Engineering Applications of Neural Networks

Robots operating in complex environments shared with humans are confronted with numerous problems. One important problem is the identification of obstacles and interaction partners. In order to reach this goal, it can be beneficial to use data from multiple available sources, which need to be processed appropriately. Furthermore, such environments are not static. Therefore, the robot needs to learn novel objects. In this paper, we propose a method for learning and identifying obstacles based on multi-modal information. As this approach is based on Adaptive Resonance Theory networks, it is inherently capable of incremental online learning.

This paper presents a mobile control system capable of learn behaviors based on human examples. Our approach is based on image processing, template matching, finite state machine, and template memory. The system proposed allows image segmentation using neural networks in order to identify navigable and non-navigable regions. It also uses supervised learning techniques which work with different levels of memory of the templates. As output our system is capable controlling speed and steering for autonomous mobile robot navigation. Experimental tests have been carried out to evaluate the learning techniques.

An application of clustering methods for production planning is proposed. Hierarchical clustering,

k

-means and SOM clustering are applied to production data from the company KGL in Slovenia. A database of 252 products manufactured in the company is clustered according to the required operations and product features. Clustering results are evaluated with an average silhouette width for a total data set and the best result is obtained by SOM clustering. In order to make clustering results applicable to industrial production planning, a percentile measure for the interpretation of SOM clusters into the production cells is proposed. The results obtained can be considered as a recommendation for production floor planning that will optimize the production resources and minimize the work and material flow transfer between the production cells.

The identification of user and usage profiles in the built environment is of vital importance both for energy performance analysis and smart control purposes. Clustering tools are a suitable means as they are able to discover representative patterns from a myriad of collected data. In this work, the methodology of an eXclusive Self-Organizing Map (XSOM) is proposed as an evolution of a Kohonen map with outlier rejection capabilities. As will be shown, XSOM characteristics fit perfectly with the targeted application areas.

The Hypercube Neural Network Algorithm is a novel supervised method for classification. One hypercube is defined per class in the attribute space based on the training data. Each dimension of a hypercube is set to cover the full range of values in the class. The hypercube learning is therefore a rapid, one-shot form of learning. This paper presents three versions of the algorithm: hypercube without neurons; with simple neurons; and with adaptive activation function neurons. The methods are tested and evaluated on several diverse publically available data sets and compared with published results obtained on these data when using alternative methods.

Liquid State Machines constitute a powerful computational tool for carrying out complex real time computations on continuous input streams. Their performance is based on two properties, approximation and separation. While the former depends on the selection of class functions for the readout maps, the latter needs to be evaluated for a particular liquid architecture. In the current paper we show how the Fisher’s Discriminant Ratio can be used to effectively measure the separation of a Liquid State Machine. This measure is then used as a fitness function in an evolutionary framework that searches for suitable liquid properties and architectures in order to optimize the performance of the trained readouts. Evaluation results demonstrate the effectiveness of the proposed approach.

The biological vision system is far more efficient than machine vision system. This is due to the former has rich neural layers for representation and process. In order to obtain a non-task-dependent image representation schema, the early phase of neural vision mechanism is worth simulating. We design a neural model to simulate non-classical receptive field of ganglion cell and its local feedback control circuit, and find it can represent image, beyond pixel level, self-adaptively and regularly. The experimental results prove this method can represent image faithfully with low cost, and can produce a com-pact and abstract approximation to facilitate successive image segmentation as well as integration operation. This representation schema is good at extracting spatial relationship from different components of image, thus it can be applied to formalize image semantics. Further it can be applied to object recognition or image classification tasks in future.

The concept of

linear separability

is used in the theory of neural networks and pattern recognition methods. This term can be related to examination of learning sets (classes) separation by hyperplanes in a given feature space. The family of

K

disjoined learning sets can be transformed into K linearly separable sets by the ranked layer of binary classifiers. Problems of the ranked layers deigning are analyzed in the paper.

Using plain, but large multi-layer perceptrons, temporal eye-tracking gaze patterns of alphabetic and logographic L1 readers were successfully classified. The Eye-tracking data was fed directly into the networks, with no need for pre-processing. Classification rates up to 92% were achieved using MLPs with 4 hidden units. By classifying the gaze patterns of interaction partners, artificial systems are able to act adaptively in a broad variety of application fields.

We present a custom hardware system for image recognition, featuring a dimensionality reduction network and a classification stage. We use Bi-Directional PCA and Linear Discriminant Analysis for feature extraction, and classify based on Manhattan distances. Our FPGA-based implementation runs at 75MHz, consumes 157.24mW of power, and can classify a 61 ×49-pixel image in 143.7

μ

s, with a sustained throughput of more than 7,000 classifications per second. Compared to a software implementation on a workstation, our solution achieves the same classification performance (93.3% hit rate), with more than twice the throughput and more than an order of magnitud less power.

Color image processing systems are used for a variety of purposes including medical imaging. Basic image processing algorithms for enhancement, restoration, segmentation and classification are modified since color is represented as a vector instead of a scalar gray level variable. Color images are regarded as two-dimensional (2-D) vector fields defined on some color space (like for example the RGB space). In bibliography, operators utilizing several distance and similarity measures are adopted in order to quantify the common content of multidimensional color vectors. Self-Organizing Feature Maps (SOFMs) are extensively used for dimensionality reduction and rendering of inherent data structures. The proposed window-based SOFM uses as multidimensional inputs color vectors defined upon spatial windows in order to capture the correlation between color vectors in adjacent pixels. A 3x3 window is used for capturing color components in uniform color space (

L

*

u

*

v

*

). The neuron featuring the smallest distance is activated during training. Neighboring nodes of the SOFM are clustered according to their statistical similarity (using the Mahalanobis distance). Segmentation results suggest that clustered nodes represent populations of pixels in rather compact segments of the images featuring similar texture.

Neural networks have represented a serious barrier-to-entry in their application in automated fraud detection due to their black box and often proprietary nature which is overcome here by combining them with symbolic rule extraction. A Sparse Oracle-based Adaptive Rule extraction algorithm is used to produce comprehensible rules from a neural network to aid the detection of credit card fraud. In this paper, a method to improve this extraction algorithm is presented along with results from a large real-world European credit card data set. Through this application it is shown that neural networks can assist in mission-critical areas of business and are an important tool in the transparent detection of fraud.

Weather operations play an important and integral part of planning, execution and sustainment of mission operations. In this paper, a neuro-fuzzy hybridization technique is applied to model the weather operations and predict its impact on the effectiveness of air tasking operations and missions. Spatio-temporal weather data from various meteorological sources are collected and used as the input to a neural network and the predicted weather conditions at a given place is classified based on fuzzy logic. The corresponding fuzzy rules are generated forming the basis for introducing the weather conditions in the evaluation of the effectiveness of the military mission plans. An agent-based architecture is proposed where agents representing the various weather sensors feed the weather data to the simulator, and a weather agent developed using neuro-fuzzy hybridization computes the weather conditions over the flight plan of the mission. These rules are then used by the Mission Planning and Execution system that evaluates the effectiveness of military missions in various weather conditions.

Audio mining is currently the subject of several research efforts, especially because of its potential to speed up search for spoken words in audio recordings. This study explores a method for approaching the problem from the bottom. It proposes a framework based on smart logic, mainly fuzzy logic, and on an audio model applicable to any kind of audio recording, including music.

Raw cow milk has short supply market in summer and over supply in winter, which causes consumers and dairy industry concern about the quality of raw milk whether is adulated with reconstituted milk (powdered milk). This study prepared 307 raw cow milk samples with various adulteration ratios 0%, 2%, 5%, 10%, 20%, 30%, 50%, 75%, and 100% of powdered milk. Least square support vector machine (LS-SVM) was applied to calibrate the prediction model for adulteration ratio. Grid search approach was used to find the better value of network parameters of

γ

and

σ

2

. Results show that R

2

ranges from 0.9662 to 0.9777 for testing data set with plate surface and four concave regions. Scatter plot of testing data showed that adulteration ratio above 10% clearly differs from 0% samples.

This research effort aims in the estimation of Wood Loss Factor by employing Support Vector Machines. For this purpose experimental data for two different wood species were used. The estimation of the dielectric properties of wood was done by using various Kernel algorithms as a function of both ambient electro-thermal conditions applied during drying of wood and basic wood chemistry. Actually the best fit neural models that were developed in a previous effort of our research team were compared to the Kernels’ approaches in order to determine the optimal ones.

In this paper, we propose a method to compare and visualize spectrograms in a low dimensional space using manifold learning. This approach is divided in two steps: a data processing and dimensionality reduction stage and a feature extraction and a visualization stage. The procedure is applied on different types of data from a hot rolling process, with the aim to detect

chatter

. Results obtained suggest future developments and applications in hot rolling and other industrial processes.

Raw goat milk pricing is based on the milk quality especially on fat, solid not fat (SNF) and density. Therefore, there is a need of approach for composition quantization. This study applied radial basis function network (RBFN) to calibrate fat, SNF, and density with visible and near infrared spectra (400~2500 nm). To find the optimal parameters of goal error and spread used in RBFN, a response surface method (RSM) was employed. Results showed that with the optimal parameters suggested by RSM analysis, R

2

difference for training and testing data set was the smallest which indicated the model was less possible of overtraining or undertraining. The R

2

for testing set was 0.9569, 0.8420 and 0.8743 for fat, SNF and density, respectively, when optimal parameters were used in RBFN.

The main objective of transfer learning is to reuse knowledge acquired in a previous learned task, in order to enhance the learning procedure in a new and more complex task. Transfer learning comprises a suitable solution for speeding up the learning procedure in Reinforcement Learning tasks. In this work, we propose a novel method for transferring models to a hybrid reinforcement learning agent. The models of the transition and reward functions of a source task, will be transferred to a relevant but different target task. The learning algorithm of the target task’s agent takes a hybrid approach, implementing both model-free and model-based learning, in order to fully exploit the presence of a model. The empirical evaluation, of the proposed approach, demonstrated significant results and performance improvements in the 3D Mountain Car task, by successfully using the models generated from the standard 2D Mountain Car.

The goal of this study is to detect anomalous queries from network logs using a dimensionality reduction framework. The fequencies of 2-grams in queries are extracted to a feature matrix. Dimensionality reduction is done by applying diffusion maps. The method is adaptive and thus does not need training before analysis. We tested the method with data that includes normal and intrusive traffic to a web server. This approach finds all intrusions in the dataset.

The purpose of this work is to further study the relevance of accelerating the Monte-Carlo calculations for the gamma rays external radiotherapy through feed-forward neural networks. We have previously presented a parallel incremental algorithm that builds neural networks of reduced size, while providing high quality approximations of the dose deposit [4]. Our parallel algorithm consists in an optimized decomposition of the initial learning dataset (also called learning domain) in as much subsets as available processors. However, although that decomposition provides subsets of similar signal complexities, their sizes may be quite different, still implying potential differences in their learning times. This paper presents an efficient data extraction allowing a good and balanced training without any loss of signal information. As will be shown, the resulting irregular decomposition permits an important improvement in the learning time of the global network.

In the past decade or so we have been experiencing an extraordinary explosion of data volumes first in wireline networks and recently even in mobile wireless networks. Optimizing bandwidth utilization is critical for planning and deploying efficient networks that are capable of delivering new services like IPTV over cost-oriented implementations. Models of distributed content caching in the access network have been employed - for example - as analytical optimization tools in order to tackle associated problems. A modified

capacitated quality-of-service network

(QoS) model is proposed herein in order to optimize the placement of the sites of surrogate media servers (central offices-COs) on the access part of a content distribution network (CDN). The novelty of the proposed approach lies in the fact that

capacitated quality-of-service network

optimization is cast as an optimization problem over two rather than one optimization variables-objectives. Implementation cost and link delay as determined by capacity/utilization requirements are the optimization functionals-objectives. Optimization of the network architecture is carried out via a multiobjective evolutionary algorithm that encodes all possible edges between the first level aggregation points of the access network. Proper priorities are assigned to different types of traffic according to class of service. Two main services are considered, namely live broadcast/IPTV and video on demand services (VoD). The media servers/COs are incorporated into the infrastructure of the access nodes in a step-by-step fashion modifying the traffic requirements between source and sink nodes of the optimal configurations of the access network. The evolution of the Pareto front is investigated in each case.

In conservation and improvement programs of bovine livestock, an important parameter is morphological assessment, which consist of scoring an animal attending to its morphology, and is always performed by highly-qualified staff.

We present in this paper a system designed to help in morphological assessment, providing a score based on a lateral image of the cow. The system consist of two main parts. First, a feature extractor stage is used to reduce the information of the cow in the image to a set of parameters (describing the shape of the profile of the cow). For this stage, a model of the object is constructed by means of point distribution models (PDM), and later that model is used in the searching process within each image, that is carried out using genetic algorithms (GAs). Second, the parameters obtained are used in the following stage, where a multilayer perceptron is trained in order to provide the desired assessment, using the scores given by experts for selected cows.

The system has been tested with 124 images corresponding to 44 individuals of a special rustic breed, with very promising results, taking into account that the information contained in only one view of the cow is not complete.

We describe a web application that takes advantage of new computer vision techniques to allow the user to make searches based on visual similarity of color and texture related to the object of interest. We use a supervised neural network strategy to segment different classes of objects. A strength of this solution is the high speed in generalization of the trained neural networks, in order to obtain an object segmentation in real time. Information about the segmented object, such as color and texture, are extracted and indexed as text descriptions. Our case study is the online commercial offers domain where each offer is composed by text and images. Many successful experiments were done on real datasets in the fashion field.

We propose a novel supervised learning rule allowing the training of a precise input-output behavior to a spiking neuron. A single neuron can be trained to associate (map) different output spike trains to different multiple input spike trains. Spike trains are transformed into continuous functions through appropriate kernels and then Delta rule is applied. The main advantage of the method is its algorithmic simplicity promoting its straightforward application to building spiking neural networks (SNN) for engineering problems. We experimentally demonstrate on a synthetic benchmark problem the suitability of the method for spatio-temporal classification. The obtained results show promising efficiency and precision of the proposed method.

The electroencephalograph (EEG) signal is one of the most widely used signals in the biomedicine field due to its rich information about human tasks. This research study describes a new approach based on i) build reference models from a set of time series, based on the analysis of the events that they contain, is suitable for domains where the relevant information is concentrated in specific regions of the time series, known as events. In order to deal with events, each event is characterized by a set of attributes. ii) Discrete wavelet transform to the EEG data in order to extract temporal information in the form of changes in the frequency domain over time- that is they are able to extract non-stationary signals embedded in the noisy background of the human brain.

The performance of the model was evaluated in terms of training performance and classification accuracies and the results confirmed that the proposed scheme has potential in classifying the EEG signals.

This paper proposes an ensemble based approach for feature selection. We aim at overcoming the problem of parameter sensitivity of feature selection approaches. To do this we employ ensemble method. We get the results per different possible threshold values automatically in our algorithm. For each threshold value, we get a subset of features. We give a score to each feature in these subsets. Finally by use of ensemble method, we select the features which have the highest scores. This method is not a parameter sensitive one, and also it has been shown that using the method based on the fuzzy entropy results in more reliable selected features than the previous methods’. Empirical results show that although the efficacy of the method is not considerably decreased in most of cases, the method becomes free from setting of any parameter.

When solving a pattern classification problem, it is common to apply a feature extraction method as a pre-processing step, not only to reduce the computation complexity but also to obtain better classification performance by reducing the amount of irrelevant and redundant information in the data. In this study, we investigate a novel schema for linear feature extraction in classification problems. The method we have proposed is based on clustering technique to realize feature extraction. It focuses in identifying and transforming redundant information in the data. A new similarity measure-based trend analysis is devised to identify those features. The simulation results on face recognition show that the proposed method gives better or competitive results when compared to conventional unsupervised methods like PCA and ICA.

Estimation of future glucose concentration is important for diabetes management. To develop a model predictive control (MPC) system that measures the glucose concentration and automatically inject the amount of insulin needed to keep the glucose level within its normal range, the accuracy of the predicted glucose level and the longer prediction time are major factors affecting the performance of the control system. The predicted glucose values can be used for early hypoglycemic/hyperglycemic alarms for adjustment of insulin injections or insulin infusion rates of manual or automated pumps. Recent developments in continuous glucose monitoring (CGM) devices open new opportunities for glycemia management of diabetic patients. In this article a new technique, which uses a recurrent neural network (RNN) and data obtained from CGM device, is proposed to predict the future values of the glucose concentration for prediction horizons (PH) of 15, 30, 45, 60 minutes. The results of the proposed technique is evaluated and compared relative to that obtained from a feed forward neural network prediction model (NNM). Our results indicate that, the RNN is better in prediction than the NNM for the relatively long prediction horizons.

Medical image segmentation is considered a very important task for diagnostic and treatment-planning purposes. Accurate segmentation of medical images helps clinicians to clarify the type of the disease and facilitates the process of efficient treatment. In this paper, we propose two different approaches to segment breast ultrasound images using neural networks. In the first approach, we use scale invariant feature transform (SIFT) to calculate a set of descriptors for a set of points inside the image. These descriptors are used to train a supervised neural network. In the second approach, we use SIFT to detect a set of key points inside the image. Texture features are then extracted from a region around each point to train the network. This process is repeated multiple times to verify the generalization ability of the network. The average segmentation accuracy is calculated by comparing every segmented image with corresponding gold standard images marked by an expert.

A novel ontology based type 2 diabetes risk analysis system framework is described, which allows the creation of global knowledge representation (ontology) and personalized modeling for a decision support system. A computerized model focusing on organizing knowledge related to three chronic diseases and genes has been developed in an ontological representation that is able to identify interrelationships for the ontology-based personalized risk evaluation for chronic diseases. The personalized modeling is a process of model creation for a single person, based on their personal data and the information available in the ontology. A transductive neuro-fuzzy inference system with weighted data normalization is used to evaluate personalized risk for chronic disease. This approach aims to provide support for further discovery through the integration of the ontological representation to build an expert system in order to pinpoint genes of interest and relevant diet components.

Brain-Computer Interfaces (BCIs) are devices offering alternative means of communication when conventional means are permanently, or nonpermanently, impaired. The latter is commonly induced in general anesthesia and is necessary for the conduction of the surgery. However, in some cases it is possible that the patient regains consciousness during surgery, but cannot directly communicate this to the anesthetist due to the induced muscle paralysis. Therefore, a BCI-based device that monitors the spontaneous brain activity and alerts the anesthetist is an essential addition to routine surgery. In this paper the use of Permutation Entropy (PE) as a feature for ‘conscious’ and ‘unconscious’ brain state classification for a BCI-based anesthesia monitor is investigated. PE is a linear complexity measure that tracks changes in spontaneous brain activity resulting from the administration of anesthetic agents. The overall classification performance for 10 subjects, as assessed with a linear Support Vector Machine, exceeds 95%, indicating that PE is an appropriate feature for such a monitoring device.

A key physical property used in the description of a soil-water regime is a soil water retention curve, which shows the relationship between the water content and the water potential of the soil. Pedotransfer functions are based on the supposed dependence of the soil water content on the available soil characteristics. In this paper, artificial neural networks (ANNs) and support vector machines (SVMs) were used to estimate a drying branch of a water retention curve. The performance of the models are evaluated and compared in case study for the Zahorska Lowland in the Slovak Republic. The results obtained show that in this study the ANN model performs somewhat better and is easier to handle in determining pedotransfer functions than the SVM models.

A locally recurrent neural network based fault detection and isolation approach is presented. A model of the system under test is created by means of a dynamic neural network. The fault detection is performed on the basis of the statistical analysis of the residual provided by the estimated density shaping of residuals in the case of nominal value of all the parameters, made of a simply neural network. The approach is illustrated by using the Rössler hyperchaotic system.

This paper presents the application of Neural Networks for the interpolation of (critical frequency)

foF2

data over Cyprus in the presence of sporadic E layer which is a frequent phenomenon during summer months causing inevitable gaps in the

foF2

data series. This ionospheric characteristic (

foF2

) constitutes the most important parameter in HF (High Frequency) communications since it is used to derive the optimum operating frequency in HF links and therefore interpolating missing data is very important in preserving the data series which is used in long-term prediction procedures and models.

The paper presents modeling of steels strength characteristics in dependence from their alloying components quantities using neural networks as nonlinear approximation functions. Further, for optimization purpose the neural network models are used. The gradient descent algorithm based on utility function backpropagation through the models is applied. The approach is aimed at synthesis of steel alloys compositions with improved strength characteristics by solving multi-criteria optimization task. The obtained optimal alloying compositions fall into martenzite region of steels. They will be subject of further experimental testing in order to synthesize new steels with desired characteristics.

Intelligent agents that seek to automate various stages of the negotiation process are often enhanced with models of computational intelligence extending the cognitive abilities of the parties they represent. This paper is focused on predictive strategies employed by automated negotiators, and particularly those based on forecasting the counterpart’s responses. In this context a strategy supporting negotiations over multiple issues is presented and assessed. Various behaviors emerge with respect to negotiator’s attitude towards risk, resulting to different utility gains. Forecasting is conducted with the use of Multilayer Perceptrons (MLPs) and the training set is extracted online during the negotiation session. Two cases are examined: in the first separate MLPs are used for the estimations of each negotiable attribute, whereas in the second a single MLP is used to estimate the counterpart’s response. Experiments are conducted to search the architecture of the MLPs.

The use of Neural Networks for modeling systems has been widespread, in particular within areas where the great amount of available data and the complexity of the systems keeps the problem very unfriendly to treat following traditional data analysis methodologies. In the last two decades, small strain shear modulus became one of the most important geotechnical parameters to characterize soil stiffness. Finite element analysis have shown that in-situ stiffness of soils and rocks is much higher than was previously thought, and that stress-strain behaviour of these materials is non-linear in most cases with small strain levels, especially in the ground around retaining walls, foundations and tunnels typically in the order of 10

− 2

to 10

− 4

of strain. Although the best approach seems to be based in measuring seismic wave velocities, deriving the parameter through correlations with in-situ tests is usually considered very useful for design practice. In this work, a new approach using Neural Networks is proposed for sedimentary soils and the results are discussed and compared with some of the most common available methodologies for this evaluation.

NNIGnets is a freeware computer program which can be used for teaching, research or business applications, of Artificial Neural Networks (ANNs). This software includes presently several tools for the application and analysis of Multilayer Perceptrons (MLPs) and Radial Basis Functions (RBFs), such as stratified Cross-Validation, Learning Curves, Adjusted Rand Index, novel cost functions, and Vapnik–Chervonenkis (VC) dimension estimation, which are not usually found in other ANN software packages. NNIGnets was built following a software engineering approach which decouples operative from GUI functions, allowing an easy growth of the package. NNIGnets was tested by a variety of users, with different backgrounds and skills, who found it to be intuitive, complete and easy to use.

This talk summarizes several points that have been learned about applying Artificial Neural Networks in the chemical industry. Artificial Neural Networks are one of the major tools of

Empirical Process Modeling

, but not the only one. To properly assess the appropriate model complexity, combine information about both the

Training

and the

Test

data sets. A neural network, or any other empirical model, is better at making predictions than the comparison between modeled and observed data shows. Finally, it is important to exploit synergies with other disciplines and practitioners to stimulate the use of Neural Networks in industry.

Acute graft-versus-host disease (aGvHD) is a serious systemic complication of allogeneic hematopoietic stem cell transplantation (HSCT) that occurs when alloreactive donor-derived T cells recognize host-recipient antigens as foreign. The early events leading to GvHD seem to occur very soon, presumably within hours from the graft infusion. Therefore, when the first signs of aGvHD clinically manifest, the disease has been ongoing for several days at the cellular level, and the inflammatory cytokine cascade is fully activated. So, it comes as no surprise that to identify biomarker signatures for approaching this very complex task is a critical issue. In the past, we have already approached it through joint molecular and computational analyses of gene expression data proposing a computational framework for this disease. Notwithstanding this, there aren’t in literature quantitative measurements able to identify patterns or rules from these biomarkers or from aGvHD data, thus this is the first work about the issue. In this paper first we have applied different feature selection techniques, combined with different classifiers to detect the aGvHD at onset of clinical signs, then we have focused on the aGvHD scenario and in the knowledge discovery issue of the classification techniques used in the computational framework.

In Wireless Local Area Networks (WLAN), beacon, probe request and response messages are unprotected, so the information is visible to sniffers. Probe requests can be sent by anyone with a legitimate Media Access Control (MAC) address, as association to the network is not required at this stage. Legitimate MAC addresses can be easily spoofed to bypass Access Point (AP) access lists. Attackers take advantage of these vulnerabilities and send a flood of probe request frames which can lead to a Denial-of-Service (DoS) to legitimate stations. This paper discusses an intelligent approach to recognise probe request attacks in WLANs. The research investigates and analyses WLAN traffic captured on a home wireless network, and uses supervised feedforward neural network with 4 input neurons, 2 hidden layers and an output neuron to determine the results. The computer simulation results demonstrate that this approach improves detection of MAC spoofing and probe request attacks considerably.

Computer text entry may be full of noises – for example, computer keyboard users inevitably make typing mistakes and their typing stream implies all users’ self rectification actions. These may produce a great negative influence on the accessibility and usability of applications. This research develops an original Intelligent Keyboard hybrid framework, which can be used to analyze users’ typing stream, and accordingly correct typing mistakes and predict users typing intention. An extendable Focused Time-Delay Neural Network (FTDNN) n-gram prediction algorithm is developed to learn from the users’ typing history and produce text entry prediction and correction based on historical typing data. The results show that FTDNN is an efficient tool to model typing stream. Also, the computer simulation results demonstrate that the proposed framework performs better than using the conventional keyboard.

A mobile network provides a continuous stream of data describing the performance of its cells. Most of the data describes cells with acceptable performance. Detecting and analysing mobile network cells with quality problems from the data stream is a tedious and continuous problem for network operators. Anomaly detection can be used to identify cells, whose performance deviates from the average and which are potentially having some sub-optimal configuration or are in some error condition. In this paper we provide two methods to detect such anomalously behaving cells. The first method estimates the distance from a cell to an optimal state and the second one is based on detecting the support of the data distribution using One-Class Support Vector Machine (OC-SVM). We use the methods to analyse a data sample from a live 3G network and compare the analysis results. We also show how clustering of found anomalies can be used to find similarly behaving cells that can benefit from the same corrective measures.

The amount of data collected from telecommunications networks has increased significantly during the last decade. In comparison to the earlier networks, present-day 3G networks are able to provide more complex and detailed data, such as distributions of quality indicators. However, the operators lack proper tools to efficiently utilize these data in monitoring and analyzing the networks. Classification of the network elements (cells) into groups of similar behavior provides valuable information for a radio expert, who is responsible of hundreds or thousands of elements.

In this paper we propose fuzzy methods applied to 3G network channel quality distributions for analyzing the network performance. We introduce a traditional fuzzy inference system based on features extracted from the distributional data. We provide interpretation of the resulting classes to demonstrate their usability on network monitoring. Constructing and maintaining fuzzy rule sets are laborious tasks, therefore there is a demand for data driven methods that can provide similar information to the experts. We apply fuzzy C-means clustering to create performance classes. Finally, we introduce further analysis on how the performance of individual network elements varies between the classes in the course of time.

The underlying technologies driving the World Wide Web are largely based on the assumption of wired communications and powerful desktop hardware. This is not true when a user is accessing the Web pages and documents using a PDA and moving from wireless network in a mall to a neighboring office environment. Taking into consideration on the information needs of the user, we have developed a meta-search application that can be executed on mobile devices with limited hardware capabilities. By running in the auto-configuration mode, the system detects the dynamic contextual information, and presents the results of meta-searching in a suitable format to cope with the limitations of the mobile stations. This is achieved by an adaptive service composition that is defined according to the adaptation model defined in a system configuration file.

This paper describes statistical analysis of web data traffic to identify the probability distribution best fitting the connection arrivals and to evaluate whether data traffic traces follow heavy-tail distributions – an indication of fractal behaviour, which is in contrast to conventional data traffic.

Modelling of the fractal nature of web data traffic is used to specify classes of fractal computing methods which are capable of accurately describing the burstiness behaviour of the measured data, thereby establishing web data traffic patterns.

We present mathematical models that describe individual neural networks of the Central Nervous System. Three cases are examined, varying in each case the values of the refractory period and the synaptic delay of a neuron. In the case where both the refractory period and the synaptic delay are bigger than one, we split the population of neurons into sub-groups with their own distinct synaptic delay. It is shown that the proposed approach describes the neural activity of the network efficiently, especially in the case mentioned above. Various examples with different network parameters are presented to investigate the network’s behavior.

Assessment of the reliability of microarray experiments as well as their cross-laboratory/platform reproducibility rise as the major need. A critical challenge concerns the design of optimal Universal Reference RNA (URR) samples in order to maximize detectable spots in two-color/channel microarray experiments, decrease the variability of microarray data, and finally ease the comparison between heterogeneous microarray datasets. Towards this target we devised and present an

in-silico

(binary) optimization process the solutions of which present optimal URR sample designs. Setting a cut-off threshold value over which a gene is considered as detectably expressed enables the process. Experimental results are quite encouraging and the related discussion highlights the suitability and flexibility of the approach.

Prior work has shown that the sensitivity of a tumour to a specific drug can be predicted from a molecular signature of gene expressions. This is an important finding for improving drug efficacy and personalizing drug use. In this paper, we present an analysis strategy that, compared to prior work, maintains more information and leads to improved chemosensitivity prediction. Specifically we show (a) that prediction is improved when the GI50 value of a drug is estimated by all available measurements and fitting a sigmoid curve and (b) application of regression techniques often results in more accurate models compared to classification techniques. In addition, we show that (c) modern variable selection techniques, such as MMPC result in better predictive performance than simple univariate filtering. We demonstrate the strategy on 59 tumor cell lines after treatment with 118 fully characterized drugs obtained by the National Cancer Institute (NCI 60 screening) and biologically comment on the identified molecular signatures of the best predicted drugs.

In this manuscript, a Genetic Algorithm is applied on a filter in order to optimize the selection of clusters having a high probability to represent protein complexes. The filter was applied on the results (obtained by experiments made on five different yeast datasets) of three different algorithms known for their efficiency on protein complex detection through protein interaction graphs. The derived results were compared with three popular clustering algorithms, proving the efficiency of the proposed method according to metrics such as successful prediction rate and geometrical accuracy.

The analysis of protein-protein interactions (PPIs) is crucial to the understanding of cellular organizations, processes and functions. The reliability of the current experimental approaches interaction data is prone to error. Thus, a variety of computational methods have been developed to supplement the interactions that have been detected experimentally. The present paper’s main objective is to present a novel classification framework for predicting PPIs combining the advantages of two algorithmic methods’ categories (heuristic methods, adaptive filtering techniques) in order to produce high performance classifiers while maintaining their interpretability. Our goal is to find a simple mathematical equation that governs the best classifier enabling the extraction of biological knowledge. State-of-the-art adaptive filtering techniques were combined with the most contemporary heuristic methods which are based in the natural selection process. To the best of our knowledge, this is the first time that the proposed classification framework is applied and analyzed extensively for the problem of predicting PPIs. The proposed methodology was tested with a commonly used data set using all possible combinations of the selected adaptive filtering and heuristic techniques and comparisons were made. The best algorithmic combinations derived from these procedures were Genetic Algorithms with Extended Kalman Filters and Particle Swarm Optimization with Extended Kalman Filters. Using these algorithmic combinations high accuracy interpretable classifiers were produced.

Airborne pollen has been associated with allergic symptoms in sensitized individuals, whereas atmospheric pollution indisputably aggravates the impact on the overall quality of life. Therefore, it is of major importance to correlate, forecast and disseminate information concerning high concentration levels of allergic pollen types and air pollutants to the public, in order to safeguard the quality of life of the population. In this study, we investigate the relationship between the Defined Daily Dose (DDD) given to patients in a triggered allergy reaction and the different levels of air pollutants and pollen types. By profiling specific atmospheric conditions, specialists may define the need for medication to individuals suffering from pollen allergy, not only according to their personal medical record but also to the existing air quality observations. Paper results indicate some interesting interrelationships between the use of medication and atmospheric quality conditions and shows that the forecasting of daily medication is possible with the aid of proper algorithms.

This paper aims to shed light on the contribution of determinants to the health status of the population and to provide evidence on whether or not these determinants are producing similar results from two different statistical methods, across OECD countries. In this study, one output – Life Expectancy (LE) at birth of the total population – and three inputs are included. The inputs represent the three main dimensions of health outcome production: health resources (measured by health spending or the number of health practitioners), socioeconomic environment (pollution, education and income) and lifestyle (tobacco, alcohol and diet). A variable expressing country specificities is also used. Two independent statistical analyses, resulted that health resources and country specific effects are more closely related to LE.

The paper describes an environment quality analysis system based on a combination of some artificial intelligence techniques, artificial neural networks and rule-based expert systems. Two case studies of the system use are discussed: air pollution analysis and flood forecasting with their impact on the environment and on the population health. The system can be used by an environmental decision support system in order to manage various environmental critical situations (such as floods and environmental pollution), and to inform the population about the state of the environment quality.

Allergies due to airborne pollen affect approximately 15-20% of European citizens; therefore, the provision of health related services concerning pollen-induced symptoms can improve the overall quality of life. In this paper, we demonstrate the development of personalized quality of life services by adopting a data-driven approach. The data we use consist of allergic symptoms reported by citizens as well as detailed pollen concentrations of the most allergenic taxa. We apply computational intelligence methods in order to develop models that associate pollen concentration levels with allergic symptoms on a personal level. The results for the case of Austria, show that this approach can result to accurate and reliable models; we report a correlation coefficient up to r=0.70 (average of 102 citizens). We conclude that some of these models could serve as the basis for personalized health services.

This is an intelligent modeling of the evolution of drought and forest fires, due to climate change in Cyprus. Original annual wild fire data records (1979-2009) and data regarding meteorological parameters were used. A flexible modeling approach was proposed towards the determination of drought risk indices in all of the country. Cyprus was divided in eight polygons corresponding to eight meteorological stations. A Fuzzy Inference Rule Based System (FIRBS) was developed to produce the drought risk indices vectors for the forest regions under study. An analysis of the spatial distribution of the heat index vectors was performed. Forest fires distribution through the island was addressed. All of the results were stored by using an ArcGIS, (version 9.3) spatial data base that enables more comprehensive presentation of the most risky areas. There is a significant increase of drought in the island and this has a serious effect in the problems of forest fires and heat indices.