Search

Testing interactions in multi-agent systems is a complex task because of several reasons. Agents are distributed and can move through different nodes in a network, so their interactions can occur concurrently and from many different sites. Also, agents are autonomous entities with a variety of possible behaviours, which can evolve during their lives by adapting to changes in the environment and new interaction patterns. Furthermore, the number of agents can vary during system execution, from a few dozens to thousands or more. Therefore, the number of interactions can be huge and it is difficult to follow up their occurrence and relationships. In order to solve these issues we propose the use of a set of data mining tools, the ACLAnalyser, which processes the results of the execution of large scale multi-agent systems in a monitored environment. This has been integrated with an agent development toolset, the INGENIAS Development Kit, in order to facilitate the verification of multi-agent system models at the design level rather than at the programming level. [ABSTRACT FROM AUTHOR]

In this work, we propose a new greedy clustering algorithm to identify groups of related genes. Clustering algorithms analyze genes in order to group those with similar behavior. Instead, our approach groups pairs of genes that present similar positive and/or negative interactions. Our approach presents some interesting properties. For instance, the user can specify how the range of each gene is going to be segmented (labels). Some of these will mean expressed or inhibited (depending on the gradation). From all the label combinations a function transforms each pair of labels into another one, that identifies the type of interaction. From these pairs of genes and their interactions we build clusters in a greedy, iterative fashion, as two pairs of genes will be similar if they have the same amount of relevant interactions. Initial two-genes clusters grow iteratively based on their neighborhood until the set of clusters does not change. The algorithm allows the researcher to modify all the criteria: discretization mapping function, gene-gene mapping function and filtering function, and provides much flexibility to obtain clusters based on the level of precision needed. The performance of our approach is experimentally tested on the yeast dataset. The final number of clusters is low and genes within show a significant level of cohesion, as it is shown graphically in the experiments. [ABSTRACT FROM AUTHOR]

Multiple teleoperated robots are more powerful to accomplish complex tasks than single robot does, and coordinated control of multiple robots play an important role in improving their performances. A knowledge based intelligent control method is proposed to coordinate the autonomous actions of the two teleoperated robots. In this method, each intelligent agent of the teleoperated robots has a model of other robots and a knowledge base of the task in a virtual distributed graphic predictive display system. Based on minimum distance query and collision detection, the autonomous motion of each robot is then planned via knowledge based reasoning to accomplish the given task. Experimental simulation results show that the method is effective for multiple robots to work coordinately. Keywords: Multiple robots, Virtual reality, Intelligent control. [ABSTRACT FROM AUTHOR]

The purpose of this paper is to solve Quality-of-Service (QoS) multicast routing problem by Particle Swarm Optimization (PSO). The QoS multicast routing optimization problem was transformed into a quasi-continuous problem by constructing a new integer coding and the constrained conditions in the problem were solved by the method of penalty function. The experimental results indicated that the proposed algorithm could converge to the optimal on near-optimal solution with less computational cost. It also appeared that PSO outperformed Genetic Algorithm on QoS the tested multicast routing problem. Keywords: Quality-of Service, Multicast Routing, Particle swarm optimization. [ABSTRACT FROM AUTHOR]

Conventional frequent pattern mining algorithms require users to specify some minimum support threshold, which is not easy to identify without knowledge about the datasets in advance. This difficulty leads users to dilemma that either they may lose useful information or may not be able to screen for the interesting knowledge from huge presented frequent patterns sets. Mining top-k frequent patterns allows users to control the number of patterns to be discovered for analyzing. In this paper, we propose an optimized version of the ExMiner, called OExMiner, to mine the top-k frequent patterns from a large scale dataset efficiently and effectively. In order to improve the user-friendliness and also the performance of the system we proposed other 2 methods, extended from OExMiner, called Seq-Miner and Seq-BOMA to mine top-k frequent patterns sequentially. Experiments on both synthetic and real data show that our proposed methods are much more efficient and effective compared to the existing ones. [ABSTRACT FROM AUTHOR]

We proposed a new algorithm for packet routing problems using chaotic neurodynamics and analyze its statistical behavior. First, we construct a basic neural network which works in the same way as the Dijkstra algorithm that uses information of shortest path lengths from a node to another node in a computer network. When the computer network has a regular topology, the basic routing method works well. However, when the computer network has an irregular topology, it fails to work, because most of packets cannot be transmitted to their destinations due to packet congestion in the computer network. To avoid such an undesirable problem, we extended the basic neural network to employ chaotic neurodynamics. We confirm that our proposed method exhibits good performance for complex networks, such as scale-free networks. [ABSTRACT FROM AUTHOR]

To evaluate predictability of complex behavior produced from nonlinear dynamical systems, we often use normalized root mean square error, which is suitable to evaluate errors between true points and predicted points. However, it is also important to estimate prediction intervals, where the future point will be included. Although estimation of prediction intervals is conventionally realized by an ensemble prediction, we applied the bootstrap resampling scheme to evaluate prediction intervals of nonlinear time-series. By several numerical simulations, we show that the bootstrap method is effective to estimate prediction intervals for nonlinear time-series. [ABSTRACT FROM AUTHOR]

The intuitionistic fuzzy set, developed by Atanassov [1], is a useful tool to deal with vagueness and uncertainty.Correlation analysis of intuitionistic fuzzy sets is an important research topic in the intuitionistic fuzzy set theory and has great practical potential in a variety of areas, such as engineering, decision making, medical diagnosis, pattern recognition, etc. In this paper, we propose a new method for deriving the correlation coefficients of intuitionistic fuzzy sets, which has some advantages over the existing methods. Furthermore, we extend the developed method to the interval-valued intuitionistic fuzzy set theory, and show its application in medical diagnosis. [ABSTRACT FROM AUTHOR]

MOVICAB-IDS enables the more interesting projections of a massive traffic data set to be analysed, thereby providing an overview of any possible anomalous situations taking place on a computer network. This IDS responds to the challenges presented by traffic volume and diversity. It is a connectionist agent-based model extended by means of a functional and mobile visualization interface. The IDS is designed to be more flexible, accessible and portable by running on a great variety of applications, including small mobile ones such as PDA's, mobile phones or embedded devices. Furthermore, its effectiveness has been demonstrated in different tests. Keywords: Unsupervised Learning, Neural Networks, Exploratory Projection Pursuit, Multiagent Systems, Computer Network Security, Intrusion Detection. [ABSTRACT FROM AUTHOR]

A data mining procedure for automatic determination of fuzzy decision tree structure using a genetic program is discussed. A genetic program (GP) is an algorithm that evolves other algorithms or mathematical expressions. Methods for accelerating convergence of the data mining procedure are examined. The methods include introducing fuzzy rules into the GP and a new innovation based on computer algebra. Experimental results related to using computer algebra are given. Comparisons between trees created using a genetic program and those constructed solely by interviewing experts are made. Connections to past GP based data mining procedures for evolving fuzzy decision trees are established. Finally, experimental methods that have been used to validate the data mining algorithm are discussed. Keywords: Genetic Programs, Fuzzy Logic, Data Mining, Control Algorithms, Planning Algorithms. [ABSTRACT FROM AUTHOR]

The fuzzy approach to regression has been traditionally considered as a problem of linear programming. In this work, we introduce a variety of models founded on quadratic programming together with a set of indices useful to check the quality of the obtained results. In order to test the validness of our proposal, we have done an empirical study and we have applied the models in a case with financial data: the Chilean COPEC Company stock price. [ABSTRACT FROM AUTHOR]

We propose a classification method based on a special class of feed-forward neural network, namely product-unit neural networks. They are based on multiplicative nodes instead of additive ones, where the nonlinear basis functions express the possible strong interactions between variables. We apply an evolutionary algorithm to determine the basic structure of the product-unit model and to estimate the coefficients of the model. We use softmax transformation as the decision rule and the cross-entropy error function because of its probabilistic interpretation. The empirical results over four benchmark data sets show that the proposed model is very promising in terms of classification accuracy and the complexity of the classifier, yielding a state-of-the-art performance. Keywords: Classification; Product-Unit; Evolutionary Neural Networks. [ABSTRACT FROM AUTHOR]

In the last years, much work have been done in the development of techniques for automated negotiation, and, particularly, in the automated negotiation of SLAs. However, there is no work that describes how to develop advanced software systems that are able to negotiate automatically in an open environment such as the Internet. In this work, we develop a conceptual framework for automated negotiations of SLAs that serves as a starting point by identifying the elements that must be supported in those software systems. In addition, based on that conceptual framework, we report on a set of properties for automated negotiation systems that may be used to compare different proposals. [ABSTRACT FROM AUTHOR]

The growing interest in mobile devices in new different scenarios increases the need to create context-aware multi-agent systems. Interaction between users is produced in many environments and new mobility technologies allow access information anytime and anywhere. Many researches and location systems have focused in specific domains to share knowledge between agents. This article proposes a global ontology to let agents work with heterogeneous domains using a wireless network. The intention of the proposed ontology is to provide customization about different environment services based on user location and profile. Keywords: ontology, heterogeneous domain, multi-agent system, wireless network. [ABSTRACT FROM AUTHOR]

Peer-to-peer computing has been emerged as very popular application due to the strong retrieval performance and the easiness of sharing resource and information. Nonetheless in reality p2p users are demanding more privacy rather than share of information. In this study we propose mutual anonymity technique based on length-restricted multi-path for identity of initiator and responder and the anonymous communication between initiator and responder. In this technique is not easily revealed to denial of service attack by ensuring the anonymity of initiator and responder and anonymous communication is possible by means of grouping technique of intermediate nodes. Compared with the previous studies privacy can easily be ensured without causing cryptography processing overhead by using former protocol only. Overhead, expected in the proposed technique of intermediate node grouping, is evaluated through test. [ABSTRACT FROM AUTHOR]

This article presents a hierarchical planner to actuate in uncertain domains named HIPU - Hierarchical Planner under Uncertainty. The uncertainties treated by HIPU include the probabilistic distribution of the operators effects and a distribution of probabilities on the possible initial states of the domain. HIPU automatically creates the abstraction hierarchy that will be used during planning, and for this it uses an extension of the Alpine method, adapted to act under uncertainty conditions. The planning process in HIPU happens initially in the highest level of abstraction, and the solution found in this level is refined by lower levels, until reaching the lowest level. During the search the plan evaluation is carried out, indicating if the plan achieves the goal with a success probability larger or equal to a previously defined value. To evaluate this probability, the planner uses the forward projection method. Keywords: planning, reasoning under uncertainty, machine learning. [ABSTRACT FROM AUTHOR]

We address the problem of identifying a non-redundant subset of important variables. All modern feature selection approaches including filters, wrappers, and embedded methods experience problems in very general settings with massive mixed-type data, and with complex relationships between the inputs and the target. We propose an efficient ensemble-based approach measuring statistical independence between a target and a potentially very large number of inputs including any meaningful order of interactions between them, removing redundancies from the relevant ones, and finally ranking variables in the identified minimum feature set. Experiments with synthetic data illustrate the sensitivity and the selectivity of the method, whereas the scalability of the method is demonstrated with a real car sensor data base. [ABSTRACT FROM AUTHOR]

Automatic text categorization requires the construction of appropriate surrogates for documents within a text collection. The surrogates, often called document vectors, are used to train learning systems for categorising unseen documents. A comparison of different measures (tfidf and weirdness) for creating document vectors is presented together with two different state-of-theart classifiers: supervised Kohonen's SOFM and unsupervised Vapniak's SVM. The methods are tested using two ‘gold standard' document collections and one data set from a ‘real-world' news stream. There appears to be an optimal size both for the of document vectors and for the dimensionality of each vector that gives the best compromise between categorization accuracy and training time. The performance of each of the classifiers was computed for five different surrogate vector models: the first two surrogates were created with tfidf and weirdness measures accordingly, the third surrogate was created purely on the basis of high-frequency words in the training corpus, and the fourth vector model was created from a standardised terminology database. Finally, the fifth surrogate (used for evaluation purposes) was based on a random selection of words from the training corpus. Keywords: Feature selection, Text categorisation, Information extraction, Selforganizing feature map (SOFM), Support vector machine (SVM). [ABSTRACT FROM AUTHOR]

We develop and investigate probabilistic approaches of state clustering in higher-order Markov chains. A direct extension of the Aggregate Markov model to higher orders turns out to be problematic due to the large number of parameters required. However, in many cases, the events in the finite memory are not equally salient in terms of their predictive value. We exploit this to reduce the number of parameters. We use a hidden variable to infer which of the past events is the most predictive and develop two different mixed-order approximations of the higher-order aggregate Markov model. We apply these models to the problem of community identification from event sequences produced through online computer-mediated interactions. Our approach bypasses the limitations of static approaches and offers a flexible modelling tool, able to reveal novel and insightful structural aspects of online interaction dynamics. [ABSTRACT FROM AUTHOR]

Automatic creation of syntactic and semantic word categorizations is a challenging problem for highly inflecting languages due to excessive data sparsity. Moreover, the study of colloquial language resources requires the utilization of fully corpus-based tools. We present a completely automated approach for producing word categorizations for morphologically rich languages. Self-Organizing Map (SOM) is utilized for clustering words based on the morphological properties of the context words. These properties are extracted using an automated morphological segmentation algorithm called Morfessor. Our experiments on a colloquial Finnish corpus of stories told by young children show that utilizing unsupervised morphs as features leads to clearly improved clusterings when compared to the use of whole context words as features. [ABSTRACT FROM AUTHOR]

In data mining problems, the selection of appropriate input transformations is often crucial to obtain good solutions. The purpose of such transformations is to project the original attribute space onto a new one that, being closer to the problem structure, allows for more compact and interpretable solutions. We address the problem of automatic construction of input transformations in classification problems. We use an evolutionary approach to search the space of input transformations and a linear method to perform classification on the new feature space. Our assumption is that once a proper data representation, which captures the problem structure, is found, even a linear classifier may find a good solution. Linear methods are free from local minima, while the use of a representation space closer to the problem structure will in general provide more compact and interpretable solutions. We test our method using an artificial problem and a real classification problem from the UCI database. In both cases we obtain low error solutions that in addition are compact and interpretable. [ABSTRACT FROM AUTHOR]

In this study, we propose a new classification framework, CARSVM model, which integrates association rule mining and support vector machine. The aim is to take advantages of both knowledge represented by class association rules and the power of SVM algorithm to construct an efficient and accurate classifier model. Instead of using the original training set, a set of rule-based feature vectors, which are generated based on the discriminative ability of class association rules over the training samples, are presented to the learning process of the SVM algorithm. The reported test results demonstrate the applicability, efficiency and effectiveness of the proposed model. Keywords: classification, association rule mining, associative classifiers, class association rules, support vector machine, machine learning. [ABSTRACT FROM AUTHOR]

Three methods are investigated and presented for automated learning of Relevance Vector Machines (RVM) in large scale text sets. RVM probabilistic Bayesian nature allows both predictive distributions on test instances and model-based selection yielding a parsimonious solution. However, scaling up the algorithm is not workable in most digital information processing applications. We look at the properties of the baseline RVM algorithm and propose new scaling approaches based on choosing appropriate working sets which retain the most informative data. Incremental, ensemble and boosting algorithms are deployed to improve classification performance by taking advantage of the large training set available. Results on Reuters-21578 are presented, showing performance gains and maintaining sparse solutions that can be deployed in distributed environments. [ABSTRACT FROM AUTHOR]

Context recognition in indoor and outdoor surroundings is an important area of research for the development of autonomous systems. This work describes an approach to the classification of audio signals found in both indoor and outdoor environments. Several audio features are extracted from raw signals. We analyze the relevance and importance of these features and use that information to design a multi-stage classifier architecture. Our results show that the multi-stage classification scheme is superior than a single stage classifier and it generates an 80% success rate on a 7 class problem. [ABSTRACT FROM AUTHOR]

Due to the discarded attributes, the effectual condition classes of the decision rules are highly different. To provide a unified evaluative measure, the derivation of each rule is depicted by the reduced attributes with a layered manner. Therefore, the inconsistency is divided into two primary categories in terms of the reduced attributes. We introduce the notion of joint membership function wrt. the effectual joint attributes, and a classification method extended from the default decision generation framework is proposed to handle the inconsistency. [ABSTRACT FROM AUTHOR]

A novel shape descriptor, chord length function (CLF) which can be obtained by equal arc length partitions of a contour, is proposed. The difference of two shapes is measured by the distance between their corresponding CLF. The proposed CLF is invariant to rotation, scaling and translation. It is robust to noise and simple to compute. Experimental results indicate that CLF is an effective shape descriptor. [ABSTRACT FROM AUTHOR]

The uncertain regions are modeled by broad boundary regions, while the description and reasoning of combined spatial relations between them are still unresolved problems. So far, the research about describing spatial relations mainly focuses on single-kind of spatial relation between simple and certain objects, rather than the combination of multi-kinds of spatial relations between uncertain regions. In addition, the uncertainty of formalizing combined spatial relations between uncertain regions is still a puzzle problem. Based on this point, the rough set theory is introduced into describing combined spatial relations between broad boundary regions. First, topological relations, exterior direction relations and detailed direction relations are approximated by a lower and an upper rough set, respectively, and then the difference set between the two sets indicates the uncertainty of spatial relations. Finally, the combination of the topological and direction relations between broad boundary regions is used to describe combined spatial relations. Keywords: Broad boundary region, direction relation, topological relation. [ABSTRACT FROM AUTHOR]

Identification of the footwear traces from crime scenes is an important yet largely forgotten aspect of forensic intelligence and evidence. We present initial results from a developing automatic footwear classification system. The underlying methodology is based on large numbers of localized features located using MSER feature detectors. These features are transformed into robust SIFT or GLOH descriptors with the ranked correspondence between footwear patterns obtained through the use of constrained spectral correspondence methods. For a reference dataset of 368 different footwear patterns, we obtain a first rank performance of 85% for full impressions and 84% for partial impressions. [ABSTRACT FROM AUTHOR]

Qualitative knowledge reasoning is a key content in knowledge science. Case-based reasoning is one of the main reasoning methodologies in artificial intelligence. Outranking relation methods, called ELECTRE and others, have been developed. In this research, a new algorithm of similarity measuring for qualitative problems in the presence of multiple experts based on outranking relations in case-based reasoning was proposed. Strict preference, weak preference, and indifference relations were introduced to formulate imprecision, uncertainty, incompleteness knowledge from multi-experts. Case similarities were integrated through aggregating house on the foundation of outranking relations. Experiments indicated that the new algorithm got accordant outcome with traditional quantitative similarity mode but extended its application range. [ABSTRACT FROM AUTHOR]

Statistical learning theory make large margins an important property of linear classifiers and Support Vector Machines were designed with this target in mind. However, it has been shown that large margins can also be obtained when much simpler kernel perceptrons are used together with ad-hoc updating rules, different in principle from Rosenblatt's rule. In this work we will numerically demonstrate that, rewritten in a convex update setting and using an appropriate updating vector selection procedure, Rosenblatt's rule does indeed provide maximum margins for kernel perceptrons, although with a convergence slower than that achieved by other more sophisticated methods, such as the Schlesinger-Kozinec (SK) algorithm. [ABSTRACT FROM AUTHOR]

The discovery of non-taxonomical relationships is one of the less studied knowledge acquisition tasks, even though it is a crucial point in ontology learning. We present an automatic and unsupervised methodology for extracting non-taxonomically related concepts and labelling relationships, using the whole Web as learning corpus. We also discuss how the obtained relationships may be automatically evaluated, using relatedness measures based on WordNet. [ABSTRACT FROM AUTHOR]

Although various location-sensing techniques and services have been developed, most of the conventional location-based services provide only static service. They do not consider user's preference but only a current location. Considering the trajectory might help to understand the user's intention and to provide a proper service. We propose a novel method that predicts user's mobility to provide service corresponding to the intention. The user's movement trajectory is analyzed by two stage modeling of recurrent self-organizing maps (RSOM) and Markov models. Using a GPS data set collected on the campus of Yonsei University, we have verified the usefulness of the proposed method. [ABSTRACT FROM AUTHOR]

A problem of learning from a database where each sample consists of several time series and a single response is considered. We are interested in maximum data reduction that preserves predictive power of the original time series, and at the same time allows reasonable reconstruction quality of the original signals. Each signal is decomposed into a set of wavelet features that are coded according to their importance consisting of two terms. The first depends on the influence of the feature on the expected signal reconstruction error, and the second is determined by feature importance for the response prediction. The latter is calculated by building series of boosted decision tree ensembles. We demonstrate that such combination maintains small signal distortion rates, and ensures no increase in the prediction error in contrast to the unsupervised compression with the same reduction ratio. [ABSTRACT FROM AUTHOR]

Vis/NIRs technique can be used in non-destructive measurement of the material internal quality in many fields. In this study, a mixed algorithm combined with back-propagation neural networks (BPNNs) and partial least squares (PLS) method was applied in the predicting the acidity of yogurt. The reflectance of optimal wavebands selected by PLS process were set as input neurons of BPNNs to establish the prediction model. By training the 130 yogurt samples in the BPNNs of topological structure 19:11:1, the acidity of the remaining 25 samples were predicted. The correlation between the measured and predicted values shows an excellent prediction performance with the value of 0.97, higher than the result (0.916) obtained only by PLS. Thus, it is concluded that the algorithm construct by BPNNs combined with partial least square applied to pattern recognition is an available alternative for pattern recognition based on Vis/NIRs. [ABSTRACT FROM AUTHOR]

Radar detection of targets in clutter and noise is an usual problem presented in radar systems. Several schemes based on statistical signal processing are proposed as detectors. In some cases, the Neural Networks (NNs) are applied to this problem. In this article, a radar detector based in a class of NN, the MultiLayer Perceptron (MLP), is proposed. This MLP can be trained in a supervised way to minimize the Mean Square Error (MSE) criterion. Moreover, it is demonstrated that the MLP trained in that way approximates the Neyman-Pearson detector. The NN-based detector proposed is compared with a Target Sequence Known A Priori (TSKAP) detector. The last detector is only took as reference because it is not realizable due to it is necessary to know when the target exists and its magnitude and shape. The results show how the proposed detector improves the performance of the TSKAP one for different conditions of the target measured with the Signal-to-Noise Ratio (SNR) and the skewness or shape parameter (a) of the Weibull-distributed clutter. Finally, several figures show which is the improvement of the NN-based detector. [ABSTRACT FROM AUTHOR]

Cyclic Dynamic Time Warping (CDTW) is a good measure of contour shapes dissimilarity, but it is computationally expensive. We introduce a lower bound for CDTW inspired in the Bunke and Bühler that leads to a significant speed up of contours classification in real tasks, as the experiments show. [ABSTRACT FROM AUTHOR]

Feature subset selection using a wrapper means to perform a search for an optimal set of attributes using the Machine Learning Algorithm as a black box. The Naive Bayes Classifier is based on the assumption of independence among the values of the attributes given the class value. Consequently, its effectiveness may decrease when the attributes are interdependent. We present FBL, a wrapper that uses information about dependencies to guide the search for the optimal subset of features and we use the Naive Bayes Classifier as the black-box Machine Learning algorithm. Experimental results show that FBL allows the Naive Bayes Classifier to achieve greater accuracies, and that FBL performs better than other classical filters and wrappers. Keywords: Naive Bayes, (In)Dependent Attributes, Wrapper, Feature Subset Selection, Machine Learning, Data Analysis, Data Mining. [ABSTRACT FROM AUTHOR]

We suggest and implement WPS (Wearable Personal Station) and Web-based robust Language Processing Interface (LPI) integrating speech and sign language (the Korean Standard Sign Language; KSSL). In other word, the LPI is integration language recognition and processing system that can select suitable language recognition system according to noise degree in given noise environment, and it is extended into embedded and ubiquitous-oriented the next generation language processing system that can take the place of a traditional uni-modal language recognition system using only 1 sensory channel based on desk-top PC and wire communication net. In experiment results, while an average recognition rate of uni-modal recognizer using KSSL only is 92.58% and speech only is 93.28%, advanced LPI deduced an average recognition rate of 95.09% for 52 sentential recognition models. Also, average recognition time is 0.3 seconds in LPI. [ABSTRACT FROM AUTHOR]

This study further investigates the complete-basis-functions parameterization method (CBFP) for Evolution Strategies (ES), and its application to a challenging real-life high-dimensional physics optimization problem, namely Femtosecond Laser Pulse Shaping. The CBFP method, which was introduced recently for tackling efficiently the learning task of n-variables functions, is combined here, for the first time, with niching techniques, and shown to boost the learning process of the given laser problem, and to yield satisfying multiple optima. Moreover, a technique for learning the basis-functions and improving this method is outlined. [ABSTRACT FROM AUTHOR]

Saplings Growing up Algorithm (SGA) is a novel computational intelligence method inspired by sowing and growing up of saplings. This method contains two phases: Sowing Phase and Growing up Phase. Uniformed sowing sampling is aim to scatter evenly in the feasible solution space. Growing up phase contains three operators: mating, branching, and vaccinating operator. In this study thinking capability of SGA has been defined and it has been demonstrated that sapling population generated initially has diversity. The similarity of population concludes the interaction of saplings and at consequent, they will be similar. Furthermore, the operators used in the algorithm uses similarity and hence the population has the convergence property. [ABSTRACT FROM AUTHOR]

A new approach is presented for the approximation of a scalar function defined on a discrete set of points. The method is based on the application of functional networks and the Lagrange interpolation formula. The interpolation mechanism of the separable functional networks when the neuron functions are approximated by Lagrange polynomials, is explored. The coefficients of the Lagrange interpolation formula are estimated during the learning of the functional network by simply solving a linear system of equations. Finally, several examples show the effectiveness of the proposed interpolation method. [ABSTRACT FROM AUTHOR]

In real-world applications, it has been often observed that class imbalance (significant differences in class prior probabilities) may produce an important deterioration of the classifier performance, in particular with patterns belonging to the less represented classes. This effect becomes especially significant on instance-based learning due to the use of some dissimilarity measure. We analyze the effects of class imbalance on the classifier performance and how the overlap has influence on such an effect, as well as on several techniques proposed in the literature to tackle the class imbalance. Besides, we study how these methods affect to the performance on both classes, not only on the minority class as usual. [ABSTRACT FROM AUTHOR]

Neural networks (NNs) are proposed for approximating the Average Likelihood Ratio (ALR). The detection of gaussian targets with gaussian autocorrelation function and unknown one-lag correlation coefficient, ρs, in Additive White Gaussian Noise (AWGN) is considered. After proving the low robustness of the likelihood ratio (LR) detector with respect to ρs, the ALR detector assuming a uniform distribution of this parameter in [0,1] has been studied. Due to the complexity of the involved integral, two NN based solutions are proposed. Firstly, single Multi-Layer Perceptrons (MLPs) are trained with target patterns with ρs varying in [0,1]. This scheme outperforms the LR detector designed for a fixed value of ρs. MLP with 17 hidden neurons is proposed as a solution. Then, two MLPs trained with target patterns with ρs varying in [0,0.5] and [0.5,1], respectively, are combined. This scheme outperforms the single MLP and allows to determine a solution of compromise between complexity and approximation error. A detector composed of MLPs with 17 and 8 hidden units each one is proposed. [ABSTRACT FROM AUTHOR]

We propose an enhanced QA model with combination of hetero-geneous answer acquisition methods. Our QA system is based on web encyclopedia in Korean. We investigated characteristic features of the encyclopedia and incorporate them in our answer acquisition methods. We defined three different types of answer extraction methods: learning-based, pattern-based, and traditional statistical methods. By empirical experiments, we obtained 59% improvement on MRR as well as 2.3 times speedy response. Keywords: Question Answering, Knowledge acquisition, Machine learning. [ABSTRACT FROM AUTHOR]

We review the performance function associated with the familiar K-Means algorithm and that of the recently developed K-Harmonic Means. The inadequacies in these algorithms leads us to investigate a family of performance functions which exhibit superior clustering on a variety of data sets over a number of different initial conditions. In each case, we derive a fixed point algorithm for convergence by finding the fixed point of the first derivative of the performance function. We give illustrative results on a variety of data sets. We show how one of the algorithms may be extended to create a new topology-preserving mapping. [ABSTRACT FROM AUTHOR]

This work presents a new classification method based on the iterative combination of two steps: a clustering technique and a set of one-layer neural networks. First, the clustering algorithm divides the input space in several regions (local models). Subsequently, a one-layer neural network, for each local region, is used to fit the model (classifier) for a specific group of data points. Experimental results on three different data sets are showed to verify the validity of the proposed method. Besides, a comparative study with a feedforward neural network is included. This study exhibits that the presented algorithm is a fast procedure that obtains, in many cases, better results than the other technique. [ABSTRACT FROM AUTHOR]

Measuring the length of a path that a taxi must fare is an obvious task: when driving lower than certain speed threshold the fare is time dependent, but at higher speeds the length of the path is measured, and the fare depends on such measure. When passing an indoor MOT test, the taximeter is calibrated simulating a cab run, while the taxi is placed on a device equipped with four rotating steel cylinders in touch with the drive wheels. This indoor measure might be inaccurate, as the information given by the cylinders is affected by tires inflating pressure, and only straight trajectories are tested. Moreover, modern vehicles with driving aids such as ABS, ESP or TCS might have their electronics damaged in the test, since two wheels are spinning while the others are not. To surpass these problems, we have designed a small, portable GPS sensor that periodically logs the coordinates of the vehicle and computes the length of a discretionary circuit. We will show that all the legal issues with the tolerance of such a procedure (GPS data are inherently imprecise) can be overcome if genetic and fuzzy techniques are used to process and analyze the raw data. [ABSTRACT FROM AUTHOR]

In telecom industry high installation and marketing costs make it between six to ten times more expensive to acquire a new customer than it is to retain the existing one. Prediction and prevention of customer churn is therefore a key priority for industrial research. While all the motives of customer decision to churn are highly uncertain there is lots of related temporal data sequences generated as a result of customer interaction with the service provider. Existing churn prediction methods like decision tree typically just classify customers into churners or non-churners while completely ignoring the timing of churn event. Given histories of other customers and the current customer's data, the presented model proposes a new k nearest sequence (kNS) algorithm along with temporal sequence fusion technique to predict the whole remaining customer data sequence path up to the churn event. It is experimentally demonstrated that the new model better exploits time-ordered customer data sequences and surpasses the existing churn prediction methods in terms of performance and offered capabilities. [ABSTRACT FROM AUTHOR]

The main objective of the present work is to propose and evaluate a neural stochastic optimization framework for reservoir parameter estimation, for which a history matching procedure is implemented by combining three independent sources of spatial and temporal information: production data, time-lapse seismic and sensor information. In order to efficiently perform large-scale parameter estimation, a coupled multilevel, stochastic and learning search methodology is proposed. At a given resolution level, the parameter space is globally explored and sampled by the simultaneous perturbation stochastic approximation (SPSA) algorithm. The estimation and sampling performed by SPSA is further enhanced by a neural learning engine that evaluates the objective function sensitiveness with respect to parameter estimates in the vicinity of the most promising optimal solutions. [ABSTRACT FROM AUTHOR]

NEH (Nawaz et al., 1983) is an effective heuristic for solving the permutation flowshop problem with the objective of makespan. It includes two phases. PE-VM (Li et al., 2004), which is based on NEH, is analyzed and we conclude that its good performance is partially due to the priority rule it used in phase I. Then, we propose a new measure to solve job insertion ties which may arise in phase II. The measure is based on the idea of balancing the utilization among all machines. Finally, we propose a heuristic NEH-D (NEH based on Deviation), which combines the priority rule of PE-VM and the new measure. Computational results show that the measure is effective and NEH-D is better than PE-VM. There is one parameter in NEH-D, and we design an experiment to try to find a near optimized value of it for tuning the performance of NEH-D. [ABSTRACT FROM AUTHOR]