Skip to main content
main-content

Über dieses Buch

This two-volume set LNCS 10305 and LNCS 10306 constitutes the refereed proceedings of the 14th International Work-Conference on Artificial Neural Networks, IWANN 2017, held in Cadiz, Spain, in June 2017.

The 126 revised full papers presented in this double volume were carefully reviewed and selected from 199 submissions. The papers are organized in topical sections on Bio-inspired Computing; E-Health and Computational Biology; Human Computer Interaction; Image and Signal Processing; Mathematics for Neural Networks; Self-organizing Networks; Spiking Neurons; Artificial Neural Networks in Industry ANNI'17; Computational Intelligence Tools and Techniques for Biomedical Applications; Assistive Rehabilitation Technology; Computational Intelligence Methods for Time Series; Machine Learning Applied to Vision and Robotics; Human Activity Recognition for Health and Well-Being Applications; Software Testing and Intelligent Systems; Real World Applications of BCI Systems; Machine Learning in Imbalanced Domains; Surveillance and Rescue Systems and Algorithms for Unmanned Aerial Vehicles; End-User Development for Social Robotics; Artificial Intelligence and Games; and Supervised, Non-Supervised, Reinforcement and Statistical Algorithms.

Inhaltsverzeichnis

Frontmatter

Erratum to: A Novel Technique to Estimate Biological Parameters in an Epidemiology Problem

Antone dos Santos Benedito, Fernando Luiz Pio dos Santos

Bio-inspired Computing

Frontmatter

A Parallel Swarm Library Based on Functional Programming

In this paper we present a library of parallel skeletons to deal with swarm intelligence metaheuristics. The library is implemented using the parallel functional language Eden, an extension of the sequential functional language Haskell. Due to the higher-order nature of functional languages, we simplify the task of writing generic code, and also the task of comparing different strategies. The paper illustrates how to develop new skeletons and presents empirical results.

Fernando Rubio, Alberto de la Encina, Pablo Rabanal, Ismael Rodríguez

A Parallel Island Approach to Multiobjective Feature Selection for Brain-Computer Interfaces

This paper shows that parallel processing is useful for feature selection in brain-computer interfacing (BCI) tasks. The classification problems arising in such application usually involve a relatively small number of high-dimensional patterns and, as curse of dimensionality issues have to be taken into account, feature selection is an important requirement to build suitable classifiers. As the number of features defining the search space is high, the distribution of the searching space among different processors would contribute to find better solutions, requiring similar or even smaller amount of execution time than sequential counterpart procedures. We have implemented a parallel evolutionary multiobjective optimization procedure for feature selection, based on the island model, in which the individuals are distributed among different subpopulations that independently evolve and interchange individuals after a given number of generations. The experimental results show improvements in both computing time and quality of EEG classification with features extracted by multiresolution analysis (MRA), an approach widely used in the BCI field with useful properties for both temporal and spectral signal analysis.

Julio Ortega, Dragi Kimovski, John Q. Gan, Andrés Ortiz, Miguel Damas

Deep Belief Networks and Multiobjective Feature Selection for BCI with Multiresolution Analysis

High-dimensional pattern classification problems with a small number of training patterns are difficult. This paper deals with classification of motor imagery tasks for brain-computer interfacing (BCI), which is a hard problem involving a relatively small number of high-dimensional training patterns where curse of dimensionality issue has to be taken into account and feature selection is an important requirement to build a suitable classifier. Evolutionary metaheuristics for feature selection are usually more time-consuming than other alternatives, but their high performances in terms of classification accuracy make them desirable approaches. In this paper, feature selection through a wrapper procedure based on multi-objective optimization is compared with the use of deep belief networks (DBN) that constitute powerful classifiers implementing feature selection implicitly. Two different classifiers, LDA (linear discriminant analysis) and DBN, have been used to classify EEG signals with features extracted by multiresolution analysis (MRA) and selected by a multiobjective evolutionary method that also uses LDA to implement the fitness function of the solutions. The experimental results show that DBNs usually provide better or similar classification performances without requiring an explicit feature selection phase. Nevertheless, the DBN’s classification performance significantly decreases in problems with a very large number of features. Moreover, to achieve high classification rates, it is necessary to determine a suitable structure for the DBN. Therefore, in this paper we also propose a multiobjective approach to tackle this problem.

Julio Ortega, Andrés Ortiz, Pedro Martín-Smith, John Q. Gan, Jesús González-Peñalver

IMOGA/SOM: An Intelligent Multi-objective Genetic Algorithm Using Self Organizing Map

Multi-objective Genetic Algorithms (MOGAs) are probabilistic search techniques and provide solutions of multi-objective optimization problems. When MOGA reaches near optimal regions, it may face problem in convergence due to its probabilistic nature. MOGA does not pay attention on the neighbourhood of the current population which makes the convergence slow. This scenario may also lead to premature convergence. To overcome this problem, we propose an Intelligent Multi-objective Genetic Algorithm using Self Organizing Map (IMOGA/SOM). The proposed algorithm uses the neighbourhood property of SOM. SOM is trained by the solutions generated by MOGA. SOM performs competition and cooperation among its neurons for better convergence. We have compared the results of the proposed algorithm with two existing algorithms NSGA-II and SOM-Based Multi Objective Genetic Algorithm (SBMOGA). Empirical results demonstrate the superiority of the proposed algorithm IMOGA/SOM.

Subhradip Aon, Ashis Sau, Prasenjit Dey, Tandra Pal

Solving Scheduling Problems with Genetic Algorithms Using a Priority Encoding Scheme

Scheduling problems are very hard computational tasks with several applications in multitude of domains. In this work we solve a practical problem motivated by a real industry situation, in which we apply a genetic algorithm for finding an acceptable solution in a very short time interval. The main novelty introduced in this work is the use of a priority based chromosome codification that determines the precedence of a task with respect to other ones, permitting to introduce in a very simple way all problem constraints, including setup costs and workforce availability. Results show the suitability of the approach, obtaining real time solutions for tasks with up to 50 products.

José L. Subirats, Héctor Mesa, Francisco Ortega-Zamorano, Gustavo E. Juárez, José M. Jerez, Ignacio Turias, Leonardo Franco

Tuning of Clustering Search Based Metaheuristic by Cross-Validated Racing Approach

The success of a metaheuristic is directly tied to the good configuration of its free parameters, this process is called Tuning. However, this task is, usually, a tedious and laborious work without scientific robustness for almost all researches. The absence of a formal definition of the tuning and diversity of metaheuristic research contributes to the difficulty in comparing and validating the results, making the progress slower. In this paper, a tuning method named Cross-Validated Racing (CVR) is proposed along with the so named Biased Random-Key Evolutionary Clustering Search and applied to solve instances of the Permutation Flow Shop Problem (PFSP). The proposed approach has reached $$99.1\%$$ of accuracy in predicting the optimal solution with the parameters found by Irace tuning method. Configurations generated by Irace, even different, have obtained results with the same statistical relevance.

Thiago Henrique Lemos Fonseca, Alexandre Cesar Muniz de Oliveira

A Transformation Approach Towards Big Data Multilabel Decision Trees

A large amount of the data processed nowadays is multilabel in nature. This means that every pattern usually belongs to several categories at once. Multilabel data are abundant, and most multilabel datasets are quite large. This causes that many multilabel classification methods struggle with their processing. Tackling this task by means of big data methods seems a logical choice. However, this approach has been scarcely explored by now. The present work introduces several big data multilabel classifiers, all of them based on decision trees. After detailing how they have been designed, their predictive performance, as well as the execution time, are analyzed.

Antonio Jesús Rivera Rivas, Francisco Charte Ojeda, Francisco Javier Pulgar, Maria Jose del Jesus

Evolutionary Support Vector Regression via Genetic Algorithms: A Dual Approach

Evolutionary machine learning is an emerging research area that covers any combination of evolutionary strategies and machine learning. In support vector machines, metaheuristics have been widely employed to tune parameters, select features or obtain a reduced sub-set of support vectors. However, there are only a few works that aim at embedding evolutionary strategies into the support vector regressors training process, i.e., to apply evolutionary methods to solve the quadratic optimization problem. In this paper, we intend to solve the quadratic optimization problem for support vector regression in its dual formulation by employing genetic algorithms. Our proposal was validated in real-world datasets against state-of-the-art methods, such as sequential minimal optimization, iterative single data algorithm, and a classical mathematical method. The results revealed that our proposal is a competitive alternative, which often reduced the generalization error and achieved sparse solutions.

Shara S. A. Alves, Madson L. D. Dias, Ajalmar R. da Rocha Neto, Ananda L. Freire

E-Health and Computational Biology

Frontmatter

Analysis of Electroreception with Temporal Code-Driven Stimulation

Temporal code-driven stimulation is a new closed-loop stimulation method for information processing research in biological systems. The biological signal is processed and an event-based binary digitization is performed in real time. Patterns of temporal activity in the system are matched with binary codes and stimulation is triggered after the detection of a predetermined code. This paper presents the characteristics of this closed-loop methodology together with novel analytical possibilities derived from using an information-theoretic approach. The implementation of this method for its application to the study of coding schemes in fish electroreception is presented. Finally, our preliminary results showed that code-driven stimulation decreases the discharge frequency of the electric fish and increases the probability of sparser codes. The relation between those two measures can be used to assess the analysis of factors involved in the information processing in the system.

Ángel Lareo, Caroline Garcia Forlim, Reynaldo D. Pinto, Pablo Varona, Francisco B. Rodríguez

A Novel Technique to Estimate Biological Parameters in an Epidemiology Problem

In this paper, we describe a study of a parameter estimation technique to estimate a set of unknown biological parameters of a non-linear dynamic model of dengue. We also explore a Levenberg-Marquardt (LM) algorithm to minimize the cost function. A classical mathematical model describes the dynamics of mosquitoes in water and winged phases, where the data are available. The main interest is to fit the model to the data taking into account the parameters estimated. Numerical simulations were performed and results showed the robustness of LM in estimating the important parameters in the dengue disease problem.

Antone dos Santos Benedito, Fernando Luiz Pio dos Santos

Breast Cancer Microarray and RNASeq Data Integration Applied to Classification

Although Next-Generation Sequencing (NGS) has more impact nowadays than microarray sequencing, there is a huge volume of microarray data that has not still been processed. The last represents the most important source of biological information nowadays due largely to its use over many years, and a very important potential source of genetic knowledge deserving appropriate analysis. Thanks to the two techniques, there is now a huge amount of data that allows us to obtain robust results from its integration. This paper deals with the integration of RNASeq data with microarrays data in order to find breast cancer biomarkers as expressed genes. These integrated data has been used to create a classifier for an early diagnosis of breast cancer.

Daniel Castillo, Juan Manuel Galvez, Luis Javier Herrera, Ignacio Rojas

Deep Learning Using EEG Data in Time and Frequency Domains for Sleep Stage Classification

Polysomnography analysis for sleeping disorders is a discipline that is showing interest in the development of reliable classifiers to determine the sleep stage. The most common methods shown in the literature bet for classical learning techniques and statistics that are applied to a reduced number of features in order to tackle the computational load. Nowadays, the application of deep learning to the sleep stage classification problem seems very interesting and novel, therefore, this paper presents a first approximation using a single channel and information from the current epoch to perform the classification. The complete Physionet database has been used in the experiments. Deep learning has been applied to the time and frequency domains from the EEG signal obtaining a good performance and promising further work.

Martí Manzano, Alberto Guillén, Ignacio Rojas, Luis Javier Herrera

Human Computer Interaction

Frontmatter

Application of an Eye Tracker Over Facility Layout Problem to Minimize User Fatigue

With interactive evolutionary computation it is possible to introduce the subjective preferences of the decision maker within the general algorithm evolution criteria. The problem that generates this is user fatigue, since it has to evaluate a considerable number of plants designs in each generation. To avoid user fatigue it is proposed to substitute the direct evaluation through the mouse by means of a numerical scale by an eye tracking system in which the system “captures” the evaluation that the user assigns to the plants through the gaze behavior. This article presents a first approximation to this solution. The results obtained in the experiments are promising and a clear relationship between the parameters that define the gaze behavior of the user with the score assigned to the designs can be seen.

Juan García-Saravia, Lorenzo Salas-Morera, Laura García-Hernández, Adoración Antolí Cabrera

Active Sensing in Human Activity Recognition

This work studies the problem of reducing the energy consumption of wearable sensors in a Human Activity Recognition (HAR) system. A HAR system is implemented using Hidden Markov Models, where decisions over the acquisition of new data are made based on the entropy of the posterior distribution of the activities. This problem is intractable in general, so three different active sensing algorithms are implemented to find numerically the data acquisition events. The performance of these algorithms is evaluated using a HAR database, resulting in a significant reduction on the number of observations acquired, thus reducing the energy consumption, while maintaining the performance of the system.

Alfredo Nazábal, Antonio Artés

Searching the Sky for Neural Networks

Sky computing is a new computing paradigm leveraging resources of multiple Cloud providers to create a large-scale distributed infrastructure. N2Sky is a research initiative promising a framework for the utilization of Neural Networks as services across many Clouds integrating into a Sky. This involves a number of challenges ranging from the provision, discovery and utilization of N2Sky services to the management, monitoring, metering and accounting of the N2Sky infrastructure. This paper focuses on the semantic discovery of N2Sky services through a human-centered querying mechanism termed as N2Query. N2Query allows N2Sky users to specify their problem statement as natural language queries. In response to the natural language queries, it delivers a list of ranked neural network services to the user as a solution to their stated problem. The search algorithm of N2Query is based on the semantic mapping of ontologies referring to problem and solution domains.

Erich Schikuta, Abdelkader Magdy, Irfan Ul Haq, A. Baith Mohamed, Benedikt Pittl, Werner Mach

Image and Signal Processing

Frontmatter

Non-linear Least Mean Squares Prediction Based on Non-Gaussian Mixtures

Independent Component Analyzers Mixture Models (ICAMM) are versatile and general models for a large variety of probability density functions. In this paper, we assume ICAMM to derive a closed-form solution to the optimal Least Mean Squared Error predictor, which we have named E-ICAMM. The new predictor is compared with four classical alternatives (Kriging, Wiener, Matrix Completion, and Splines) which are representative of the large amount of existing approaches. The prediction performance of the considered methods was estimated using four performance indicators on simulated and real data. The experiment on real data consisted in the recovering of missing seismic traces in a real seismology survey. E-ICAMM outperformed the other methods in all cases, displaying the potential of the derived predictor.

Gonzalo Safont, Addisson Salazar, Alberto Rodríguez, Luis Vergara

Synchronized Multi-chain Mixture of Independent Component Analyzers

This paper presents a novel method for modeling the joint behavior of a number of synchronized Independent Component Analysis Mixture Models (ICAMM), which we have named Multi-chain ICAMM (MCICAMM). This allows flexible estimation of complex densities of data, subspace classification, blind source separation, accurate local dynamic learning, and global dynamic interaction. Furthermore, the proposed method can also be used for classification following the maximum a posteriori, forward-backward, or Viterbi procedures. MCICAMM outperformed competitive methods such as ICAMM, SICAMM, and Dynamic Bayesian Networks for the classification of simulated data and the automatic staging of electroencephalographic (EEG) data from epileptic patients performing a neuropsychological test for short-term memory. Therefore, the potential of the method to suit different kind of data densities and to deal with the changing non-stationarity and non-linearity of brain dynamics was demonstrated. MCICAMM parameters provide a structured result that might be interpreted in several applications.

Gonzalo Safont, Addisson Salazar, Ahmed Bouziane, Luis Vergara

Pooling Spike Neural Network for Acceleration of Global Illumination Rendering

The generation of photo-realistic images is a major topic in computer graphics. By using the principles of physical light propagation, images that are indistinguishable from real photographs can be generated. However, this computation is a very time-consuming task. When simulating the real behavior of light, individual images can take hours to be of sufficient quality. This paper proposes a bio-inspired architecture with spiking neurons for acceleration of global illumination rendering. This architecture with functional parts of sparse encoding, learning and decoding consists of a robust convergence measure on blocks. Feature, concatenation and prediction pooling coupled with three pooling operators: convolution, average and standard deviation are used in order to separate noise from signal. The pooling spike neural network (PSNN) represents a non-linear mapping from stochastic noise features of rendering images to their quality visual scores. The system dynamic, that computes a learning parameter for each image based on its level of noise, is a consistent temporal framework where the precise timing of spikes is employed for information processing. The experiments are conducted on a global illumination set which contains diverse image distortions and large number of images with different noise levels. The result of this study is a system composed from only two spike pattern association neurons (SPANs) suitably adopted to the quality assessment task that accurately predict the quality of images with a high agreement with respect to human psycho-visual scores. The proposed spike neural network has also been compared with support vector machine (SVM). The obtained results show that the proposed method gives promising efficiency.

Joseph Constantin, Andre Bigand, Ibtissam Constantin

Automatic Recognition of Daily Physical Activities for an Intelligent-Portable Oxygen Concentrator (iPOC)

In recent years, new autonomous physiological close-loop controlled (PCLC) medical devices for oxygen delivery are being researched. Most of this PCLC devices are based on the feedback of arterial oxygen saturation, measured using a pulse oximeter. However, pulse oximeters may provide spuriously low or high SpO2 values. In this work, a different approach to adjust automatically oxygen dosing in portable oxygen concentrators (POC) according to the physical activity performed by patients with COPD is presented. To that purpose, the ability of various machine-learning algorithms to recognize four human daily activities from sensor signals collected from a single waist-worn tri-axial accelerometer is evaluated. A set of 56 features was considered and recognition accuracy of up to 91.15% on the four activities of daily living was obtained using a SVM classifier. The associated activity recognition error rate was lower than 5%, ensuring a low percentage of time wrongly assigned to a certain activity. The underlying idea is the hardware implementation of the SVM classifier to control the oxygen flow in intelligent portable oxygen concentrators.

Daniel Sanchez-Morillo, Osama Olaby, Miguel Angel Fernandez-Granero, Antonio Leon-Jimenez

Automatic Detection of Epiretinal Membrane in OCT Images by Means of Local Luminosity Patterns

This work presents a novel approach for automatic detection of the epiretinal membrane in Optical Coherence Tomography (OCT) images. A tool able to detect this pathology is very valued since it can prevent further ocular damage by doing an early detection. This approach is based in the location of the inner limiting membrane (ILM) layers of the retina. Then, the detected locations are classified using a local-feature based vector in order to determine presence of the membrane. Different tests are run and compared to establish the appropriateness of the approach as well as its practical validity.

Sergio Baamonde, Joaquim de Moura, Jorge Novo, Marcos Ortega

An Expert System Based on Using Artificial Neural Network and Region-Based Image Processing to Recognition Substantia Nigra and Atherosclerotic Plaques in B-Images: A Prospective Study

The presented paper is focused on ways of digital image analysis of ultrasound B-images based on echogenicity investigation in determined Region of Interest (ROI). An expert system has been developed in the course of the research. The goal of the paper is to demonstrate how to interconnect automatic finding of the position of the substantia nigra using Artificial Neural Network (ANN) with supervised learning and ROI-based image analysis. For substantia nigra is able to detect the position using ANN from B-image in transverse thalamic plane. From this is computed echogenicity index grade inside the ROI as parkinsonism feature. The methodology is well applicable for a set of images with the same resolution. The results have shown practical application of ANN learning in this case. The second part of the paper is focused on detection of atherosclerotic plaques. An experimental prospective study shown the using ANN can be highly time-consuming problem due to complexity of B-images. The plaques have no standardized shape and size in comparison with SN. To objective appraisal of using ANN to automatic finding atherosclerotic plaque in B-image we need a large set of images of normal and pathological state. Although it is very important using ANN, automatic detection in highly time-consuming problem for ANN training.

Jiří Blahuta, Tomáš Soukup, Jiri Martinu

Automatic Tool for Optic Disc and Cup Detection on Retinal Fundus Images

The aging of the population is a matter of concern due to its association with various diseases in humans that limit their quality of life. Among them, glaucoma is one of the leading causes of blindness in the world. To its early diagnose, retinal fundus images are visually inspected by experts. In recent years, image-based computer aided diagnosis systems have been proposed. Automatic segmentation of Optic Disc (OD) and cup areas are their first and most difficult tasks. In this paper, a computerized technique aimed to their extraction from the original images is presented. The tool is related to human perception due to the use of an advanced color metric, CIE94 within a uniform color space, CIE L*a*b* to compute pixels’ color gradients [1]. Based on this information, a classifier assigns a probability value to each of the pixels, meaning its suitability for being part of the Optic Disc and Cup border. The tool has been tested on 200 images from different public databases achieving an accuracy value of 96.63%. This quality level makes the proposed color-based image processing system capable to assist the physicians in glaucoma screening programs.

Miguel Angel Fernandez-Granero, Auxiliadora Sarmiento Vega, Anabel Isabel García, Daniel Sanchez-Morillo, Soledad Jiménez, Pedro Alemany, Irene Fondón

2C-SVM Based Radar Detectors in Gaussian and K-Distributed Real Interference

This paper tackles the design and evaluation of cost sensitive Support Vector Machine (2C-SVM) based radar detectors in presence of Gaussian and K-Distributed clutter. 2C-SVM based solutions are able to approximate the Neyman-Pearson detector for a specific false alarm rate ($$P_{FA}$$). Real data acquired in different wind conditions by a coherent, pulsed and X-Band radar were considered. A statistical analysis is carried out to design the 2C-SVM for detecting targets with unknown parameters in Gaussian and non-Gaussian interference. A grid search of the best training parameters to approximate the pair detection probability ($$P_D$$) and $$P_{FA}$$ of the NP detector is required. Results prove the capability of the 2C-SVM based detectors to maximize the $$P_D$$ for a desired $$P_{FA}$$ independently of the detection problem likelihood functions.

David Mata-Moya, Maria-Pilar Jarabo-Amores, Manuel Rosa-Zurera, Javier Rosado-Sanz, Nerea del-Rey-Maestre

Uncertainty Analysis of ANN Based Spectral Analysis Using Monte Carlo Method

Uncertainty analysis of an Artificial Neural Network (ANN) based method for spectral analysis of asynchronously sampled signals is performed. Main uncertainty components contributions, jitter and quantization noise, are considered in order to obtain the signal amplitude and phase uncertainties using Monte Carlo method. The analysis performed identifies also uncertainties main contributions depending on parameters configurations. The analysis is performed simultaneously with the proposed method and two others: Discrete Fourier Transform (DFT) and Multiharmonic Sine Fitting Method (MSFM), in order to compare them in terms of uncertainty. Results show the proposed method has the same uncertainty as DFT for amplitude values and around double uncertainty in phase values.

José Ramón Salinas, Francisco García-Lagos, Javier Díaz de Aguilar, Gonzalo Joya, Francisco Sandoval

Using Deep Learning for Image Similarity in Product Matching

Product matching aims at disambiguating descriptions of products belonging to different websites in order to be able to recognize identical elements and to merge the content from those identical items. Most approaches face this matter applying various machine learning methods to textual product descriptions. Recently some authors are including information extracted from an image associated to a textual description of a product. Modern machine learning methods, such as content based information retrieval (CBIR) or deep learning, can be applied to this type of images since they can manage very large data sets for finding hidden structure within them, and for making accurate predictions. This information could boost the performance of the traditional textual matching but at the same time increase the computational complexity of the process. In this paper we review some CBIR and deep learning models and analyse the performance of these approaches when they are applied to images for product matching. The results obtained will help to introduce a combined classifier using textual and image information.

Mario Rivas-Sánchez, Maria De La Paz Guerrero-Lebrero, Elisa Guerrero, Guillermo Bárcena-Gonzalez, Jaime Martel, Pedro L. Galindo

Enhanced Similarity Measure for Sparse Subspace Clustering Method

Trying to find clusters in high dimensional data is one of the most challenging issues in machine learning. Within this context, subspace clustering methods have showed interesting results especially when applied in computer vision tasks. The key idea of these methods is to uncover groups of data that are embedding in multiple underlying subspaces. In this spirit, numerous subspace clustering algorithms have been proposed. One of them is Sparse Subspace Clustering (SSC) which has presented notable clustering accuracy. In this paper, the problem of similarity measure used in the affinity matrix construction in the SSC method is discussed. Assessment on motion segmentation and face clustering highlights the increase of the clustering accuracy brought by the enhanced SSC compared to other state-of-the-art subspace clustering methods.

Sabra Hechmi, Abir Gallas, Ezzeddine Zagrouba

Mathematics for Neural Networks

Frontmatter

Neural Network-Based Simultaneous Estimation of Actuator and Sensor Faults

The paper is devoted to the problem of a neural network-based robust simultaneous actuator and sensor faults estimator design for the purpose of the Fault Diagnosis (FD) of non-linear systems. In particular, the methodology of designing a neural network-based $$\mathcal {H_\infty }$$ fault estimator is developed. The main novelty of the approach is associated with possibly simultaneous sensor and actuator faults. For this purpose, a Linear Parameter Varying (LPV) description of a Recurrent Neural Network (RNN) is exploited. The proposed approach guaranties a predefined disturbance attenuation level and convergence of the estimator. The final part of the paper presents an illustrative example concerning the application of the proposed approach to the multi-tank system fault diagnosis.

Marcin Pazera, Marcin Witczak, Marcin Mrugalski

Exploring a Mathematical Model of Gain Control via Lateral Inhibition in the Antennal Lobe

Bioinspired Neural Networks have in many instances paved the way for significant discoveries in Statistical and Machine Learning. Among the many mechanisms employed by biological systems to implement learning, gain control is a ubiquitous and essential component that guarantees standard representation of patterns for improved performance in pattern recognition tasks. Gain control is particularly important for the identification of different odor molecules, regardless of their concentration. In this paper, we explore the functional impact of a biologically plausible model of the gain control on classification performance by representing the olfactory system of insects with a Single Hidden Layer Network (SHLN). Common to all insects, the primary olfactory pathway starts at the Antennal Lobes (ALs) and, then, odor identity is computed at the output of the Mushroom Bodies (MBs). We show that gain-control based on lateral inhibition in the Antennal Lobe robustly solves the classification of highly-concentrated odors. Furthermore, the proposed mechanism does not depend on learning at the AL level, in agreement with biological literature. Due to its simplicity, this bioinspired mechanism may not only be present in other neural systems but can also be further explored for applications, for instance, involving electronic noses.

Aaron Montero, Thiago Mosqueiro, Ramon Huerta, Francisco B. Rodriguez

Optimal Spherical Separability: Artificial Neural Networks

In this research paper, the concept of hyper-spherical/hyper-ellipsoidal separability is introduced. Method of arriving at the optimal hypersphere (maximizing margin) separating two classes is discussed. By projecting the quantized patterns into higher dimensional space (as in encoders of error correcting code), the patterns are made hyper-spherically separable. Single/multiple layers of spherical/ellipsoidal neurons are proposed for multi-class classification. An associative memory based on hyper-ellipsoidal neuron is proposed.

Rama Murthy Garimella, Ganesh Yaparla, Rhishi Pratap Singh

Pre-emphasizing Binarized Ensembles to Improve Classification Performance

Machine ensembles are learning architectures that offer high expressive capacities and, consequently, remarkable performances. This is due to their high number of trainable parameters.In this paper, we explore and discuss whether binarization techniques are effective to improve standard diversification methods and if a simple additional trick, consisting in weighting the training examples, allows to obtain better results. Experimental results, for three selected classification problems, show that binarization permits that standard direct diversification methods (bagging, in particular) achieve better results, obtaining even more significant performance improvements when pre-emphasizing the training samples. Some research avenues that this finding opens are mentioned in the conclusions.

Lorena Álvarez-Pérez, Anas Ahachad, Aníbal R. Figueiras-Vidal

Dynamics of Quaternionic Hopfield Type Neural Networks

In this research paper, a novel ordinary quaternionic hopfield type network is proposed and the associated convergence theorem is proved. Also, a novel structured quaternionic recurrent hopfield network is proposed. It is proved that in the parallel mode of operation, such a network converges to a cycle of length 4.

Rama Murthy Garimella, Rayala Anil

Quasi-Newton Learning Methods for Quaternion-Valued Neural Networks

This paper presents the deduction of the quasi-Newton learning methods for training quaternion-valued feedforward neural networks, using the framework of the HR calculus. Since these algorithms yielded better training results than the gradient descent for the real- and complex-valued cases, an extension to the quaternion-valued case is a natural idea to enhance the performance of quaternion-valued neural networks. Experiments done on four time series prediction applications show a significant improvement over the quaternion gradient descent algorithm.

Călin-Adrian Popa

Exponential Stability for Delayed Octonion-Valued Recurrent Neural Networks

Over the last few years, neural networks with values in multidimensional domains have gained a lot of interest. A non-associative normed division algebra which generalizes the complex and quaternion algebras is represented by the octonion algebra. It does not fall into the category of Clifford algebras, which are associative. Delayed octonion-valued recurrent neural networks are introduced, for which the states and weights are octonions. A sufficient criterion is given in the form of linear matrix inequalities, which assures the global exponential stability of the equilibrium point for the proposed networks. Lastly, a numerical example illustrates the correctness of the theoretical results.

Călin-Adrian Popa

Forward Stagewise Regression on Incomplete Datasets

The Forward Stagewise Regression (FSR) algorithm is a popular procedure to generate sparse linear regression models. However, the standard FSR assumes that the data are fully observed. This assumption is often flawed and pre-processing steps are applied to the dataset so that FSR can be used. In this paper, we extend the FSR algorithm to directly handle datasets with partially observed feature vectors, dismissing the need for the data to be pre-processed. Experiments were carried out on real-world datasets and the proposed method reported promising results when compared to the usual strategies for handling incomplete data.

Marcelo B. A. Veras, Diego P. P. Mesquita, João P. P. Gomes, Amauri H. Souza Junior, Guilherme A. Barreto

Convolutional Neural Networks with the F-transform Kernels

We propose a new convolutional neural network – the FTNet and explain its theoretical background referring to the theory of a higher degree F-transform. The FTNet is parametrized by kernel sizes, on/off activation of weights learning, the choice of strides or pooling, etc. It is trained on the database MNIST and tested on handwritten inputs. The obtained results demonstrate that the FTNet has better recognition accuracy than the automatically trained LENET-5. We have also analyzed the FTNet and LENET-5 rotation invariance.

Vojtech Molek, Irina Perfilieva

Class Switching Ensembles for Ordinal Regression

The term ordinal regression refers to classification tasks in which the categories have a natural ordering. The main premise of this learning paradigm is that the ordering can be exploited to generate more accurate predictors. The goal of this work is to design class switching ensembles that take into account such ordering so that they are more accurate in ordinal regression problems. In standard (nominal) class switching ensembles, diversity among the members of the ensemble is induced by injecting noise in the class labels of the training instances. Assuming that the classes are interchangeable, the labels are modified at random. In ordinal class switching, the ordering between classes is taken into account by reducing the transition probabilities to classes that are further apart. In this manner smaller label perturbations in the ordinal scale are favoured. Two different specifications of these transition probabilities are considered; namely, an arithmetic and a geometric decrease with the absolute difference of the class ranks. These types of ordinal class switching ensembles are compared with an ensemble method that does not consider class-switching, a nominal class-switching ensemble, an ordinal variant of boosting, and two state-of-the-art ordinal classifiers based on support vector machines and Gaussian processes, respectively. These methods are evaluated and compared in a total of 15 datasets, using three different performance metrics. From the results of this evaluation one concludes that ordinal class-switching ensembles are more accurate than standard class-switching ones and than the ordinal ensemble method considered. Furthermore, their performance is comparable to the state-of-the-art ordinal regression methods considered in the analysis. Thus, class switching ensembles with specifically designed transition probabilities, which take into account the relationships between classes, are shown to provide very accurate predictions in ordinal regression problems.

Pedro Antonio Gutiérrez, María Pérez-Ortiz, Alberto Suárez

Attractor Basin Analysis of the Hopfield Model: The Generalized Quadratic Knapsack Problem

The Continuous Hopfield Neural Network (CHN) is a neural network which can be used to solve some optimization problems. The weights of the network are selected based upon a set of parameters which are deduced by mapping the optimization problem to its associated CHN. When the optimization problem is the Traveling Salesman Problem, for instance, this mapping process leaves one free parameter; as this parameter decreases, better solutions are obtained. For the general case, a Generalized Quadratic Knapsack Problem (GQKP), there are some free parameters which can be related to the saddle point of the CHN. Whereas in simple instances of the GQKP, this result guarantees that the global optimum is always obtained, in more complex instances, this is far more complicated. However, it is shown how in the surroundings of the saddle point the attractor basins for the best solutions grow as the free parameter decreases, making saddle point neighbors excellent starting point candidates for the CHN. Some technical results and some computational experiences validate this behavior.

Lucas García, Pedro M. Talaván, Javier Yáñez

A Systematic Approach for the Application of Restricted Boltzmann Machines in Network Intrusion Detection

A few exploratory works studied Restricted Boltzmann Machines (RBMs) as an approach for network intrusion detection, but did it in a rather empirical way. It is possible to go one step further taking advantage from already mature theoretical work in the area. In this paper, we use RBMs for network intrusion detection showing that it is capable of learning complex datasets. We also illustrate an integrated and systematic way of learning. We analyze learning procedures and applications of RBMs and show experimental results for training RBMs on a standard network intrusion detection dataset.

Arnaldo Gouveia, Miguel Correia

Selecting the Coherence Notion in Multi-adjoint Normal Logic Programming

This paper is focused on looking for an appropriate coherence notion which allows us to deal with inconsistent information included in multi-adjoint normal logic programs. Different definitions closely related to the inconsistency concept have been studied and an adaptation of them to our logic programming framework has been included. A detailed reasoning is presented in order to motivate and justify the suitability of the chosen coherence notion.

M. Eugenia Cornejo, David Lobo, Jesús Medina

Gaussian Opposite Maps for Reduced-Set Relevance Vector Machines

The Relevance Vector Machine is a bayesian method. This model represents its decision boundary using a subset of points from the training set, called relevance vectors. The training algorithm of that is time consuming. In this paper we propose a technique for initialize the training process using the points of an opposite map in classification problems. This solution approximate the relevance points of the solutions obtained by Support Vector Machines. In order to assess the performance of our proposal, we carried out experiments on well-known datasets against the original RVM and SVM. The GOM-RVM achieved accuracy equivalent or superior than to SVM and RVM with fewer relevance vectors.

Lucas Silva de Sousa, Ajalmar Rêgo da Rocha Neto

Self-organizing Networks

Frontmatter

Massive Parallel Self-organizing Map and 2-Opt on GPU to Large Scale TSP

This paper proposes a platform both for parallelism of self-organizing map (SOM) and the 2-opt algorithm to large scale 2-Dimensional Euclidean traveling salesman problems. This platform makes these two algorithms working in a massively parallel way on graphical processing unit (GPU). Advantages of this platform include its flexibly topology preserving network, its fine parallel granularity and it allows maximum (N / 3) 2-opt optimization moves to be executed with O(N) complexity within one tour orientation and does not cut the integral tour. The parallel technique follows data decomposition and decentralized control. We test this optimization method on large TSPLIB instances, experiments show that the acceleration factor we obtained makes the proposed method competitive, and allows for further increasing for very large TSP instances along with the quantity increase of physical cores in GPU systems.

Wen-bao Qiao, Jean-charles Créput

Finding Self-organized Criticality in Collaborative Work via Repository Mining

In order to improve team productivity and the team interaction itself, as well as the willingness of occasional volunteers, it is interesting to study the dynamics underlying collaboration in a repository-mediated project and their mechanisms, because the mechanisms producing those dynamics are not explicit or organized from the top, which allows self organization to emerge from the collaboration and the way it is done. This is why finding if self-organization takes place and under which conditions will yield some insights on this process, and, from this, we can deduce some hints on how to improve it. In this paper we will focus on the former, examining repositories where collaborative writing of scientific papers by our research team is taking place show the characteristics of a critical state, which can be measured by the existence of a scale-free structure, long-distance correlations and pink noise when analyzing the size of changes and its time series. This critical state is reached via self-organization, which is why it is called self-organized criticality. Our intention is to prove that, although with different characteristics, most repositories independently of the number of collaborators and their real nature, self-organize, which implies that it is the nature of the interactions, and not the object of the interaction, which takes the project to a critical state. This critical state has already been established in a number of repositories with different types of projects, such as software or even literary works; we will also find if there is any essential difference between the macro measures of the states reached by these and the object of this paper.

J. J. Merelo, Pedro A. Castillo, Mario García-Valdez

Capacity and Retrieval of a Modular Set of Diluted Attractor Networks with Respect to the Global Number of Neurons

The modularity and hierarchical structures in associative networks can replicate parallel pattern retrieval and multitasking abilities found in complex neural systems. These properties can be exhibited in an ensemble of diluted Attractor Neural Networks for pattern retrieval. It has been shown in a previous work that this modular structure increases the single attractor storage capacity using a divide-and-conquer approach of subnetwork diluted modules. Each diluted module in the ensemble learns disjoint subsets of unbiased binary patterns. The present article deals with an ensemble of diluted Attractor Neural Networks which is studied for different values of the global number of network units, and their performance is compared with a single fully connected network keeping the same cost (total number of connections). The ensemble system more than doubles the maximal capacity of the single network with the same wiring cost. The presented approach can be useful for engineering applications to limited memory systems such as embedded systems or smartphones.

Mario González, David Dominguez, Ángel Sánchez, Francisco B. Rodríguez

Opposite-to-Noise ARTMAP Neural Network

Fuzzy ARTMAP (FAM) aims to solve the stability-plasticity dilemma by the adaptive resonance theory (ART). Despite this advantage, category proliferation is an important drawback in Fuzzy ARTMAP due mostly to the overlapping region (noise) between classes. In such a region, the match tracking mechanism is often triggered by raising the vigilance parameter value to avoid future learning errors. In order to overcome this drawback, we propose a Fuzzy ARTMAP-based architecture robust to noise, named OnARTMAP. Our proposal has a two-stage learning process. The first stage requires two new modules, the overlapping region detection module (ORDM) and another one very similar to $$ART_a$$, called $$ART_c$$. The ORDM finds the overlapping region between categories and the second one ($$ART_c$$) computes and stores special categories for overlapping areas (overlapping categories). In the second stage, the weights for conventional categories are estimated from data outside the overlapping area. Consequently, by not considering noise data, the number of categories drops considerably. We can infer from achievements that our proposal in general outperformed Fuzzy ARTMAP, ART-EMAP, $$\mu $$ARTMAP, and BARTMAP and achieved good data generalization with fewer categories and robustness on noise.

Alan Matias, Ajalmar Rocha Neto, Atslands Rocha

Accuracy Improvement of Neural Networks Through Self-Organizing-Maps over Training Datasets

Although it is not a novel topic, pattern recognition has become very popular and relevant in the last years. Different classification systems like neural networks, support vector machines or even complex statistical methods have been used for this purpose. Several works have used these systems to classify animal behavior, mainly in an offline way. Their main problem is usually the data pre-processing step, because the better input data are, the higher may be the accuracy of the classification system. In previous papers by the authors an embedded implementation of a neural network was deployed on a portable device that was placed on animals. This approach allows the classification to be done online and in real time. This is one of the aims of the research project MINERVA, which is focused on monitoring wildlife in Doñana National Park using low power devices. Many difficulties were faced when pre-processing methods quality needed to be evaluated. In this work, a novel pre-processing evaluation system based on self-organizing maps (SOM) to measure the quality of the neural network training dataset is presented. The paper is focused on a three different horse gaits classification study. Preliminary results show that a better SOM output map matches with the embedded ANN classification hit improvement.

Daniel Gutierrez-Galan, Juan Pedro Dominguez-Morales, Ricardo Tapiador-Morales, Antonio Rios-Navarro, Manuel Jesus Dominguez-Morales, Angel Jimenez-Fernandez, Alejandro Linares-Barranco

Spiking Neurons

Frontmatter

Computing with Biophysical and Hardware-Efficient Neural Models

In this paper we evaluate how seminal biophysical Hodgkin Huxley model and hardware-efficient TrueNorth model of spiking neurons can be used to perform computations on spike rates in frequency domain. This side-by-side evaluation allows us to draw connections how fundamental arithmetic operations can be realized by means of spiking neurons and what assumptions should be made on input to guarantee the correctness of the computed result. We validated our approach in simulation and consider this work as a first step towards FPGA hardware implementation of neuromorphic accelerators based on spiking models.

Konstantin Selyunin, Ramin M. Hasani, Denise Ratasich, Ezio Bartocci, Radu Grosu

A SpiNNaker Application: Design, Implementation and Validation of SCPGs

In this paper, we present the numerical results of the implementation of a Spiking Central Pattern Generator (SCPG) on a SpiNNaker board. The SCPG is a network of current-based leaky integrate-and-fire (LIF) neurons, which generates periodic spike trains that correspond to different locomotion gaits (i.e. walk, trot, run). To generate such patterns, the SCPG has been configured with different topologies, and its parameters have been experimentally estimated. To validate our designs, we have implemented them on the SpiNNaker board using PyNN and we have embedded it on a hexapod robot. The system includes a Dynamic Vision Sensor system able to command a pattern to the robot depending on the frequency of the events fired. The more activity the DVS produces, the faster that the pattern that is commanded will be.

Brayan Cuevas-Arteaga, Juan Pedro Dominguez-Morales, Horacio Rostro-Gonzalez, Andres Espinal, Angel F. Jimenez-Fernandez, Francisco Gomez-Rodriguez, Alejandro Linares-Barranco

Smart Hardware Implementation of Spiking Neural Networks

During last years a lot of attention have been focused to the hardware implementation of Artificial Neural Networks (ANN) to efficiently exploit the inherent parallelism associated to these systems. From the different types of ANN, the Spiking Neural Networks (SNN) arise as a promising bio-inspired model that is able to emulate the expected neural behavior with a high confidence. Many works are centered in using analog circuitry to reproduce SNN with a high degree of precision, while minimizing the area and the energy costs. Nevertheless, the reliability and flexibility of these systems is lower if compared with digital implementations. In this paper we present a new, low-cost bio-inspired digital neural model for SNN along with an auxiliary Computer Aided Design (CAD) tool for the efficient implementation of high-volume SNN.

Fabio Galán-Prado, Josep L. Rosselló

An Extended Algorithm Using Adaptation of Momentum and Learning Rate for Spiking Neurons Emitting Multiple Spikes

This paper presents two methods of using the dynamic momentum and learning rate adaption, to improve learning performance in spiking neural networks where neurons are modelled as spiking multiple times. The optimum value for the momentum factor is obtained from the mean square error with respect to the gradient of synaptic weights in the proposed algorithm. The delta-bar-delta rule is employed as the learning rate adaptation method. The XOR and Wisconsin breast cancer (WBC) classification tasks are used to validate the proposed algorithms. Results demonstrate no error and a minimal error of 0.08 are achieved for the XOR and WBC classification tasks respectively, which are better than the original Booij’s algorithm. The minimum number of epochs for XOR and Wisconsin breast cancer tasks are 35 and 26 respectively, which are also faster than the original Booij’s algorithm – i.e. 135 (for XOR) and 97 (for WBC). Compared with the original algorithm with static momentum and learning rate, the proposed dynamic algorithms can control the convergence rate and learning performance more effectively.

Yuling Luo, Qiang Fu, Junxiu Liu, Jim Harkin, Liam McDaid, Yi Cao

Development of Doped Graphene Oxide Resistive Memories for Applications Based on Neuromorphic Computing

Resistive random access memory ReRAM has attracted great attention due to its potential for flash memory replacement in next generation nonvolatile memory applications. Among the main characteristics of this type of memory, we have: low energy consumption, high-speed switching, durability, scalability and friendly manufacturing process. This device is based on resistive switching phenomenon for operation, which is reversible and can be played back repeatedly. In this work, eight different devices are developed and fabrication is made as follows: thin films are obtained by dip coating technique. The dip coating apparatus basically consists of a clamp which holds the substrate is dipped in a GO solution (graphene oxide) which containing dopant (cupper, iron or silver) or CuO (copper oxide). ITO (indium tin oxide) and aluminum contacts were evaporated. The devices were developed with purpose: intention is record and read information dynamically with appropriate algorithm. There is even the possibility of storing images. With these functions, it would be promising to enter the neuromorphic computing area that is one of the resistive memory applications. ReRAM technology advent represents a paradigm shift for artificial neural networks, being the best candidate for emulation of synaptic plasticity and learning mode.

Marina Sparvoli, Mauro F. P. Silva, Mario Gazziro

Artificial Neural Networks in Industry ANNI’17

Frontmatter

Performance Study of Different Metaheuristics for Diabetes Diagnosis

The problem of medical data classification involves an optimization phase that may be solved through metaheuristic approaches. In this work, we evaluate the performance in diagnosis of diabetes disease, using Particle Swarm Optimization (PSO), Firefly (FF) and Homogeneity-Based Algorithm (HBA) metaheuristics in conjunction with fuzzy system. Here, the fitness function in the optimization process is the total misclassification cost that is in term of false positive, false negative and unclassifiable rates. The results prove that HBA approach achieves better results than the other metaheuristics. With execution time, FF was faster than the PSO and HBA methods.

Fatima Bekaddour, Mohamed Ben Rahmoune, Chikhi Salim, Ahmed Hafaifa

Randomized Neural Networks for Recursive System Identification in the Presence of Outliers: A Performance Comparison

In this paper, randomized single-hidden layer feedforward networks (SLFNs) are extended to handle outliers sequentially in online system identification tasks involving large-scale datasets. Starting from the description of the original batch learning algorithms of the evaluated randomized SLFNs, we discuss how these neural architectures can be easily adapted to cope with sequential data by means of the famed least mean squares (LMS). In addition, a robust variant of this rule, known as the least mean M-estimate (LMM) rule, is used to cope with outliers. Comprehensive performance comparison on benchmarking datasets are carried out in order to assess the validity of the proposed methodology.

César Lincoln C. Mattos, Guilherme A. Barreto, Gonzalo Acuña

Neural Network Overtopping Predictor Proof of Concept

Wave overtopping is a dangerous phenomenon. When it occurs in a commercial port environment, the best case scenario will be the disruption of activities and even this best case scenario has a negative financial repercussion.Being in disposal of a system that predicts overtopping events would provide valuable information, allowing the minimization of the impact of overtopping: the financial impact, the property damage or even physical harm to port workers.We designed an overtopping predictor and implemented a proof of concept based on neural networks. To carry out the proof of concept of the system, we created a series of tests in a scaled breakwater physical model, placed on a wave basin. We used a multidirectional wavemaker and video cameras to identify the overtopping events. Using all of the collected data we trained a neural network model that predicts an overtopping based on the simulated sea state.Once the validity of this approach is determined, we propose the real system design and the resources needed for its implementation.

Alberto Alvarellos, Enrique Peña, Andrés Figuero, José Sande, Juan Rabuñal

Artificial Neural Networks Based Approaches for the Prediction of Mean Flow Stress in Hot Rolling of Steel

The problem of the estimation of mean flow stress within a hot rolling mill plant for flat steel products is faced, as the correct estimation of this measure can improve the quality of the final product. Various approaches, from standard empirical methods to advanced architectures based on neural networks, have been tested on industrial data. The results of these tests put into evidence the limit of empirical techniques and the big advantages deriving from the application of neural networks, which are able to efficiently combine process knowledge and data driven models tuning. The best performing approaches reduce the estimation error to one third with respect to standard techniques.

Marco Vannucci, Valentina Colla, Vincenzo Iannino

Machine Learning for Renewable Energy Applications

Frontmatter

State of Health Estimation of Zinc Air Batteries Using Neural Networks

One major problem of energy storages is degradation. Degradation leads to a loss of capacity and a higher series resistance. One possibility to determine the state of health is the electrochemical impedance spectroscopy. The ac resistance is therefore measured for a set of different frequencies. Previous approaches match the measured impedances with a nonlinear equivalent circuit, which needs a lot of time to solve a nonlinear least squares problem. This paper combines the electrochemical impedance spectroscopy with neural networks to directly model the state of health in order to speed up the estimation. Zinc air batteries are exemplary used as energy storage, as other problems exists, that can be solved by impedance measurements. Optimizing a cost function is used to determine the fastest combination of examined frequencies.

Andre Loechte, Daniel Heming, Klaus T. Kallis, Peter Gloesekoetter

Bayesian Optimization of a Hybrid Prediction System for Optimal Wave Energy Estimation Problems

In the last years, Bayesian optimization (BO) has emerged as a practical tool for high-quality parameter selection in prediction systems. BO methods are useful for optimizing black-box objective functions that either lack an analytical expression, or are very expensive to evaluate. In this paper we show how BO can be used to obtain optimal parameters of a prediction system for a problem of wave energy flux prediction. Specifically, we propose the Bayesian optimization of a hybrid Grouping Genetic Algorithm with an Extreme Learning Machine (GGA-ELM) approach. The system uses data from neighbor stations (usually buoys) in order to predict the wave energy at a goal marine energy facility. The proposed BO methodology has been tested in a real problem involving buoys data in the Western coast of the USA, improving the performance of the GGA-ELM without a BO approach.

Laura Cornejo-Bueno, Eduardo C. Garrido-Merchán, Daniel Hernández-Lobato, Sancho Salcedo-Sanz

Hybrid Model for Large Scale Forecasting of Power Consumption

After the electricity liberalization in Europe, the electricity market moved to a more competitive supply market with higher efficiency in power production. As a result of this competitiveness, accurate models for forecasting long-term power consumption become essential for electric utilities as they help operating and planning of the utility’s facilities including Transmission and Distribution (T&D) equipments. In this paper, we develop a multi-step statistical analysis approach to interpret the correlation between power consumption of residential as well as industrial buildings and its main potential driving factors using the dataset of the Irish Commission for Energy Regulation (CER). In addition we design a hybrid model for forecasting long-term daily power consumption on the scale of portfolio of buildings using the models of conditional inference trees and linear regression. Based on an extensive evaluation study, our model outperforms two robust machine learning algorithms, namely random forests (RF) and conditional inference tree (ctree) algorithms in terms of time efficiency and prediction accuracy for individual buildings as well as for a portfolio of buildings. The proposed model reveals that dividing buildings in homogeneous groups, based on their characteristics and inhabitants demographics, can increase the prediction accuracy and improve the time efficiency.

Wael Alkhatib, Alaa Alhamoud, Doreen Böhnstedt, Ralf Steinmetz

A Coral Reef Optimization Algorithm for Wave Height Time Series Segmentation Problems

Time series segmentation can be approached using metaheuristics procedures such as genetic algorithms (GAs) methods, with the purpose of automatically finding segments and determine similarities in the time series with the lowest possible clustering error. In this way, segments belonging to the same cluster must have similar properties, and the dissimilarity between segments of different clusters should be the highest possible. In this paper we tackle a specific problem of significant wave height time series segmentation, with application in coastal and ocean engineering. The basic idea in this case is that similarity between segments can be used to characterise those segments with high significant wave heights, and then being able to predict them. A recently metaheuristic, the Coral Reef Optimization (CRO) algorithm is proposed for this task, and we analyze its performance by comparing it with that of a GA in three wave height time series collected in three real buoys (two of them in the Gulf of Alaska and another one in Puerto Rico). The results show that the CRO performance is better than the GA in this problem of time series segmentation, due to the better exploration of the search space obtained with the CRO.

Antonio Manuel Durán-Rosal, David Guijo-Rubio, Pedro Antonio Gutiérrez, Sancho Salcedo-Sanz, César Hervás-Martínez

Satellite Based Nowcasting of PV Energy over Peninsular Spain

In this work we will study the use of satellite-measured irradiances as well as clear sky radiance estimates as features for the nowcasting of photovoltaic energy productions over Peninsular Spain. We will work with three Machine Learning models (Lasso and linear and Gaussian Support Vector Regression-SVR) plus a simple persistence model. We consider prediction horizons of up to three hours, for which Gaussian SVR is the clear winner, with a quite good performance and whose errors increase slowly with time. Possible ways to further improve these results are also proposed.

Alejandro Catalina, Alberto Torres-Barrán, José R. Dorronsoro

A Study on Feature Selection Methods for Wind Energy Prediction

This work deals with wind energy prediction using meteorological variables estimated by a Numerical Weather Prediction model in a grid around the wind farm of interest. Two machine learning techniques have been tested, Support Vector Machine and Gradient Boosting Regression, in order to study their performance and compare the results. The use of meteorological variables estimated in a grid generally implies a large number of inputs to the models and the performance of models might decrease. Hence, in this context, the use of feature selection algorithms might be interesting in order to improve the generalization capability of models and/or reduce the number of attributes. We have compared three feature selection techniques based on different paradigms: Principal Components Analysis, ReliefF, and Sequential Forward Selection. Energy production data has been obtained from the Sotavento experimental wind farm. Meteorological variables have been obtained from European Centre for Medium-Range Weather Forecasts, for a 5$$\,\times \,$$5 grid around Sotavento.

Rubén Martín-Vázquez, Ricardo Aler, Inés M. Galván

Combining Reservoir Computing and Over-Sampling for Ordinal Wind Power Ramp Prediction

Wind power ramp events (WPREs) are strong increases or decreases of wind speed in a short period of time. Predicting WPREs is of vital importance given that they can damage the turbines in a wind farm. In contrast to previous binary approaches (ramp versus non-ramp), a three-class prediction is proposed in this paper by considering: negative ramp, non-ramp and positive ramp, where the natural order of the events is clear. The independent variables used for prediction include past ramp function values and meteorological data obtained from physical models (reanalysis data). The proposed methodology is based on reservoir computing and an over-sampling process for alleviating the high degree of unbalance of the dataset (non-ramp events are much more frequent than ramps). The reservoir computing model is a modified echo state network composed by: a recurrent neural network layer, a nonlinear kernel mapping and an ordinal logistic regression, in such a way that the order of the classes can be exploited. The standard synthetic minority oversampling technique (SMOTE) is applied to the reservoir activations, given that the direct application over the input variables would damage its temporal structure. The performance of this proposal is compared to the original dataset (without over-sampling) and to nominal logistic regression, and the results obtained with the oversampled dataset and ordinal logistic regression are found to be more robust.

Manuel Dorado-Moreno, Laura Cornejo-Bueno, Pedro Antonio Gutiérrez, Luis Prieto, Sancho Salcedo-Sanz, César Hervás-Martínez

Arbitrated Ensemble for Solar Radiation Forecasting

Utility companies rely on solar radiation forecasting models to control the supply and demand of energy as well as the operability of the grid. They use these predictive models to schedule power plan operations, negotiate prices in the electricity market and improve the performance of solar technologies in general. This paper proposes a novel method for global horizontal irradiance forecasting. The method is based on an ensemble approach, in which individual competing models are arbitrated by a metalearning layer. The goal of arbitrating individual forecasters is to dynamically combine them according to their aptitude in the input data. We validate our proposed model for solar radiation forecasting using data collected by a real-world provider. The results from empirical experiments show that the proposed method is competitive with other methods, including current state-of-the-art methods used for time series forecasting tasks.

Vítor Cerqueira, Luís Torgo, Carlos Soares

Modeling the Transformation of Olive Tree Biomass into Bioethanol with Reg-CORBFN

Research in renewable energies is a global trend. One remarkable area is the biomass transformation into biotehanol, a fuel that can replace fossil fuels. A key step in this process is the pretreatment stage, where several variables are involved. The experimentation for determining the optimal values of these variables is expensive, therefore it is necessary to model this process. This paper focus on modeling the production of biotehanol from olive tree biomass by data mining methods. Notably, the authors present Reg-CO$$^2$$RBFN, an adaptation of a cooperative-competitive designing method for radial basis function networks. One of the main drawbacks in this modeling is the low number of instances in the data sets. To compare the results obtained by Reg-CO$$^2$$RBFN, other well-known data mining regression methods are used to model the transformation process.

Francisco Charte Ojeda, Inmaculada Romero Pulido, Antonio Jesús Rivera Rivas, Eulogio Castro Galiano

A Hybrid Neuro-Evolutionary Algorithm for Wind Power Ramp Events Detection

In this work, a hybrid system for wind power ramps events prediction in wind farms is proposed. The system is based on modelling the prediction problem as a binary classification problem from atmospheric reanalysis data inputs. On the other hand, a hybrid neuro-evolutive algorithm is proposed, which combines Artificial Neuronal Networks such as Extreme Learning Machines, with evolutionary algorithms to optimize the trained models. The phenomenon under study occurs with a very low probability, for this reason the problem is so unbalanced, and it is necessary to resort to techniques focused on obtain good results by means of a reduction of the samples from the majority class, as the SMOTE approach. A feature selection is performed by the evolutionary algorithm in order to choose the best trained model. Finally, this model is evaluated by a test set and its accuracy performance is given. The accuracy obtained in the results is quite good in terms of classification performance.

Laura Cornejo-Bueno, Adrián Aybar-Ruiz, Carlos Camacho-Gómez, Luis Prieto, Alberto Barea-Ropero, Sancho Salcedo-Sanz

Backmatter

Weitere Informationen

Premium Partner

    Bildnachweise