control the wheelchair freely, similar to controlling it with a joystick.

A novel multi-resolution approach is presented for vessel segmentation using multi-scale fuzzy clustering and vessel enhancement filtering. According to geometric shape analysis of the vessel structure with different scale, a new fuzzy inter-scale constraint based on antistrophic diffusion linkage model is introduced which builds an efficient linkage relationship between the high resolution feature images and low resolution ones. Meanwhile, this paper develops two new fuzzy distances which describe the fuzzy similarity of line-like structure in adjacent scales effectively. Moreover, a new multiresolution framework combining the inter- and intra-scale constraints is presented. The proposed framework is robust to noisy vessel images and low contrast ones, such as medical images. Segmentation of a number of vessel images shows that the proposed approach is robust and accurate.

The use of domain ontologies is becoming increasingly popular in Medical Natural Language Processing Systems. A wide variety of knowledge bases in multiple languages has been integrated into the Unified Medical Language System (UMLS) to create a huge knowledge source that can be accessed with diverse lexical tools. MetaMap (and its java version MMTx) is a tool that allows extracting medical concepts from free text, but currently there not exists a Spanish version. Our ongoing research is centered on the application of biomedical concepts to cross-lingual text classification, what makes it necessary to have a Spanish MMTx available. We have combined automatic translation techniques with biomedical ontologies and the existing English MMTx to produce a Spanish version of MMTx. We have evaluated different approaches and applied several types of evaluation according to different concept representations for text classification. Our results prove that the use of existing translation tools such as Google Translate produce translations with a high similarity to original texts in terms of extracted concepts.

We study averaging schemes that are specifically adapted to the analysis of electroencephalographic data for the purpose of interpreting temporal information from single trials. We find that a natural assumption about processing speed in the subjects yields a complex but nevertheless robust algorithm for the analysis of electrophysiological data.

Quantitative descriptors of intrinsic properties of imaging data can be obtained from the theory of random matrices (RMT). Based on theoretical results for standardized data, RMT offers a systematic approach to surrogate data which allows us to evaluate the significance of deviations from the random baseline. Considering exemplary fMRI data sets recorded at a visuo-motor task and rest, we show their distinguishability by RMT-based quantities and demonstrate that the degree of sparseness and of localization can be evaluated in a strict way, provided that the data are sufficiently well described by the pairwise cross-correlations.

This paper proposes a robust algorithm for adaptive modelling of EEG signal classification using a modified Extended Kalman Filter (EKF). This modified EKF combines Radial Basis functions (RBF) and Autoregressive (AR) modeling and obtains better classification performance by truncating the filtering distribution when new observations are very informative.

XML query languages such as XQuery, XSLT and SQL/XML are mainly dependent on XPath as the search and extraction language. XPath expressions often define complicated navigations which require expensive query processing costs especially when they are executed over large collections of XML documents. In this paper, we describe an approach of exploiting materialized XPath views to improve the efficiency of relational query processing of XML queries. The main contribution of this paper is to show that an intuitive and

very cheap

Data Guide synopsis of XML path summaries in addition a

light-wight

tracing of XPath steps can significantly reduce the XML query-evaluation costs in the relational hosts. Our experiments shows that the overhead introduced by the use of path summaries and an additional path identifier of node-based relational encoding of the XML documents is negligible but can result in significant reduction of the processing costs of relational evaluation of XML queries.

Relevance vector machines (RVM) is a machine learning approach with good nonlinear approximation capacity and generalization performance. In order to solve the inverse model for nonlinear systems, a multiple relevance vector machines (MRVM) based inverse dynamics model identification approach was presented. The input and output variables were allocated into multiple calculational subspaces according to their differential orders for the system. The RVM was put forward to identify the influence of the outputs to the inputs with a certain differential order in each subspace. Moreover, another RVM was delivered to connect all subspaces, such that the MRVM based inverse dynamics identification model for the nonlinear systems was constructed. At last it was applied to identify the inverse dynamics of a high temperature exchanger for the generator. And the result validates the effectiveness of the proposed approach.

Real world inconsistent information often has to do with not only what conflicting circumstances are but also when they happen. In this paper we present our research work on the temporal characteristics of inconsistent information that can exist in an intelligent system. To facilitate the discussions, we use knowledge base (KB) to refer to the component in an intelligent system that contains knowledge about a problem domain. Knowledge in a KB can be represented in terms of different formalisms, and plays a pivotal role in how an intelligent system accomplishes its intended tasks. The main results reported in this paper include: (1) establishing a formal definition for temporal inconsistency for knowledge in a KB in terms of the interval temporal logic; (2) describing a systematic approach to identifying conflicting intervals for temporally inconsistent assertions in a KB; and (3) delineating the semantic difference between the classical and temporal inconsistency.

The balance of electric power supply and demand is the important precondition of the sustainable development of power industry and national economy. In China, the government adjusts industry structure by putting macro-policy in practice for the balance. But the relationship between macro-policy and electric power supply and demand is non-linear and complicated, namely that policy simulation is a semi-structure problem, and many non-linear relationships and uncertain factors can’t be simulated in traditional linear model. Intelligent Engineering (IE) is a kind of methodology, System Engineering (SE) offers direction for perfecting and applying IE, This paper focuses on the combination of Distributed Artificial Intelligence (DAI) and Cybernetics, and does an innovative research about Distributed Intelligent Control (DIC). Based on IE and optimized solution theory, the controllability of DIC system are defined for the first time. As a sample, agent-based intelligent simulation system is built to simulate the influence from macro policy to electric power supply and demand.

As the ubiquitous computing is rapidly changing, the research on agent technologies is constantly being conducted. In a multi-agent system environment, each agent is registered in directory facilitator in a fixed from for service it provides, and gains a service function capable of modifying and deleting the service. Other agents may inquire about a service they want and receive necessary information on the service. By using this directory facilitator, the user can retrieve the most appropriate service. In this paper, we propose an efficient directory facilitator architecture that can improve the existing agent-based service discovery.

Electricity demand will be affected by national policies and other factors. There are many semi-structure problems in the electricity demand forecasting, which are very difficult to be solved by the use of traditional methods. In this paper, intelligent engineering is developed. It adopts theory and technique of artificial intelligent, soft computing, uncertain theory, and multi-agent system. Three fundamental problems and generalized model are proposed in intelligent space. As a case, inspired by the physical experiment, the laboratory of policy study is built based on intelligent engineering to simulate the impact of the national policy on electricity demand forecasting. A case study in China has been shown in the paper.

A new type of adaptive evolutionary algorithm that combines two genetic algorithms using mutation matrix is developed based on an adaptive resource allocation of CPU time. Performance evaluations are made on the airport scheduling problem with constraint. The two genetic algorithms used are based on the construction of the mutation matrix M(t), which is problem independent as it uses the fitness distribution in the population and the statistical information of the locus only. The mutation matrix is parameter free and adaptive since the matrix elements are time dependent and inherits the information accumulated from past generations. A self-adaptive time sharing method is introduced to allocate resource to the two different strategies, which uses the theory of mean-variance analysis in portfolio management. The application to airport scheduling demonstrates that the self-adaptive mutation only genetic algorithm is able to provide quality solutions efficiently.

There are many well-formedness rules of each UML element in UML specification[1], but there are not any rules that check the consistency among UML diagrams. Therefore, in this paper, we propose several checking rules to improve the consistency among UML diagrams, especially between UML static diagram and dynamic diagram. So we make explicit some requirements on consistency of UML diagrams that are buried in the original well-formedness rules of UML specification and derive some checking rules. Finally, we examine the usefulness of the derived rules through a case study.

In this article we investigate the problem of detection the statistically significant dependences of stock trading return, which occur in particular days of the week (usually the first or the last trading day), and which could be important for creating profitable investment strategies. The identifying such days of the week (day-of-the-week effect) is performed by using artificial neural networks. The research results helped to conclude the effectiveness of application of neural networks, as compared to the traditional linear statistical methods for finding stock trading anomalies. The effectiveness of the method was confirmed by exploring impact of different variables to the day-of-the-week effect, evaluation of their influence and sensitivity analysis, and by selecting best performing neural network type. The experimental verification was implemented by using Vilnius Stock Exchange trading data.

We discuss Bregman divergences and the very close relationship between a class of these divergences and the regular family of exponential distributions before applying them to various topology preserving dimension reducing algorithms. We apply these methods to identification of structure in magnetic resonance images of the brain and show that different divergences reveal different structure in these images.

We review the recent technique of two dimensional canonical correlation analysis and illustrate its use as a method for identification of the location of particular features in a data set.

Classification of entities, for example, into national states, into social groups, into business enterprises and into scientific taxa, is an enduring problem in neural computing. In this paper, we look at the problems faced by researchers in developing a taxonomy of ’multinationality’ and explore the use of hierarchical SOMs in ’discovering’ a taxonomy of multinational corporations (MNCs).

This paper presents an adaptive image watermarking scheme. Watermark bits are embedded adaptively into the non-separable wavelet domain based on the Human Visual System (HVS) model trained by Support Vector Regression (SVR). Unlike conventional separable wavelet filter banks that limit the ability in capturing directional information, non-separable wavelet filter banks contain the basis elements oriented at a variety of directions and different filter banks are able to capture different detail information. After removing the high frequency components, the low frequency subband used for watermark embedding is more robust against noise and other distortions. In addition, owing to the good generalization ability of the support vector machine, watermark embedding strength can be adjusted according to the HVS value. The superiority of non-separable wavelet transform (DNWT) in capturing image features combined with the good generalization ability of support vector regression provide us with a promising way to design a more robust watermarking algorithm featuring a better trade-off between the robustness and imperceptivity, the main duality of watermarking algorithms. Experimental results show that the DNWT watermarking scheme is robust to noising, JPEG compression, and cropping. In particular, it is more resistant to JPEG compression and noise than the discrete separable wavelet transform based scheme.

This work attempts to take advantage of the properties of Kohonen’s Self-Organizing Map (SOM) to perform the cluster analysis of remotely sensed images. A clustering method which automatically finds the number of clusters as well as the partitioning of the image data is proposed. The data clustering is made using the SOM. Different partitions of the trained SOM are obtained from different segmentations of the U-matrix (a neuron-distance image) that are generated by means of mathematical morphology techniques. The different partitions of the trained SOM produce different partitions for the image data which are evaluated by cluster validity indexes. Seeking to guarantee even greater efficiency in the image categorization process, the proposed method extracts information from the image by means of pixel windows, in order to incorporate textural information. The experimental results show an application example of the proposed method on a TM-Landsat image.

Weighted Voting Superposition (WeVoS) is a novel summarization algorithm that may be applied to the results of an ensemble of topology preserving maps in order to identify the lowest topographical error in a map and thereby, to calculate the best possible visualization of the internal structure of its datasets. It is applied in this research to the food industry field that is studying the olfactory properties of Spanish dry-cured ham. The datasets used for the analysis are taken from the readings of an electronic nose, a device that can be used to recognize the sensory smellprints of Spanish dry-cured ham samples. They are then automatically analyzed using the previously mentioned techniques, in order to detect those batches with an anomalous smell (acidity, rancidity and different type of taints).. The Weighted Voting Superposition of ensembles of Self-Organising Maps (SOMs) is used here for visualization purposes, and is compared with the simple version of the SOM. The results clearly demonstrate how the WeVoS-SOM outperforms the simple SOM method.

Laser milling is a relatively new micromanufacturing technique in the production of copper and other metallic components. This study presents multidisciplinary research, which is based on unsupervised connectionist architectures in conjunction with modelling systems, on the determination of the optimal operating conditions in this industrial process. Sensors on a laser milling centre relay the data used in this industrial case study of a machine-tool that manufactures copper components for high value micro-coolers. The two-phase application of the connectionist architectures is capable of identifying a model for the laser-milling process based on low-order models such as Black Box. The final system is capable of approximating the optimal form of the model. Finally, it is shown that the Box-Jenkins algorithm, which calculates the function of a linear system from its input and output samples, is the most appropriate model to control these industrial tasks.

As part of a multidisciplinary research project on relevant applications of Exploratory Projection Pursuit, this study sets out to examine levels of country and political risk that are assumed by a sample of Spanish Multinational Enterprises (MNEs). It analyses information pertaining to points such as decisions over the localization of subsidiary firms in various regions across the world, the importance accorded to such decisions and the driving forces behind them. The specific variables under study are economic freedoms, perceived levels of corruption and the constraints affecting the host governments in a sample of 1773 Spanish MNE subsidiaries throughout the world. Several neural projection models are applied, and we are able to conclude that these connectionist techniques help analyse the relevant data to identify the internationalization strategies of Spanish MNEs, their underlying motives and the goals they pursue.

This paper presents a novel neural network with only one layer which can compete with multi-layer neural nets. This novel neural net is called a double-threshold single-layer neural net. The theoretical analysis and experiments show that it can demonstrate similar performance as multi-layer neural nets.

In many face recognition and other classification applications, there exist unlabelled data available for training along with labelled data. The use of unlabelled data can improve the performance of the classifier. In this paper, a semi-supervised growing neural gas is proposed for such applications. The classifier is first trained on the labelled data and then gradually unlabelled data is classified and added to the training data. The proposed algorithm is demonstrated, on both artificial and real datasets, to significantly boost the classification rate with the use of unlabelled data. The improvement is particularly great when the labelled dataset is small. The algorithm is computationally simple and easy to implement.