Engineering Applications of Neural Networks

Mechanical water stress assessment is needed in agriculture to mechanically cultivate high-sugar-content crops. Although previous methods estimate water stress accurately, no method has been practically applied yet due to the high cost of equipment. Thus, the previous methods have a trade-off relationship between cost and estimation accuracy. In this paper, we propose a method for estimating water stress on the basis of plant images and sensor data collected from inexpensive equipment. Specifically, a motion-specialized deep convolutional descriptor (MDCD), which is a novel image descriptor that extracts motion features among multiple sequential images without considering appearance in each image, expresses plant wilt strongly related to water stress. Implicit exclusion of appearance enables extraction of general features of plant wilt, which is insulated from the effect of differences in shapes and colors of places and individual plants. We evaluated the performance of the proposed method using enormous agricultural data collected from a greenhouse. Accordingly, the proposed method reduced the error of mean absolute error (MAE) by approximately 25% compared with a naive convolutional neural network (CNN) using original images. The results show that the MDCD enhances temporal information, while reducing spatial information, and expresses the features of plant wilt appropriately.

Heating, Ventilation, and Air Conditioning (HVAC) systems are generally built in a modular manner, comprising several identical subsystems in order to achieve their nominal capacity. These parallel subsystems and elements should have the same behavior and, therefore, differences between them can reveal failures and inefficiency in the system. The complexity in HVAC systems comes from the number of variables involved in these processes. For that reason, dimensionality reduction techniques can be a useful approach to reduce the complexity of the HVAC data and study their operation. However, for most of these techniques, it is not possible to project new data without retraining the projection and, as a result, it is not possible to easily compare several projections. In this paper, a method based on deep autoencoders is used to create a reference model with a HVAC system and new data is projected using this model to be able to compare them. The proposed approach is applied to real data from a chiller with 3 identical compressors at the Hospital of León.

This study focuses on the design of a Neural Network (NN) model for the prediction of interpolated values of polyvinylacetate (PVAc) nanofiber diameters produced by the electrospinning process and it supposes to be a preliminary work for future and industrial applications. The experimental data gathered from the literature form the basis for generating a more consistent sample through standard interpolation. The inputs of the NN are the polymer concentration, the applied voltage, the nozzle-collector distance and the flow rate parameters of the process, whereas the average diameter acts as the unique output of the network. The generated model is able to approximate the mapping between process parameters and fiber morphology, which is of practical importance to help prepare homogeneous nano-fibers. The reliability of the model was tested by 7-fold cross validation as well as leave-one-out method, showing good performance in terms of both average RMSE (0.109, corresponding to 138.51 nm) and correlation coefficient (0.905) between the desired and the predicted diameters when a White Gaussian Noise with 2% power (WGN_2%) is applied to the interpolations.

The prediction of electricity consumption has become an important part of managing the smart grid. Smart grid management involves energy production (from traditional and renewable sources), transportation and measurements (smart meters). Storing large amounts of electrical energy is not possible, therefore it is necessary to precisely predict energy consumption. Nowadays deep learning approaches are successfully used in different artificial intelligence areas. Deep neural network architecture called WaveNet was designed for text to speech task, improving speech quality over currently used approaches. In this paper, we present modification of the WaveNet architecture from speech (sound waves) generation to energy load prediction.

The global shift towards digitization has resulted in intensive research on Cryptographic techniques. Chaotic neural networks, augment the process of cryptography by providing increased security. In this paper, a description of an algorithm for the generation of an initial value for encryption using neural network involving memristor and chaotic polynomials is provided. The chaotic series that is obtained is combined with nonlinear 1 Dimensional and 2 Dimensional chaotic equations for the encryption process. A detailed analysis is performed to find the fastest converging neural network, complemented by the chaotic equations to produce least correlated ciphertext and plaintext. The use of Memristor in Neural Network as a generator for chaotic initial value as the encryption key and the involvement of nonlinear equations for encryption, makes the communication more confidential. The network can further be used for secure multi receiver systems.

Prediction, utilizing machine learning and data mining techniques is a significant tool, offering a first step and a helping hand for educators to early recognize those students who are likely to exhibit poor performance. In this work, we introduce a new decision support software for predicting the students’ performance at the final examinations. The proposed software is based on a novel 2-level classification technique which achieves better performance than any examined single learning algorithm. Furthermore, significant advantages of the presented tool are its simple and user-friendly interface and that it can be deployed in any platform under any operating system.

Students’ performance prediction in higher education has been identified as one of the most important research problems in machine learning. Educational data mining constitutes an important branch of machine learning trying to effectively analyze students’ academic behavior and predict their performance. Over recent years, several machine learning methods have been effectively used in the educational field with remarkable results, and especially supervised classification methods. The early identification of in case fail students is of utmost importance for the academic staff and the universities. In this paper, we investigate the effectiveness of active learning methodologies in predicting students’ performance in distance higher education. As far as we are aware of there exists no study dealing with the implementation of active learning methodologies in the educational field. Several experiments take place in our research comparing the accuracy measures of familiar active learners and demonstrating their efficiency by the exploitation of a small labeled dataset together with a large pool of unlabeled data.

The demand for digitisation of complex engineering drawings becomes increasingly important for the industry given the pressure to improve the efficiency and time effectiveness of operational processes. There have been numerous attempts to solve this problem, either by proposing a general form of document interpretation or by establishing an application dependant framework. Moreover, text/graphics segmentation has been presented as a particular form of addressing document digitisation problem, with the main aim of splitting text and graphics into different layers. Given the challenging characteristics of complex engineering drawings, this paper presents a novel sequential heuristics-based methodology which is aimed at localising and detecting the most representative symbols of the drawing. This implementation enables the subsequent application of a text/graphics segmentation method in a more effective form. The experimental framework is composed of two parts: first we show the performance of the symbol detection system and then we present an evaluation of three different state of the art text/graphic segmentation techniques to find text on the remaining image.

In the context of intrinsic plagiarism detection, we are trying to discover plagiarised passages in a text, based on the stylistic changes and inconsistencies within the document itself. The main idea consists in profiling the style of the original author and marking as outliers the passages that seem to differ significantly. Besides some novel stylistic and semantic features, the present work proposes a new approach to the problem, where machine learning plays a significant role. Notably, we also consider, for the first time, the reality of unbalanced training dataset in intrinsic plagiarism detection as a major parameter of the problem. Our detection system is tested on the data corpora of PAN Webis intrinsic plagiarism detection’s shared tasks of 2009 and 2011 and is compared to the results of the highest score participations.

One of the most common approaches for handling the multi-class classification problem is to divise the original data set into binary subclasses and to use a set of binary classifiers in order to solve the binarization problem. A new method for solving multi-class classification problems is proposed, by incorporating random resampling techniques in the one-versus-all strategy. Specifically, the division used by the proposed method is based on the one-versus-all binarization technique using random resampling for handling the class-imbalance problem arising due to the one-versus-all binarization. The method has been tested extensively on several multiclass classification problems using Support Vector Machines with four different kernels. Experimental results show that the proposed method exhibits a better performance compared to the simple one-versus-all.

Developments and upgrades in the field of industrial information technology, particularly those relating to information systems’ technologies for the collection and processing of real-time data, have introduced a large number of new threats. These threats are primarily related to the specific tasks these applications perform, such as their distinct design specifications, the specialized communication protocols they use and the heterogeneous devices they are required to interconnect. In particular, specialized attacks can undertake mechanical control, dynamic rearrangement of centrifugation or reprogramming of devices in order to accelerate or slow down their operations. This may result in total industrial equipment being destroyed or permanently damaged. Cyber-attacks against Industrial Control Systems which mainly use Supervisory Control and Data Acquisition (SCADA) combined with Distributed Control Systems are implemented with Programmable Logic Controllers. They are characterized as Advanced Persistent Threats. This paper presents an advanced Spiking One-Class Anomaly Detection Framework (SOCCADF) based on the evolving Spiking Neural Network algorithm. This algorithm implements an innovative application of the One-class classification methodology since it is trained exclusively with data that characterize the normal operation of ICS and it is able to detect divergent behaviors and abnormalities associated with APT attacks.

In this paper we present a new ensemble method, called Boosted Residual Networks, which builds an ensemble of Residual Networks by growing the member network at each round of boosting. The proposed approach combines recent developements in Residual Networks - a method for creating very deep networks by including a shortcut layer between different groups of layers - with the Deep Incremental Boosting, which has been proposed as a methodology to train fast ensembles of networks of increasing depth through the use of boosting. We demonstrate that the synergy of Residual Networks and Deep Incremental Boosting has better potential than simply boosting a Residual Network of fixed structure or using the equivalent Deep Incremental Boosting without the shortcut layers.

We introduce a convolutional network architecture aimed at performing token-level processing in natural language applications. We tune this architecture for a specific task - multiword expression detection - and we compare our results to state-of-the-art systems on the same datasets. The approach is multilingual and we rely on automatically extracted word embeddings from Wikipedia dumps. We also show that task-driven lexical features embeddings increase the speed and robustness of the system versus sparse encodings.

This study explored the application of a deep neural network to the task of recognising tea types from their aroma. The aroma was measured from tea leaves using an array of quartz crystal resonators coated with plasma organic polymer films. Frequency analysis based on continuous wavelet transform, with the Morlet function as the mother wavelet, was applied to the sensor signals to construct the input vectors of the deep neural network. Experiments were conducted using oolong, jasmine and pu’erh teas as the samples and dehumidified indoor air as the base gas. The deep neural network achieved a recognition accuracy of 96.3% for the three tea types and the base gas. The experimental results demonstrated the effectiveness of applying a deep neural network to this task.

Baby cry sound detection allows parents to be automatically alerted when their baby is crying. Current solutions in home environment ask for a client-server architecture where an end-node device streams the audio to a centralized server in charge of the detection. Even providing the best performances, these solutions raise power consumption and privacy issues. For these reasons, interest has recently grown in the community for methods which can run locally on battery-powered devices. This work presents a new set of features tailored to baby cry sound recognition, called hand crafted baby cry (HCBC) features. The proposed method is compared with a baseline using mel-frequency cepstrum coefficients (MFCCs) and a state-of-the-art convolutional neural network (CNN) system. HCBC features result to be on par with CNN, while requiring less computation effort and memory space at the cost of being application specific.

Detecting fire in images using image processing and computer vision techniques has gained a lot of attention from researchers during the past few years. Indeed, with sufficient accuracy, such systems may outperform traditional fire detection equipment. One of the most promising techniques used in this area is Convolutional Neural Networks (CNNs). However, the previous research on fire detection with CNNs has only been evaluated on balanced datasets, which may give misleading information on real-world performance, where fire is a rare event. Actually, as demonstrated in this paper, it turns out that a traditional CNN performs relatively poorly when evaluated on the more realistically balanced benchmark dataset provided in this paper. We therefore propose to use even deeper Convolutional Neural Networks for fire detection in images, and enhancing these with fine tuning based on a fully connected layer. We use two pretrained state-of-the-art Deep CNNs, VGG16 and Resnet50, to develop our fire detection system. The Deep CNNs are tested on our imbalanced dataset, which we have assembled to replicate real world scenarios. It includes images that are particularly difficult to classify and that are deliberately unbalanced by including significantly more non-fire images than fire images. The dataset has been made available online. Our results show that adding fully connected layers for fine tuning indeed does increase accuracy, however, this also increases training time. Overall, we found that our deeper CNNs give good performance on a more challenging dataset, with Resnet50 slightly outperforming VGG16. These results may thus lead to more successful fire detection systems in practice.

Among the most crucial components of an intelligent system capable of assisting drone-based cinematography is estimating the pose of the main actors. However, training deep CNNs towards this task is not straightforward, mainly due to the noisy nature of the data and instabilities that occur during the learning process, significantly slowing down the development of such systems. In this work we propose a temporal averaging technique that is capable of stabilizing as well as speeding up the convergence of stochastic optimization techniques for neural network training. We use two face pose estimation datasets to experimentally verify that the proposed method can improve both the convergence of training algorithms and the accuracy of pose estimation. This also reduces the risk of stopping the training process when a bad descent step was taken and the learning rate was not appropriately set, ensuring that the network will perform well at any point of the training process.

Accurate face recognition is vital in person identification tasks and may serve as an auxiliary tool to opportunistic video shooting using Unmanned Aerial Vehicles (UAVs). However, face recognition methods often require complex Machine Learning algorithms to be effective, making them inefficient for direct utilization in UAVs and other machines with low computational resources. In this paper, we propose a method of training Autoencoders (AEs) where the low-dimensional representation is learned in a way such that the various classes are more easily discriminated. Results on the ORL and Yale datasets indicate that the proposed AEs are capable of producing low-dimensional representations with enough discriminative ability such that the face recognition accuracy achieved by simple, lightweight classifiers surpasses even that achieved by more complex models.

In this paper, we analysed the performance of deep convolutional neural networks on noisy images of fish species. Thorough experiments using four variants of noisy and challenging dataset was carried out. Different deep convolutional models were evaluated. Firstly, we trained models on noisy dataset of fishing boat images. Our second approach trained the models on a new dataset generated by annotating fish instances only from the initial set of images. Lastly, we trained the models by synthesizing more data through the application of affine transforms and random noise. Results indicate that deep convolutional network performance deteriorate in the absence of well annotated training set. This opens direction for future research in automatic image annotation.

In this paper, a novel multiple-layer neuro-fuzzy network is proposed to model/predict shoreline realignment at a highly touristic island beach (Kamari beach, Santorini, Greece). A specialized experimental setup was deployed to generate a set of input-output data that comprise parameters describing the beach morphology and wave conditions and the cross-shore shoreline position at 30 cross-sections of the beach extracted from coastal video imagery, respectively. The proposed network consists of three distinct modules. The first module concerns the network representation of a fuzzy model equipped with a typical inference mechanism. The second module implements a novel competitive learning network to generate initial values for the rule base antecedent parameters. These parameters are, then, used to facilitate the third module that employs particle swarm optimization to perform a stochastic search for optimal parameter estimation. The network is compared favorably to two other neural networks: a radial basis function neural network and a feedforward neural network. Regarding the effectiveness of the proposed network to model shoreline re-alignment, the RMSE found (7.2–7.7 m, depending on the number of rules/nodes), reflects the high variability of the shoreline position of the Kamari beach during the period of observations: the RMSE is of a similar order to the standard deviation (up to 8 m) of the cross-shore shoreline position. The results are encouraging and the effectiveness of the proposed network could be further improved by changes (fine-tuning) of the input variables.

Fuzzy implications are widely used in applications where propositional logic is applicable. In cases where a variety of fuzzy implications can be used for a specific application, it is important that the optimal candidate be chosen in order valuable inference be drawn from a given set of data. This study introduces a method for detecting the most suitable fuzzy implication among others under consideration, which incorporates an algorithm forthe separation of two extreme cases. According to the truth values of the corresponding fuzzy propositions the optimal implication is one of these two extremes. An example involving five such relations is used to illustrate the procedure of the method. The results obtained verify that the resulting implication is the optimal operator for inference making from the data.

Seismocardiogram (SCG) recording is a novel method for the prolonged monitoring of the cardiac mechanical performance during spontaneous behavior. The continuous monitoring results in a collection of thousands of beats recorded during a variety of physical activities so that the automatic analysis and processing of such data is a challenging task due to the presence of artefactual beats and morphological changes over time that currently request the human expertise. On this premise, we propose the use of the Evolving Fuzzy Neural Network (EFuNN) paradigm for the automatic artifact detection in the SCG signal. The fuzzy logic processing method can be applied to model the human expertise knowledge using the learning capabilities of an artificial neural network. The evolving capability of the EFuNN paradigm has been applied to solve the issue of the physiological variability of the SGC waveform. Preliminary tests have been carried out to validate this approach and the obtained results demonstrate the effectiveness of the method and its scalability.

The prominent feature is the nonlinear characteristics as the squaring and rectification functions, which are observed in the retinal and visual cortex networks. Conventional model for motion processing in cortex, uses a symmetric quadratic functions with Gabor filters. This paper proposes a new motion processing model in the asymmetric networks. First, the asymmetric network is analyzed using Wiener kernels. It is shown that the asymmetric network with nonlinearities is effective and general for generating the directional movement compared with the conventional quadratic model. Second, independence maximization of data is an important issue in computational neural networks. To make clear the characteristics of the asymmetric network with Gabor functions, orthogonality is computed, which shows independent characteristics of the asymmetric network without maximizing optimization of independence in the quadratic model. The orthogonal analyses for the independence of the asymmetric networks are applied to the V1 and MT neural networks to generate independent subspaces by using selective Gabor functions.

This study investigates whether feedforward neural networks with two hidden layers generalise better than those with one. In contrast to the existing literature, a method is proposed which allows these networks to be compared empirically on a hidden-node-by-hidden-node basis. This is applied to ten public domain function approximation datasets. Networks with two hidden layers were found to be better generalisers in nine of the ten cases, although the actual degree of improvement is case dependent. The proposed method can be used to rapidly determine whether it is worth considering two hidden layers for a given problem.

In this paper we present a new strategy game, with machine learning computer players, which have been developed using temporal difference reinforcement learning coupled with neural networks; the latter are used for value approximation and for storing the players’ knowledge. We set out the game rules and then design and implement a comprehensive experimentation session to allow us to explore a large state space for investigating learning and playing behavior, without placing unreasonable demands on speed and accuracy. Our experiments demonstrate how computer players manage to adapt to their environment and improve their tactic over time, based on experience only, while still accommodating a variety of behaviors which are tuned via the conventional parameters of the reinforcement learning and neural network mechanisms.

Processing generalized sound events with the purpose of predicting the emotion they might evoke is a relatively young research field. Tools, datasets, and methodologies to address such a challenging task are still under development, far from any standardized format. This work aims to cover this gap by revealing and exploiting potential similarities existing during the perception of emotions evoked by sound events and music. o this end we propose (a) the usage of temporal modulation features and (b) a transfer learning module based on an Echo State Network assisting the prediction of valence and arousal measurements associated with generalized sound events. The effectiveness of the proposed transfer learning solution is demonstrated after a thoroughly designed experimental phase employing both sound and music data. The results demonstrate the importance of transfer learning in the specific field and encourage further research on approaches which manage the problem in a cooperative way.

Inversion of a neural network trained on some classification problem has been an important issue related to the explanation of the neural classification function. Inversion based on Interval Analysis (IA) [1] showed that a reliable estimation of the neural network domain of validity is feasible and a number of quantitative issues arise from this inversion. This paper deals with the investigation of these quantitative issues and more precisely with those concerning the evaluation of the neural network classification function in terms of generalization, comparison of different network models and classification accuracy. Preliminary experimental results indicate that the IA-based inversion can offer a solid basis towards reliable evaluation of the neural classification function.

In this work an adaptive learning rate algorithm for Convolutional Neural Networks is presented. Harvesting already computed first order information of the gradient vectors of three consecutive iterations during the training phase, an adaptive learning rate is calculated. The learning rate is increasing proportionally to the similarity of the direction of the gradients in an attempt to accelerate the convergence and locate a good solution. The proposed algorithm is suitable for the time-consuming training of the Convolutional Neural Networks, alleviating the exhaustive and critical for the performance of trained network heuristic search for a suitable learning rate. The experimental results indicate that the proposed algorithm produces networks having good classification accuracy, regardless the initial learning rate value. Moreover, the training procedure is similar or better to the gradient descent algorithm with fixed heuristically chosen learning rate.

Limitations of capabilities of shallow networks to represent sparsely real-valued functions on finite domains is investigated. Influence of sizes of function domains and of sizes dictionaries of computational units on sparsity of networks computing finite mappings is explored. It is shown that when dictionary is not sufficiently large with respect to the size of the finite domain, then almost any uniformly randomly chosen function on the domain either cannot be sparsely represented or its computation is unstable.

Detection of Fraudulent Financial Statements (FFS), or simpler fraud detection problem, refers to the falsification of financial statements with the aim either to demonstrate larger positive rates, such as assets and profit, or to conceal negative factors, such as expenses and losses. Since the expansion of contemporary markets and multinational trade are real phenomena, production of large volumes of data under which the operation of the current firms is facilitated constitutes a resulting consequence. Thus, analog upgrade of the antifraud mechanisms should be adopted, enabling the introduction of Machine Learning tools in the related field. However, because of the inability to collect trustworthy datasets that describe the corresponding ratios of a firm that has conducted fraud actions, strategies that exploit the existence of a few labeled instances for discovering useful patterns from a pool of unlabeled data could be proved really efficient. In this work, comparisons of algorithms that operate under Active Learning theory against their supervised variants are being conducted, using data extracted from Greek firms. To the best of our knowledge, this is the first study that uses Active Learning for predicting FFS. The obtained results prove the superior performance of the corresponding active learners.

In this paper we compare a selection of artificial neural networks when applied for short-term stock market price prediction. The networks are selected due to their expected relevance to the problem. Further, the work aims at covering recent advances in the field of artificial neural networks. The networks considered include: Feed forward neural networks, echo state networks, conditional restricted Boltzmann machines, time-delay neural networks and convolutional neural networks. These models are also compared to another type of machine learning algorithm, support vector machine. The models are trained on daily stock exchange data, to make short-term predictions for one day and two days ahead, respectively. Performance is evaluated by following the models directly in a simple financial strategy; trade every prediction they make once during each day.Possibly due to the noisy nature of stock data, the results are slightly inconsistent between different data sets. If performance is averaged across all the data sets, the feed forward network generates most profit during the three year test period: 23.13% and 30.43% for single-step and double-step prediction, respectively. Convolutional networks get close to the feed forward network in terms of profitability, but are found unreliable due to their unreasonable bias towards predicting positive price changes. The support vector machine delivered average profits of 17.28% for single-step and 11.30% for double-step, respectively. Low profits or large deviations were observed for the other models.

Several computer-based medical systems have been proposed for automatic detection of abnormalities in a variety of medical imaging domains. The majority of these systems are based on binary supervised classification algorithms capable of discriminating abnormal from normal image patterns. However, this approach usually does not take into account that the normal content of images is diverse, including various kinds of tissues and artifacts. In the context of gastrointestinal video-endoscopy, which is addressed in this study, the semantics of the normal content include mucosal tissues, the hole of the lumen, bubbles, and debris. In this paper we investigate such a semantic interpretation of the endoscopy video content as an approach to improve lesion detection in a weakly supervised framework. This framework is based on a novel salient point detection algorithm, the bag-of-words image representation technique and multi-label classification. Advantages of the proposed method include: (a) It does not require detailed, pixel-level annotation of training images, instead image-level annotations are sufficient; (b) It enables a richer description of image content, which is beneficial for the discrimination of lesions. The annotation of the multi-labeled training images was performed using a novel annotation tool called RATStream. The results of the experiments performed in a wireless capsule endoscopy dataset with inflammatory lesions promises an improved performance for future generation diagnostic systems.

A novel system based on deep neural networks is presented, that performs analysis of medical imaging data. The aim is to study structural and functional alterations of the human brain in patients with Parkinson’s Disease and to correlate them with epidemiological and clinical data. A new medical database, which is presently under development, is used for training the system and testing its performance. Preliminary experimental results are provided which illustrate the capability of the proposed system to analyze and provide an accurate estimation of the status of the disease.

In this work, we report the use of convolutional neural networks for the detection of malignant melanomas against nevus skin lesions in a dataset of dermoscopic images of the same magnification. The technique of transfer learning is utilized to compensate for the limited size of the available image dataset. Results show that including transfer learning in training CNN architectures improves significantly the achieved classification results.

In this study, we introduce a new population based optimization algorithm named Sonar Inspired Optimization (SIO). This algorithm is based on the underwater acoustics that war ships use for reckoning targets and obstacles. The advantage of the proposed method is the ability for performing wider range of search during the iterations of the algorithm. The proposed algorithm is tested in known benchmarks and results are encouraging.

Missing hydrometric data is a critical issue for water resources management projects and problems related to flow damage and risk assessment. Though numerous ways can be found in the literature to impute them (i.e. Box-Jenkins models, Linear regression models, case deletion, listwise and pairwise deletion, etc.), not all will render effective on a given dataset. in tropical river basin, it’s still needed to develop proven and simplified methods to deal with hydrometric data missingness and scarcity. This paper presents the analyses including an assessment of the condition of the existing hydrometric data and works related to the way in which the record was treated for flow forecasting purposes and the construction of the artificial neural network (ANN) models used for predicting the flows. The study was led based on 15-min rainfall, water surface elevation and discharge data, derived from the continuous real-time monitoring station located in the del Medio River Basin from the years 2012 to 2016. As a result, the proposed modeling approach followed two modeling methods, one employing the missing data record and the other was used a multiple imputation (MI) technique to impute the missing data and forecast flow for 1, 2 and 4 h ahead under each approach. The statistical metrics results for the two-modeling approaches, suggest the non-imputed data scenario to rule out the imputed data. This means it is recommended to further optimize the MI technique if to be used effectively to fill in the missing required days of measurements for estimating H3 gaps and afterwards to forecast the flow employing multilayer perceptron (MLP), artificial neural networks (ANNs) with 10-fold cross-validation.

Clustering algorithms like k-means, BIRCH, CLARANS and DBSCAN are designed to be scalable and they are developed to discover clusters in the full dimensional space of a database. Nevertheless their characteristics depend upon the size of the database. A DB/data warehouse may store terabytes of data. Complex data analysis (mining) may take a very long time to run on the complex dataset. One has to obtain a reduced representation of the dataset that is much smaller in volume - but yet produces the same or almost the same analytical results - in order to accelerate information processing. Reduced representations yield simplified models that are easier to interpret, avoid the curse of dimensionality and enhance generalization by reducing overfitting. Data reduction methods include data cube aggregation, attribute subset selection, fitting data into models, dimensionality reduction, hierarchies as well as other approaches. Feature selection is considered as a specific case of a more general paradigm which is called Structure Learning in cases of an outcome associated to a set of attributes. Feature selection aims at selecting a minimum set of features such that the probability distribution of different classes given the values of those features is as close as possible to the original distribution given the values of all features. A combined approach based upon representing complex datasets in DB as a minimal set of connected attribute sets of reduced dimensions is herein proposed. Value-Difference (VD) Metrics based upon binary, categorical and continuous values are used for subspace clustering. Each cluster can be represented by a different set of object features/attributes maximizing the information which is rendered by the cluster representation. Numerical data regarding a test-bed system for anomaly detection are provided in order to illustrate the aforementioned approach.

In this paper we address the problem of declarative modeling of freight transportation brokering using agents and constraints. Our model can be used for the optimization of vehicle assignments to customer orders that request the transportation of freight from source to destination points. The advantage is that a single vehicle can serve multiple customer orders on its multi-hop route that is part of a solution schedule. Our model is mapped to the ECLiPSe constraint logic programming system such that optimal schedules can be automatically computed using the available constraint solvers. We propose a method and protocol for integrating this constraint-based scheduler into a multi-agent system.

EEG techniques have been widely used for mental fatigue monitoring, which is an important factor for driving safety. In this work, we performed an experiment involving one hour driving simulation. Based on EEG recordings, we created brain functional networks in alpha power band with three different methods, partial directed coherence (PDC), direct transfer function (DTF) and phase lag index (PLI). Then, we performed feature selection and classification between alertness and fatigue states, using the functional connectivity as features. High accuracy (84.7%) was achieved, with 22 discriminative connections from PDC network. The selected features revealed alterations of the functional network due to mental fatigue and specifically reduction of information flow among areas. Finally, a feature ranking is provided, which can lead to electrode minimization for real-time fatigue monitoring applications.

The popularity of recommendation systems has made them a substantial component of many applications and projects. This work proposes a framework for package recommendations that try to meet users’ preferences as much as possible through the satisfaction of several criteria. This is achieved by modeling the relation between the items and the categories these items belong to aiming to recommend to each user the top-k packages which cover their preferred categories and the restriction of a maximum package cost. Our contribution includes an optimal and a greedy solution. The novelty of the optimal solution is that it combines the collaborative filtering predictions with a graph based model to produce recommendations. The problem is expressed through a minimum cost flow network and is solved by integer linear programming. The greedy solution performs with a low computational complexity and provides recommendations which are close to the optimal solution. We have evaluated and compared our framework with a baseline method by using two popular recommendation datasets and we have obtained promising results on a set of widely accepted evaluation metrics.

Finite Mixture of Regression (FMR) models account for the target variable and cluster the data points based on the relationship between inputs and target variables. However, extant FMR models mostly rely on linear regression models. We propose a competitive learning algorithm to perform FMR modeling, which allows nonlinear models such a Multi-Layer Perceptrons (MLPs) to carry out the regression. We demonstrate the proposed method using a real-world case study that aims to derive tailored recommendations for franchises to improve their profitability and sales effectiveness.

Recommendation systems offer valuable assistance with selecting products and services. This work checks the hypothesis that taking personality into account can improve recommendation quality. Our main goal is to examine the role of personality in Movie Recommender systems. We introduce the concept of combining collaborative techniques with a personality test to provide more personalized movie recommendations. Previous research attempted to incorporate personality in Recommender systems, but no actual implementation appears to have been achieved. We propose a method and developed the 50/50 recommender system, which combines the Big Five personality test with an existing movie recommender, and used it on a renowned movie dataset. Evaluation results showed that users preferred the 50/50 system 3.6% more than the state of the art method. Our findings show that personalization provides better recommendations, even though some extra user input is required upfront.

Modern information society depends on reliable functionality of information systems infrastructure, while at the same time the number of cyber-attacks has been increasing over the years and damages have been caused. Furthermore, graphs can be used to show paths than can be exploited by attackers to intrude into systems and gain unauthorized access through vulnerability exploitation. This paper presents a method that builds attack graphs using data supplied from the maritime supply chain infrastructure. The method delivers all possible paths that can be exploited to gain access. Then, a recommendation system is utilized to make predictions about future attack steps within the network. We show that recommender systems can be used in cyber defense by predicting attacks. The goal of this paper is to identify attack paths and show how a recommendation method can be used to classify future cyber-attacks. The proposed method has been experimentally evaluated and it is shown that it is both practical and effective.

Pattern recognition is a branch of machine learning that focuses on the recognition of patterns and regularities in data for object recognition, classification and computer vision segmentation. Features are extracted from input data and used for object classification purposes. Artificial Neural Networks (ANNs) and Deep Neural Networks (DNNs) are popular tools for pattern recognition applications. The performance of the networks is usually defined in terms of the classification accuracy. However, there are no real design guidelines for training and testing protocols. This research set out to evaluate the effect on accuracy of the design parameters, including: size of the database, number of classes, quality of images, type of network, nature of training and testing strategy. A coin recognition task was used for the evaluation. A set of guidelines for part recognition tasks is presented based on experience with this task.

The approaches in this work combine the swarm behavior principles of Craig W. Reynolds with space filling curves movements. We intend to evaluate how the entire swarm moves by including a deterministic leader behavior for some agents. Therefore, we examine different combinations of Hilbert Curves with the classical swarm algorithms. We introduce a practical problem, the collection of manganese nodules on the sea ground by using autonomous agents. Some relevant experiments, combining different parameters for the leaders were run and the results are evaluated and described. Finally, we propose further developments and ideas to continue this research.

Reactive spatial robot navigation in goal-directed tasks such as phonotaxis requires generating consistent and stable trajectories towards an acoustic target while avoiding obstacles. High-level goal-directed steering behaviour can steer a robot towards the target by mapping sound direction information to appropriate wheel velocities. However, low-level obstacle avoidance behaviour based on distance sensors may significantly alter wheel velocities and temporarily direct the robot away from the sound source, creating conflict between the two behaviours. How can such a conflict in reactive controllers be resolved in a manner that generates consistent and stable robot trajectories? We propose a neural circuit that minimises this conflict by learning sensorimotor mappings as neuronal transfer functions between the perceived sound direction and wheel velocities of a simulated non-holonomic mobile robot. These mappings constitute the high-level goal-directed steering behaviour. Sound direction information is obtained from a model of the lizard peripheral auditory system. The parameters of the transfer functions are learned via an online unsupervised correlation learning algorithm through interaction with obstacles in the form of low-level obstacle avoidance behaviour in the environment. The simulated robot is able to navigate towards a virtual sound source placed 3 m away that continuously emits a tone of frequency 2.2 kHz, while avoiding randomly placed obstacles in the environment. We demonstrate through two independent trials in simulation that in both cases the neural circuit learns consistent and stable trajectories as compared to navigation without learning.

Complex networks can be considered as a new field of scientific research inspired by the empirical study of real-world networks such as computer, social as well as biological ones. More to this point, the study of complex networks has expanded in many disciplines including mathematics, physics, biology, telecommunications, computer science, sociology, epidemiology and others. An important type of complex networks are called biological dealing with the mathematical analysis of connections - interfaces that are ecological, evolutionary and physiological studies, such as neural networks or network epidemic models. The analysis of biological networks in connection with human diseases has led to expand science and examine medical supplies networks for their deeper understanding. In this paper, an implementation of epidemic/networks models is introduced concerning the HIV spreading in a sample of people who are needle drug users.

Algorithmic music composition has long been in the spotlight of music information research and Long Short-Term Memory (LSTM) neural networks have been extensively used for this task. However, despite LSTM networks having proven useful in learning sequences, no methodology has been proposed for learning sequences conditional to constraints, such as given metrical structure or a given bass line. In this paper we examine the task of conditional rhythm generation of drum sequences with Neural Networks. The proposed network architecture is a combination of LSTM and feed forward (conditional) layers capable of learning long drum sequences, under constraints imposed by metrical rhythm information and a given bass sequence. The results indicate that the role of the conditional layer in the proposed architecture is crucial for creating diverse drum sequences under conditions concerning given metrical information and bass lines.

A closed string contains a proper factor occurring as both a prefix and a suffix but not elsewhere in the string. Closed strings were introduced by Fici (WORDS 2011) as objects of combinatorial interest. In this paper, we extend this definition to k-closed strings, for which a level of approximation is permitted up to a number of Hamming distance errors, set by the parameter k. We then address the problem of identifying whether or not a given string of length n over an integer alphabet is k-closed and additionally specifying the border resulting in the string being k-closed. Specifically, we present an $$\mathcal {O}(kn)$$-time and $$\mathcal {O}(n)$$-space algorithm to achieve this along with the pseudocode of an implementation.

Nowadays, digital data are the most valuable asset of almost every organization. Database management systems are considered as storing systems for efficient retrieval and processing of digital data. However, effective operation, in terms of data access speed and relational database is limited, as its size increases significantly [6]. Bloom filter is a special data structure with finite storage requirements and rapid control of an object membership to a dataset. It is worth mentioning that the Bloom filter structure has been proposed with a view to constructively increase data access in relational databases. Since the characteristics of a Bloom filter are consistent with the requirements of a fast data access structure, we examine the possibility of using it in order to increase the SQL query execution speed in a database. In the context of this research, a database in a RDBMS SQL Server that includes big data tables is implemented and in following the performance enhancement, using Bloom filters, in terms of execution time on different categories of SQL queries, is examined. We experimentally proved the time effectiveness of Bloom filter structure in relational databases when dealing with large scale data.

Remote care and telemonitoring have become essential component of current geriatric medicine. Intelligent use of wireless sensors is a major issue in relevant computational studies to realize these concepts in practice. While there has been a growing interest in recognizing daily activities of patients through wearable sensors, the efforts towards utilizing the streaming data from these sensors for clinical practices are limited. Here, we present a practical application of clinical data mining from wearable sensors with a particular objective of diagnosing Parkinson’s Disease from gait analysis through a sets of ground reaction force (GRF) sensors worn under the foots. We introduce a supervised learning method based on Random Forests that analyze the multi-sensor data to classify the person wearing these sensors. We offer to extract a set of time-domain and frequency-domain features that would be effective in distinguishing normal and diseased people from their gait signals. The experimental results on a benchmark dataset have shown that proposed method can significantly outperform the previous methods reported in the literature.

In this work, we consider a special type of uncertain sequence called weighted string. In a weighted string every position contains a subset of the alphabet and every letter of the alphabet is associated with a probability of occurrence such that the sum of probabilities at each position equals 1. Usually a cumulative weight threshold is specified, and one considers only strings that match the weighted string with probability at least . We provide an $$\mathcal {O}(nz)$$-time and $$\mathcal {O}(nz)$$-space off-line algorithm, where n is the length of the weighted string and is the given threshold, to compute a smallest maximal palindromic factorization of a weighted string. This factorization has applications in hairpin structure prediction in a set of closely-related DNA or RNA sequences. Along the way, we provide an $$\mathcal {O}(nz)$$-time and $$\mathcal {O}(nz)$$-space off-line algorithm to compute maximal palindromes in weighted strings.

Multimodal social networks are omnipresent in Web 2.0 with virtually every human communication action taking place there. Nonetheless, language remains by far the main premise such communicative acts unfold upon. Thus, it is statutory to discover language communities especially in social data stemming from historically multilingual countries such as Luxemburg. An adjacency tensor is especially suitable for representing such spatiosocial data. However, because of its potentially large size, heuristics should be developed for locating community structure efficiently. Linguistic structure discovery has a plethora of applications including digital marketing and online political campaigns, especially in case of prolonged and intense cross-linguistic contact. This conference paper presents TENSOR-G, a flexible genetic algorithm for approximate tensor clustering along with two alternative fitness functions derived from language variation or diffusion properties. The Kruskal tensor decomposition serves as a benchmark and the results obtained from a set of trilingual Luxemburgian tweets are analyzed with linguistic criteria.

The evolution of mobile technologies gives the opportunity for innovative approaches to the content, the process and the evaluation of the educational activity. In this paper we study the particular situation where learner (student) is in displacement, in various conditions (walking in the rain, driving a car, riding a bicycle etc.), and receives the educational content through his or her smart mobile devices. Such a dynamic educational process, that we call mobile learning, presents many new aspects to study. We focus on the study of the interaction of the student with the educational process, running under various influencing factors and restrictions. In particular, the research carried out on Greek students who are learning Russian language.

Based on the context of the original SESAME project research effort, in the present work we examined the implications of multi-tenancy upon the Radio Resources Management (RRM) and Self-x functions that support mobility control, as the latter is a fundamental functionality to ensure a seamless experience to the user equipments of the different operators when moving across the cells of a shared RAN (Radio Access Network) and when entering and leaving the shared infrastructure.

This paper focuses on multi-tenant 5G networks with virtualization and mobile edge computing capabilities, in the scope of cloud-enabled small cell deployments. In this context, the work here presented deals with the service management and orchestration challenges that arise when handling service mapping on the multi-tenant distributed cloud-enabled radio access network architecture. For that aim, once analysed cloud edge services management and 5G network instantiation in the OpenStack platform, we modify the provided virtual infrastructure manager so as to incorporate virtual network function placement features of the SESAME environment. As main contributions, we adapt the OpenStack application instances to 5G Network Service instantiation, and we include an energy-aware and latency-constrained placement solution.

The application of Artificial Intelligence (AI)-based knowledge discovery mechanisms for supporting the automation of wireless network operations is envisaged to fertilize in future Fifth Generation (5G) systems due to the stringent requirements of these systems and to the advent of big data analytics. This paper intends to elaborate on the demonstration of knowledge discovery capabilities in the context of the architecture proposed by the Small cEllS coordinAtion for Multi-tenancy and Edge services (SESAME) project that deals with multi-operator cloud-enabled small cells. Specifically, the paper presents the considered demonstration framework and particularizes it for supporting an energy saving functionality through the classification of cells depending on whether they can be switched off during certain times. The framework is illustrated with some results obtained from real small cell deployments.

Although 5G promises advanced features such as low latency, high data rates and reliability as well as high socio-economic value, the business opportunities of the proposed solutions have not yet been examined. In this paper, the SESAME approach along with spectrum sharing options and indicative use cases are initially described. The incentives for SESAME adoption along with the value proposition and creation are analyzed. Finally, a reference model describing the role and interactions of the involved players as well as the revenue streams is provided.

The core challenges in the actual SESAME EU-funded project is to develop an ecosystem to sustain network infrastructure openness, built on the pillars of network functions virtualization (NFV), mobile-edge computing (MEC) capabilities and cognitive network management that will provide multi-tenancy and flexible cloud-network interaction with highly-predictable and flexible end-to-end performance characteristics. Based on this aspect, we discuss the potential benefits of including NFV and MEC in a modern mobile communications infrastructure, through Small Cells coordination and virtualization, also focused upon realistic 5G-oriented considerations. Within the proposed SESAME architecture, we also assess the various advantages coming from a more enhanced network operation and management of resources, as it appears with the incorporation of cognitive capabilities embracing knowledge and intelligence.

Based on the actual framework of the SESAME 5G-PPP EU-funded project, we identify the importance of the related wireless backhauling within the broader 5G innovative framework, with the pure aim of using small cells together with suitable network virtualization techniques for serving multiple tenants in a modern architectural approach. The virtualization of the network nodes and the wireless links allow for the development of a suitable SDN controller intending to perform network slicing, where the wireless backhaul resources are shared and assigned on a per-tenant basis. In order to apply SON features as they are also applied at the access radio level, the SDN controller is responsible for collecting and evaluating status information of the network (link qualities, status of wireless interfaces, ongoing traffic), thus resulting to self-planning, self-optimization and self-healing attributes.

Virtualisation is playing a fundamental role in the evolution of telecommunication services and infrastructures, bringing to rethink some of the traditional design paradigms of the mobile network and enabling those functionalities necessary for supporting new complex ecosystems where multiple actors can participate in a dynamic and secure environment. In Small Cell enabled Mobile Edge Computing deployments, the impact of virtualization technologies is significant in two main aspects: the design and deployment of the telecommunication infrastructure, and the delivery of edge services. Besides, the adoption of virtualization technologies has implications also in the implementation of Self Organizing Network (SON) services and in the enforcement of Service Level Agreement (SLA) policies, both critical in the automation of the delivery of multi-tenant oriented services in such complex infrastructure. From the work performed by the H2020 SESAME project, the beneficial use of virtualization techniques emerges in adding network intelligence and services in the network edge. SESAME relays on virtualization for providing Small Cell as a Service (SCaaS) and per operator Edge Computing services, consolidating the emerging multi-tenancy driven design paradigms in communication infrastructures.