Skip to main content
main-content

Über dieses Buch

This book constitutes the refereed proceedings of the 13th International Conference on Engineering Applications of Neural Networks, EANN 2012, held in London, UK, in September 2012. The 49 revised full papers presented were carefully reviewed and selected from numerous submissions. The papers describe the applications of neural networks and other computational intelligence approaches to intelligent transport, environmental engineering, computer security, civil engineering, financial forecasting, virtual learning environments, language interpretation, bioinformatics and general engineering.

Inhaltsverzeichnis

Frontmatter

Elimination of a Catastrophic Destruction of a Memory in the Hopfield Model

For the standard Hopfield model a catastrophic destruction of the memory has place when the last is overfull (so called

catastrophic forgetting

). We eliminate the catastrophic forgetting assigning different weights to input patterns. As the weights one can use the frequencies of appearance of the patterns during the learning process. We show that only patterns whose weights are larger than some critical weight would be recognized. The case of the weights that are the terms of a geometric series is studied in details. The theoretical results are in good agreement with computer simulations.

Iakov Karandashev, Boris Kryzhanovsky, Leonid Litinskii

An Operational Riverflow Prediction System in Helmand River, Afghanistan Using Artificial Neural Networks

This study uses historical flow record to establish an operational riverflow prediction model in Helmand River using artificial neural networks (ANNs). The tool developed for this research demonstrates that the ANN model produces results with a very short turn-around time and with good accuracy. This river system used for this demonstration is quite complex and contains uncertainties associated with the historical record. These uncertainties include downstream flow rates that are not always higher than the combined upper stream values and only one continuously operating stream gage in the headwaters. With these characteristics, improvements in the hydrologic predictions are achieved by using a best additional gage search and a two-layered ANN strategy. Despite the gains demonstrated in this research, better simulation accuracy can be achieved by constructing a new knowledge base using more recent information on the hydrologic/hydraulic condition changes that have occurred since the available period for 1979. Follow-on research can also include developing extrapolation procedures for desired project events outside the range of the historical data and predictive error correction analysis.

Bernard Hsieh, Mark Jourdan

Optimization of Fuzzy Inference System Field Classifiers Using Genetic Algorithms and Simulated Annealing

A classification system that would aid businesses in selecting calls for analysis would improve the call recording selection process. This would assist in developing good automated self service applications. This paper details such a classification system for a pay beneficiary application. Fuzzy Inference System (FIS) classifiers were created. These classifiers were optimized using Genetic Algorithm (GA) and Simulated Annealing (SA). GA and SA performance in FIS classifier optimization were compared. Good results were achieved. In regards to computational efficiency, SA outperformed GA. When optimizing the FIS ’Say account’ and ’Say confirmation’ classifiers, GA is the preferred technique. Similarly, SA is the preferred method in FIS ’Say amount’ and ’Select beneficiary’ classifier optimization. GA and SA optimized FIS field classifier outperformed previously developed FIS field classifiers.

Pretesh B. Patel, Tshilidzi Marwala

Information Theoretic Self-organised Adaptation in Reservoirs for Temporal Memory Tasks

Recurrent neural networks of the Reservoir Computing (RC) type have been found useful in various time-series processing tasks with inherent non-linearity and requirements of temporal memory. Here with the aim to obtain extended temporal memory in generic delayed response tasks, we combine a generalised intrinsic plasticity mechanism with an information storage based neuron leak adaptation rule in a self-organised manner. This results in adaptation of neuron local memory in terms of leakage along with inherent homeostatic stability. Experimental results on two benchmark tasks confirm the extended performance of this system as compared to a static RC and RC with only intrinsic plasticity. Furthermore, we demonstrate the ability of the system to solve long temporal memory tasks via a simulated T-shaped maze navigation scenario.

Sakyasingha Dasgupta, Florentin Wörgötter, Poramate Manoonpong

Fuzzy-Logic Inference for Early Detection of Sleep Onset in Car Driver

Heart rate variability (HRV) is an important sign because it reflects the activity of the autonomic nervous system (ANS), which controls most of the physiological activity of the subjects, including sleep. The balance between the sympathetic and parasympathetic branches of the nervous system is an effective indicator of heart rhythm and, indirectly, heart rhythm is related to a patient’s state of wakefulness or sleep. In this paper we present a research that models a fuzzy logic inference engine for early detection of the onset of sleep in people driving a car or a public transportation vehicle. ANS activity reflected in the HRV signal is measured by electrocardiogram (ECG). Power spectrum density (PSD) is computed from the HRV signal and ANS frequency activity is then measured. Crisp measurements such as very low, low, and high HRV and low-to-high frequency ratio variability are fuzzified and evaluated by a set of fuzzy-logic rules that make inferences about the onset of sleep in automobile drivers. An experimental test environment has been developed to evaluate this method and its effectiveness.

Mario Malcangi, Salvatore Smirne

Object-Oriented Neurofuzzy Modeling and Control of a Binary Distillation Column by Using MODELICA

Neurofuzzy networks offer an alternative approach both for the identification and the control of nonlinear processes in process engineering. The lack of software tools for the design of controllers based on hybrid neural networks and fuzzy models is particularly pronounced in this field. MODELICA is an oriented-object environment widely used which allows system-level developers to perform rapid prototyping and testing. Such programming environment offers an intuitive approach to both adaptive modeling and control in a great variety of engineering disciplines. In this paper we have developed an oriented-object model of binary distillation column with nonlinear dynamics, and an ANFIS (Adaptive-Network-based Fuzzy Inference System) neurofuzzy scheme has been applied to derive both an identification model and a adaptive controller to regulate distillation composition. The results obtained demonstrate the effectiveness of the neurofuzzy control scheme when the plant’s dynamics is given by a set of nonlinear differential algebraic equations (DAE).

Javier Fernandez de Canete, Alfonso Garcia-Cerezo, Inmaculada Garcia-Moral, Pablo del Saz, Ernesto Ochoa

Evolving an Indoor Robotic Localization System Based on Wireless Networks

This work addresses the evolution of an Artificial Neural Network (ANN) to assist in the problem of indoor robotic localization. We investigate the design and building of an autonomous localization system based on information gathered from Wireless Networks (WN). The paper focuses on the evolved ANN which provides the position of one robot in a space, as in a Cartesian plane, corroborating with the Evolutionary Robotic research area and showing its practical viability. The proposed system was tested on several experiments, evaluating not only the impact of different evolutionary computation parameters but also the role of the transfer functions on the evolution of the ANN. Results show that slight variations in the parameters lead to huge differences on the evolution process and therefore in the accuracy of the robot position.

Gustavo Pessin, Fernando S. Osório, Jefferson R. Souza, Fausto G. Costa, Jó Ueyama, Denis F. Wolf, Torsten Braun, Patrícia A. Vargas

Employing ANN That Estimate Ozone in a Short-Term Scale When Monitoring Stations Malfunction

This paper describes the design, development and application, of an intelligent system (operating dynamically in an iterative manner) capable of short term forecasting the concentration of dangerous air pollutants in major urban centers. This effort is the first phase of the partial fulfillment of a wider research project that is related to the development of a real time multi agent network serving the same purpose. Short term forecasting of air pollutants is necessary for the proper feed of the real time multi agent system, when one or more sensors are damaged or malfunctioning. From this point of view the potential outcome of this research is very useful towards real time air pollution monitoring. A vast volume of actual data vectors are combined from several measurement stations located in the center of Athens. The final target is the continuous estimation of Ozone (O

3

) in the historical city center, considering the effect of primitive pollutants and meteorological conditions from neighboring stations. A group comprising of hundreds artificial neural networks has been developed, capable of estimating effectively the concentration of O

3

at a specific temporal point and also after 1, 2, 3 and 6 hours.

Antonios Papaleonidas, Lazaros Iliadis

An Ontology Based Approach to Designing Adaptive Lesson Plans in Military Training Simulators

Several classes of simulators have been designed in the military domain for training and operational analysis. Joint Operations Simulation System (JOpsSS) is a virtual warfare analysis system that has been developed for planning, operational analysis and evaluating joint operations. A major concern in the design and development of these simulators is the training lesson plans for trainees with different backgrounds. The design of intelligent lesson plans that are adaptable to the varied needs of the trainees is a challenging task. In this paper, we propose an ontology based design of training, learning and evaluation agents that has been used to design intelligent training systems. This approach has proved to be very effective in modelling complex and adaptive warfare scenarios. An ontology that represents the military domain concepts, context and data is used to represent and store the knowledge-base required for intelligent training simulators and designing intelligent lesson plans. The Instructor agent assesses each trainee from the past credentials for the level of lesson plans and builds a

concept- graph

. This is dynamically adapted to suite the level of trainee based on the responses and bridges the gap between the expected and actual competency level.

D. Vijay Rao, Ravi Shankar, Lazaros Iliadis, V. V. S. Sarma

A Continuous-Time Model of Analogue K-Winners-Take-All Neural Circuit

A continuous-time model of analogue K-winners-take-all (KWTA) neural circuit which is capable to extraction the K largest from any finite value N unknown distinct inputs, where 1 ≤ K < N, is presented. The model is described by one state equation with discontinuous right-hand side and output equation. A corresponding functional block diagram of the model is given as N feedforward and one feedback hardlimiting neurons, which is used to determine the dynamic shift of inputs. The model combines such properties as high accuracy and convergence speed, low computational and hardware implementation complexity, and independency on initial conditions. Simulation examples demonstrating the model performance are provided.

Pavlo V. Tymoshchuk

Network Intrusion Detection System Using Data Mining

The aim of this study is to simulate a network traffic analyzer that is part of an Intrusion Detection System - IDS, the main focus of research is data mining and for this type of application the steps that precede the data mining : data preparation (possibly involving cleaning data, data transformations, selecting subsets of records, data normalization) are considered fundamental for a good performance of the classifiers during the data mining stage. In this context, this paper discusses and presents as a contribution not only the classifiers that were used in the problem of intrusion detection, but also the initial stage of data preparation. Therefore, we tested the performance of three classifiers on the KDDCUP’99 benchmark intrusion detection dataset and selected the best classifiers. We initially tested a Decision Tree and a Neural Network using this dataset, suggesting improvements by reducing the number of attributes from 42 to 27 considering only two classes of detection, normal and intrusion. Finally, we tested the Decision Tree and Bayesian Network classifiers considering five classes of attack: Normal, DOS, U2R, R2L and Probing. The experimental results proved that the algorithms used achieved high detection rates (DR) and significant reduction of false positives (FP) for different types of network intrusions using limited computational resources.

Lídio Mauro Lima de Campos, Roberto Célio Limão de Oliveira, Mauro Roisenberg

A Near Linear Algorithm for Testing Linear Separability in Two Dimensions

We present a near linear algorithm for determining the linear separability of two sets of points in a two-dimensional space. That algorithm does not only detects the linear separability but also computes separation information. When the sets are linearly separable, the algorithm provides a description of a separation hyperplane. For non linearly separable cases, the algorithm indicates a negative answer and provides a hyperplane of partial separation that could be useful in the building of some classification systems.

Sylvain Contassot-Vivier, David Elizondo

A Training Algorithm for Locally Recurrent NN Based on Explicit Gradient of Error in Fault Detection Problems

In this work a diagnostic approach for nonlinear systems is presented. The diagnosis is performed resorting to a neural predictor of the output of the system, and by using the error prediction as a feature for the diagnosis. A locally recurrent neural network is used as predictor, after it has been trained on a reference behavior of the system. In order to model the system under test a novel training algorithm that uses an explicit calculation of the cost function gradient is proposed. The residuals of the prediction are affected by the deviation of the parameters from their nominal values. In this way, by a simple statistical analysis of the residuals, we can perform a diagnosis of the system. The Rössler hyperchaotic system is used as benchmark problem in order to validate the diagnostic neural approach proposed.

Sara Carcangiu, Augusto Montisci, Patrizia Boi

Measurement Correction for Multiple Sensors Using Modified Autoassociative Neural Networks

In industrial plants, the analysis of signals provided by monitoring sensors is a difficult task due to the high dimensionality of the data. This work proposes the use of Autoassociative Neural Networks trained with a Modified Robust Method in an online monitoring system for fault detection and self-correction of measurements generated by a large number of sensors. Unlike the existing models, the proposed system aims at using only one neural network to reconstruct faulty sensor signals. The model is evaluated with the use of a database containing measurements collected by industrial sensors that control and monitor an internal combustion engine. Results show that the proposed model is able to map and correct faulty sensor signals and achieve low error rates.

Javier Reyes Sanchez, Marley Vellasco, Ricardo Tanscheit

Visual Based Contour Detection by Using the Improved Short Path Finding

Contour detection is an important characteristic of human vision perception. Humans can easily find the objects contour in a complex visual scene; however, traditional computer vision cannot do well. This paper primarily concerned with how to track the objects contour using a human-like vision. In this article, we propose a biologically motivated computational model to track and detect the objects contour. Even the previous research has proposed some models by using the Dijkstra algorithm [1], our work is to mimic the human eye movement and imitate saccades in our humans. We use natural images with associated ground truth contour maps to assess the performance of the proposed operator regarding the detection of contours while suppressing texture edges. The results show that our method enhances contour detection in cluttered visual scenes more effectively than classical edge detectors proposed by other methods.

Jiawei Xu, Shigang Yue

Analysis of Electricity Consumption Profiles by Means of Dimensionality Reduction Techniques

The analysis of the daily electricity consumption profile of a building and its correlation with environmental factors make it possible to estimate its electricity demand. As an alternative to the traditional correlation analysis, a new approach is proposed to provide a detailed and visual analysis of the correlations between consumption and environmental variables. Since consumption profiles are normally characterized by many electrical variables, i.e., a high dimensional space, it is necessary to apply dimensionality reduction techniques that enable a projection of these data onto an easily interpretable 2D space. In this paper, several dimensionality reduction techniques are compared in order to determine the most appropriate one for the stated purpose. Later, the proposed approach uses the chosen algorithm to analyze the profiles of two public buildings located at the University of León.

Antonio Morán, Juan J. Fuertes, Miguel A. Prada, Serafín Alonso, Pablo Barrientos, Ignacio Díaz

Neural Networks for the Analysis of Mine-Induced Vibrations Transmission from Ground to Building Foundation

Problem of the transmission of mine-induced ground vibrations to building foundation is analysed in the paper. The maximal values of horizontal vibrations velocities (horizontal vibration components and resultant vibrations) are taken into account. Application of neural networks for the prediction of building foundation vibrations on the basis of ground vibrations is proposed. Standard back-propagation neural networks as well as recurrent cascade neural network systems were used. Experimental data obtained from the measurements of ground and actual structure vibrations were applied as the neural network training, validating and testing patterns. The obtained results lead to a conclusion that the neural technique gives results accurate enough for engineering practice.

Krystyna Kuzniar, Lukasz Chudyba

Backpropagation Neural Network Applications for a Welding Process Control Problem

The aim of this study is to develop predictive Artificial Neural Network (ANN) models for welding process control of a strategic product (155 mm. artillery ammunition) in armed forces’ inventories. The critical process about the production of product is the welding process. In this process, a rotating band is welded to the body of ammunition. This is a multi-input, multi-output process. In order to tackle problems in the welding process 2 different ANN models have been developed in this study. Model 1 is a Backpropagation Neural Network (BPNN) application used for classification of defective and defect-free products. Model 2 is a reverse BPNN application used for predicting input parameters given output values. In addition, with the help of models developed mean values of best values of some input parameters are found for a defect-free weld operation.

Adnan Aktepe, Süleyman Ersöz, Murat Lüy

Elastic Nets for Detection of Up-Regulated Genes in Microarrays

DNA analysis by microarrays is a powerful tool that allows replication of the RNA of hundreds of thousands of genes at the same time, generating a large amount of data in multidimensional space that must be analyzed using informatics tools. Various clustering techniques have been applied to analyze the microarrays, but they do not offer a systematic form of analysis. This paper proposes the use of Zinovyev’s

Elastic Net

in an iterative way to find patterns of up-regulated genes. The new method proposed has been evaluated with up-regulated genes of the Escherichia Coli k12 bacterium and is compared with the Self-Organizing Maps (SOM) technique frequently used in this kind of analysis. The results show that the proposed method finds

87%

of the up-regulated genes, compared to

65%

of genes found by the SOM. A comparative analysis of Receiver Operating Characteristic with SOM shows that the proposed method is

12%

more effective.

Marcos Levano, Alejandro Mellado

Detection and Classification of ECG Chaotic Components Using ANN Trained by Specially Simulated Data

This paper presents the use of simulated ECG signals with known chaotic and random noise combination for training of an Artificial Neural Network (ANN) as a classification tool for analysis of chaotic ECG components. Preliminary results show about 85% overall accuracy in the ability to classify signals into two types of chaotic maps – logistic and Henon. Robustness to random noise is also presented. Future research in the form of raw data analysis is proposed, and further features analysis is needed.

Polina Kurtser, Ofer Levi, Vladimir Gontar

Automatic Landmark Location for Analysis of Cardiac MRI Images

This paper addresses the problem of automatic location of landmarks used for the analysis of MRI cardiac images. Typically the landmarks of shapes in MRI images are located manually which is a time consuming process requiring human expertise and attention to detail. As an alternative a number of researchers use shape modelling and image search techniques for locating the required landmarks automatically. Usually these techniques require human expertise for initializing the search and in addition they require high quality, noise free images so that the image-based landmark location is successful. With our work we propose the use of neural network methods for learning the geometry of sets of points so that it is possible to predict the positions of all required landmarks based on the positions of a small subset of the landmarks rather than using image-data during the process of landmark-location. As part of our work the performance of neural network methods like Multilayer Perceptrons, Radial Basis Functions and Support Vector Machines is evaluated. Quantitative and visual results demonstrate the potential of using such methods for locating the required landmarks on endo-cardial and epicardial landmarks of the left ventricle of MRI cardiac images.

Chrisina Jayne, Andreas Lanitis, Chris Christodoulou

Learning of Spatio-temporal Dynamics in Thermal Engineering

Thermal engineering deals with the estimation of the temperature at different points and instants for a given set of boundary and initial conditions. For this, an analytic model replaces accurate but time-expensive numerical simulation models; it is independent of the boundary conditions and parameterized by the statistical learning of multidimensional temporal trajectories. This black-box model is a recursive neural network emulating the temperatures of interest over time from the only knowledge of initial conditions and exogenous variables.

The number of hidden neurons is selected by a non-asymptotic approach based upon the minimization of a penalyzed criterion. Methods like the slope heuristic and the dimension jump enable the calibration of the penalty constant in presence of a

n

-sample. In practice, their extrapolation to dependent data gives accurate results in the sense of the mean square error.

The surrogate model and the model selection are successfully applied to an industrial benchmark.

Matthias De Lozzo, Patricia Klotz, Béatrice Laurent

Neural Adaptive Control in Application Service Management Environment

This paper presents a method and a framework for adaptive control in Application Service Management environments. The controlled system is treated as a “black-box” by observing its operation during normal work or load conditions. Run-time metrics are collected and persisted creating a Knowledge Base of actual system states. Equipped with such knowledge we define system inputs, outputs and effectively select high/low Service Level Agreements values, and good/bad control actions from the past. On the basis of gained knowledge a training set is constructed, which determines the operation of a neural controller deployed in the application run-time. Control actions are executed in the background of the current system state, which is then again monitored and stored extending the states repository, giving views on the appropriateness of the control, which is frequently evaluated.

Tomasz Sikora, George D. Magoulas

Using Varying Negative Examples to Improve Computational Predictions of Transcription Factor Binding Sites

The identification of transcription factor binding sites (TFBSs ) is a non-trivial problem as the existing computational predictors produce a lot of false predictions. Though it is proven that combining these predictions with a meta-classifier, like Support Vector Machines (SVMs), can improve the overall results, this improvement is not as significant as expected. The reason for this is that the predictors are not reliable for the negative examples from non-binding sites in the promoter region. Therefore, using negative examples from different sources during training an SVM can be one of the solutions to this problem. In this study, we used different types of negative examples during training the classifier. These negative examples can be far away from the promoter regions or produced by randomisation or from the intronic region of genes. By using these negative examples during training, we observed their effect in improving predictions of TFBSs in the yeast. We also used a modified cross-validation method for this type of problem. Thus we observed substantial improvement in the classifier performance that could constitute a model for predicting TFBSs. Therefore, the major contribution of the analysis is that for the yeast genome, the position of binding sites could be predicted with high confidence using our technique and the predictions are of much higher quality than the predictions of the original prediction algorithms.

Faisal Rezwan, Yi Sun, Neil Davey, Rod Adams, Alistair G. Rust, Mark Robinson

Visual Analysis of a Cold Rolling Process Using Data-Based Modeling

In this paper, a method to characterize the chatter phenomenon in a cold rolling process is proposed. This approach is based on obtaining a global nonlinear dynamical MISO model, relating four input variables and the exit strip thickness as the output variable. In a second stage, local linear models are obtained for all working points using sensitivity analysis on the nonlinear model to get input/output small signal models. Each local model is characterized by a high dimensional vector containing the frequency response functions (FRF) of the four SISO resulting models. Finally, the FRF’s are projected on a 2D space, using the

t

-SNE algorithm, in order to visualize the dynamical changes of the process. Our results show a clear separation between chatter condition and other vibration states, allowing an early detection of chatter as well as being a visual analysis tool to study the chatter phenomenon.

Daniel Pérez, Francisco J. García-Fernández, Ignacio Díaz, Abel A. Cuadrado, Daniel G. Ordonez, Alberto B. Díez, Manuel Domínguez

Wind Power Forecasting to Minimize the Effects of Overproduction

Wind power generation increases very rapidly in the past few years. The available wind energy is random due to the intermittency and variability of the wind speed. This poses difficulty in the energy dispatched and cause costs, as the wind energy is not accurately scheduled in advance. This paper presents a short-term wind speed forecasting that uses a Kalman filter approach to predict the power production of wind farms. The prediction uses wind speed values measured over a year in a site, on the case study of Portugal. A method to group wind speeds by their similarity in clusters is developed together with a Kalman filter model that uses each cluster as an input to perform the wind power forecasting.

Fernando Ribeiro, Paulo Salgado, João Barreira

Using RISE Observer to Implement Patchy Neural Network for the Identification of “Wing Rock” Phenomenon on Slender Delta 80° Wings

In this paper, the “wing rock” phenomenon is described for slender delta 80° wing aircrafts on the roll axis. This phenomenon causes the aircraft to undergo a strong oscillatory movement with amplitude dependent on the angle of attack. The objective is to identify “wing rock” using the Patchy Neural Network (PNN), which is a new form of neural nets. For the update of the weights of the network, an observer called RISE (Robust Integral of Sign Error) and equations of algebraic form are used. This causes the PNN to be fast, efficient and of a low computational cost.

Paraskevas M. Chavatzopoulos, Thomas Giotis, Manolis Christodoulou, Haris Psillakis

Models Based on Neural Networks and Neuro-Fuzzy Systems for Wind Power Prediction Using Wavelet Transform as Data Preprocessing Method

Several studies have shown that the Brazilian wind potential can contribute significantly to the electricity supply, especially in the Northeast, where winds present an important feature of being complementary in relation to the flows of the San Francisco River. This work proposes and develops models to forecast hourly average wind speeds and wind power generation based on Artificial Neural Networks, Fuzzy Logic and Wavelets. The models were adjusted for forecasting with variable steps up to twenty-four hours ahead. The gain of some of the developed models in relation to the reference model was of approximately 80% for forecasts in a period of one hour ahead. The results showed that a wavelet analysis combined with artificial intelligence tools provides more reliable forecasts than those obtained with the reference models, especially for forecasts in a period of 1 to 6 hours ahead.

Ronaldo R. B. de Aquino, Hugo T. V. Gouveia, Milde M. S. Lira, Aida A. Ferreira, Otoni Nobrega Neto, Manoel A. Carvalho

Neural Networks for Air Data Estimation: Test of Neural Network Simulating Real Flight Instruments

In this paper virtual air data sensors have been modeled using neural networks in order to estimate the aircraft angles of attack and sideslip. These virtual sensors have been designed and tested using the aircraft mathematical model of the De Havilland DHC-2. The aim of the work is to evaluate the degradation of neural network performance, which is supposed to occur when real flight instruments are used instead of simulated ones. The external environment has been simulated, and special attention has been devoted to electronic noise that affects each input signals examining modern devices.. Neural networks, trained with noise free signals, demonstrate satisfactory agreement between theoretical and estimated angles of attack and sideslip.

Manuela Battipede, Piero Gili, Angelo Lerro

Direct Zero-Norm Minimization for Neural Network Pruning and Training

Designing a feed-forward neural network with optimal topology in terms of complexity (hidden layer nodes and connections between nodes) and training performance has been a matter of considerable concern since the very beginning of neural networks research. Typically, this issue is dealt with by pruning a fully interconnected network with “many” nodes in the hidden layers, eliminating “superfluous” connections and nodes. However the problem has not been solved yet and it seems to be even more relevant today in the context of deep learning networks. In this paper we present a method of direct zero-norm minimization for pruning while training a Multi Layer Perceptron. The method employs a cooperative scheme using two swarms of particles and its purpose is to minimize an aggregate function corresponding to the total risk functional. Our discussion highlights relevant computational and methodological issues of the approach that are not apparent and well defined in the literature.

S. P. Adam, George D. Magoulas, M. N. Vrahatis

3D Vision-Based Autonomous Navigation System Using ANN and Kinect Sensor

In this paper, we present an autonomous navigation system based on a finite state machine (FSM) learned by an artificial neural network (ANN) in an indoor navigation task. This system uses a kinect as the only sensor. In the first step, the ANN is trained to recognize the different specific environment configurations, identifying the different environment situations (states) based on the kinect detections. Then, a specific sequence of states and actions is generated for any route defined by the user, configuring a path in a topological like map. So, the robot becomes able to autonomously navigate through this environment, reaching the destination after going through a sequence of specific environment places, each place being identified by its local properties, as for example, straight path, path turning to left, path turning to right, bifurcations and path intersections. The experiments were performed with a Pioneer P3-AT robot equipped with a kinect sensor in order to validate and evaluate this approach. The proposed method demonstrated to be a promising approach to autonomous mobile robots navigation.

Daniel Sales, Diogo Correa, Fernando S. Osório, Denis F. Wolf

Hybrid Computational Model for Producing English Past Tense Verbs

In this work, we explore the use of artificial neural networks (ANN) as computational models for producing English past tense verbs by combining them with the genetic algorithms (GA). The principal focus was to model the population variability exhibited by children in learning the past tense. This variability stems from genetic and environmental origins.We simulated the effects of genetic influences via variations in the neuro computational parameters of the ANNs, and the effects of environmental influences via a filter applied to the training set, implementing variation in the information available to the child produced by, for example, differences in socio-economic status. In the model, GA served two main purposes - to create the population of artificial neural networks and to encode the neuro computational parameters of the ANN into the genome. English past tense provides an interesting training domain in that it comprises a set of quasi-regular mappings. English verbs are of two types, regular verbs and irregular verbs. However, a similarity gradient also exists between these two classes. We consider the performance of the combination of ANN and GA under a range of metrics. Our tests produced encouraging results as to the utility of this method, and a foundation for future work in using a computational framework to capture population-level variability.

Maitrei Kohli, George D. Magoulas, Michael Thomas

Characterizing Mobile Network Daily Traffic Patterns by 1-Dimensional SOM and Clustering

Mobile network traffic produces daily patterns. In this paper we show how exploratory data analysis can be used to inspect the origin of the daily patterns. We use a 1-dimensional self-organizing map to characterize the patterns. 1-dimensional map enables compact visualization that is especially suitable for data where the variables are not independent but form a pattern. We introduce a stability index for analyzing the variation of the daily patterns of network elements along the days of the week. We use clustering to construct profiles for the network elements to study the stability of the traffic patterns within each element. We found out that the day of the week is the main explanation for the traffic patterns on weekends. On weekdays the traffic patterns are mostly specific to groups of networks elements, not the day of the week.

Pekka Kumpulainen, Kimmo Hätönen

Dipolar Designing Layers of Formal Neurons

A layer of formal neurons can perform separable data aggregation. The term (separable data aggregation( means that a number of input vectors belonging to one category (class) are merged by the layer in one output vector with an additional condition that input vectors belonging to different categories are not aggregated. Dipolar principles of separable layers designing are examined in the paper. Hierarchical networks can be designed from separable layers and used for aggregation of all input vectors belonging to one category in an output vector.

Leon Bobrowski

Evaluating the Impact of Categorical Data Encoding and Scaling on Neural Network Classification Performance: The Case of Repeat Consumption of Identical Cultural Goods

This article investigated the impact of categorical input encoding and scaling approaches on neural network sensitivity and overall classification performance in the context of predicting the repeat viewing propensity of movie goers. The results show that neural network out of sample minimum sensitivity and overall classification performance are indifferent to the scaling of the categorical inputs. However, the encoding of inputs had a significant impact on classification accuracy and utilising ordinal or thermometer encoding approaches for categorical inputs significantly increases the out of sample accuracy of the neural network classifier. These findings confirm that the impact of categorical encoding is problem specific for an ordinal approach, and support thermometer encoding as most suitable for categorical inputs. The classification performance of neural networks was compared against a logistic regression model and the results show that in this instance, the non-parametric approach does not offer any advantage over standard statistical models.

Elena Fitkov-Norris, Samireh Vahid, Chris Hand

A Hybrid Neural Emotion Recogniser for Human-Robotic Agent Interaction

This paper presents a hybrid neural approach to emotion recognition from speech, which combines feature selection using principal component analysis (PCA) with unsupervised neural clustering through self-organising map (SOM). Given the importance that is associated with emotions in humans, it is unlikely that robots will be accepted as anything more that machines if they do not express and recognise emotions. In this paper, we describe the performance of an unsupervised approach to emotion recognition that achieves similar performance to current supervised intelligent approaches. Performance, however, reduces when the system is tested using samples from a male volunteer not in the training set using a low cost microphone. Through the use of an unsupervised neural approach, it is possible to go beyond the basic binary classification of emotions to consider the similarity between emotions and whether speech can express multiple emotions at the same time.

Alexandru Traista, Mark Elshaw

Ambient Intelligent Monitoring of Dementia Suffers Using Unsupervised Neural Networks and Weighted Rule Based Summarisation

This paper investigates the development of a system for monitoring of dementia suffers living in their own homes. The system uses unobtrusive pervasive sensor and actuator devices that can be deployed within a patient’s home grouped and accessed via standardized platforms. For each sensor group our system uses unsupervised neural networks to identify the patient’s habitual behaviours based on their activities in the environment. Rule-based summarisation is used to provide descriptive rules representing the intra and inter activity variations within the discovered behaviours. We propose a model comparison mechanism to facilitate tracking of behaviour changes, which could be due to the effects of cognitive decline. We demonstrate using user data acquired from a real pervasive computing environment, how our system is able to identify the user’s prominent behaviours enabling assessment and future tracking.

Faiyaz Doctor, Chrisina Jayne, Rahat Iqbal

On the Intelligent Machine Learning in Three Dimensional Space and Applications

The engineering applications of high dimensional neural network are becoming very popular in almost every intelligence system design. Just to name few, computer vision, robotics, biometric identification, control, communication system and forecasting are some of the scientific fields that take advantage of artificial neural networks (ANN) to emulate intelligent behavior. In computer vision the interpretation of 3D motion, 3D transformations and 3D face or object recognition are important tasks. There have been many methodologies to solve them but these methods are time consuming and weak to noise. The advantage of using neural networks for object recognition is the feasibility of a training system to capture the complex class conditional density of patterns. It will be desirable to explore the capabilities of ANN that can directly process three dimensional information. This article discusses the machine learning from the view points of 3D vector-valued neural network and corresponding applications. The learning and generalization capacity of high dimensional ANN is confirmed through diverse simulation examples.

Bipin K. Tripathi, Prem K. Kalra

Knowledge Clustering Using a Neural Network in a Course on Medical-Surgical Nursing

This paper presents a neural network-based intelligent data analysis for knowledge clustering in an undergraduate nursing course. A MCQ (Multiple Choice Question) test was performed to evaluate medical-surgical nursing knowledge in a second-year course. A total of 23 pattern groups were created from the answers of 208 students. Data collected were used to provide customized feedback which guide students towards a greater understanding of particular concepts. The pattern groupings can be integrated with an on-line (MCQ) system for training purposes.

José Luis Fernández-Alemán, Chrisina Jayne, Ana Belén Sánchez García, Juan M. Carrillo-de-Gea, Ambrosio Toval Alvarez

A New Approach in Stability Analysis of Hopfield-Type Neural Networks: Almost Stability

In this paper,we presented a new stability concept for neural networks: almost stability. The necessary and sufficient conditions of almost stability of the Hopfield-type neural networks were proposed. Examples were also given to our conditions.

Kaiming Wang

A Double Layer Dementia Diagnosis System Using Machine Learning Techniques

Studies show that dementia is highly age-associated. Early diagnosis can help patient to receive timely treatment and slow down the deterioration. This paper proposed a hierarchical double layer structure with multi-machine learning algorithms for early stage dementia diagnosis. Fuzzy cognitive map (FCM) and probability neural networks (PNNs) were adopted to give initial diagnosis at based-layer, and then Bayesian networks (BNs) was used to make a final diagnosis at top-layer. Diagnosis results, “proposed treatment” and “no treatment required” can be used to provide self-testing or secondary dementia diagnosis to medical institutions. To demonstrate the reliability of the proposed system, a clinical data provided by the Cheng Kung University Hospital was examined. The accuracy of this system was as high as 83%, which showed that the proposed system was reliable and flexible.

Po-Chuan Cho, Wen-Hui Chen

Workshop on Applying Computational Intelligence Techniques in Financial Time Series Forecasting and Trading

A Hybrid Radial Basis Function and Particle Swarm Optimization Neural Network Approach in Forecasting the EUR/GBP Exchange Rates Returns

The motivation for this paper is to introduce in Finance a hybrid Neural Network architecture of Adaptive Particle Swarm Optimization and Radial Basis Function (ARBF-PSO) and a Neural Network fitness function for financial forecasting purposes. This is done by benchmarking the ARBF-PSO results with those of three different Neural Networks architectures and three statistical/technical models. As it turns out, the ARBF-PSO architecture outperforms all other models in terms of statistical accuracy and trading efficiency in the examined forecasting task.

Georgios Sermpinis, Konstantinos Theofilatos, Andreas Karathanasopoulos, Efstratios Georgopoulos, Christian Dunis

Forecasting and Trading the High-Low Range of Stocks and ETFs with Neural Networks

Intraday trading has some appealing characteristics. For example, overnight gap risks are greatly reduced. Intraday trading strategies tend to achieve better risk adjusted returns. However, academic literature on intraday trading strategies is relatively scarce compared to a significant amount of literature based on daily closing data. This may be partly related to the increased difficulty of dealing with intraday data. In the present paper we expand on a novel approach that builds an intraday trading strategy on open-high-low-close (OHLC) data. OHLC data is easily available from most database vendors. We use OHLC data to train neural networks that forecast the day’s high and low of liquid US stocks and ETFs. The resulting long-short strategy tries to take advantage of the daily trading range of a security and exits all positions at the close. A volatility filter further improves risk-adjusted returns.

Hans-Jörg von Mettenheim, Michael H. Breitner

Kalman Filters and Neural Networks in Forecasting and Trading

The motivation of this paper is to investigate the use of a Neural Network (NN) architecture, the Psi Sigma Neural Network, when applied to the task of forecasting and trading the Euro/Dollar exchange rate and to explore the utility of Kalman Filters in combining NN forecasts. This is done by benchmarking the statistical and trading performance of PSN with a Naive Strategy and two different NN architectures, a Multi-Layer Perceptron and a Recurrent Network. We combine our NN forecasts with Kalman Filter, a traditional Simple Average and the Granger- Ramanathan’s Regression Approach. The statistical and trading performance of our models is estimated throughout the period of 2002-2010, using the last two years for out-of-sample testing. The PSN outperforms all models’ individual performances in terms of statistical accuracy and trading performance. The forecast combinations also present improved empirical evidence, with Kalman Filters outperforming by far its benchmarks.

Georgios Sermpinis, Christian Dunis, Jason Laws, Charalampos Stasinakis

Short-Term Trading Performance of Spot Freight Rates and Derivatives in the Tanker Shipping Market: Do Neural Networks Provide Suitable Results?

In this paper we investigate the forecasting and trading performance of linear and non-linear methods, in order to generate short-term forecasts in the dirty tanker shipping market. We attempt to uncover the benefits of using several time series models and the potential of neural networks. Maritime forecasting studies using neural networks are rare and only focus on spot rates. We build on this kind of investigation, but we extend our study on freight rates derivatives or Forward Freight Agreements (FFA) in a simple trading simulation. Our conclusion is, that non-linear methods like neural networks are suitable for short-term forecasting and trading freight rates, as their results match or improve on those of other models. Nevertheless, we think that further research with freight rates and corresponding derivatives is developable for decision and trading applications with enhanced forecasting models.

Christian von Spreckelsen, Hans-Jörg von Mettenheim, Michael H. Breitner

Modelling and Trading the DJIA Financial Index Using Neural Networks Optimized with Adaptive Evolutionary Algorithms

In the current paper we study an evolutionary framework for the optimization of various types Neural Networks’ structure and parameters. Two different adaptive evolutionary algorithms, named as adaptive Genetic Algorithms (aGA) and adaptive Differential Evolution (aDE), were developed to optimize the structure and the parameters of two different types of Neural Networks: Multilayer Perceptron (MLPs) and Wavelet Neural Networks (WNN). Wavelet neural networks have been introduced as an alternative to MLPs to overcome their shortcomings presenting more compact architecture and higher learning speed. Furthermore, the evolutionary algorithms, which were implemented, are both adaptive in terms that their most important parameters (Mutation and Crossover probabilities) are assigned using a self adaptive scheme. The motivation of this paper is to uncover novel hybrid methodologies for the task of forecasting and trading DJIE financial index. This is done by benchmarking the forecasting performance the four proposed hybrid methodologies (aGA-MLP, aGA-WNN, aDE-MLP and aDE-WNN) with some traditional techniques, either statistical such as a an autoregressive moving average model (ARMA), or technical such as a moving average covcergence/divergence model (MACD). The trading performance of all models is investigated in a forecast and trading simulation on our time series over the period 1997-2012. As it turns out, the aDE-WNN hybrid methodology does remarkably well and outperforms all other models in simple trading simulation exercises. (This paper is submitted for the ACIFF workshop).

Konstantinos Theofilatos, Andreas Karathanasopoulos, Georgios Sermpinis, Thomas Amorgianiotis, Efstratios Georgopoulos, Spiros Likothanassis

Workshop on Computational Intelligence Applications in Bioinformatic

Applying Kernel Methods on Protein Complexes Detection Problem

During the last years, various methodologies have made possible the detection of large parts of the protein interaction network of various organisms. However, these networks are containing highly noisy data, degrading the quality of information they carry. Various weighting schemes have been applied in order to eliminate noise from interaction data and help bioinformaticians to extract valuable information such as the detection of protein complexes. In this contribution, we propose the addition of an extra step on these weighting schemes by using kernel methods to better assess the reliability of each pairwise interaction. Our experimental results prove that kernel methods clearly help the elimination of noise by producing improved results on the protein complexes detection problem.

Charalampos Moschopoulos, Griet Laenen, George Kritikos, Yves Moreau

Efficient Computational Prediction and Scoring of Human Protein-Protein Interactions Using a Novel Gene Expression Programming Methodology

Proteins and their interactions have been proven to play a central role in many cellular processes. Thus, many experimental methods have been developed for their prediction. These experimental methods are uneconomic and time consuming in the case of low throughput methods or inaccurate in the case of high throughput methods. To overcome these limitations, many computational methods have been developed to predict and score Protein-Protein Interactions (PPIs) using a variety of functional, sequential and structural data for each protein pair. Existing computational methods can still be enhanced in terms of classification performance and interpretability. In the present paper we present a novel Gene Expression Programming (GEP) algorithm, named as jGEPModelling 2.0, and apply it to the problem of PPI prediction and scoring. jGEPModelling2.0 is a variation of the classic GEP algorithm to make it suitable for the problem of PPI prediction and enhance its classification performance. To test its efficiency, we applied it to a public available dataset and compared it to two other state-of-the-art PPI prediction models. Experimental results proved that jGEPModelling2.0 outperformed existing methodologies in terms of classification performance and interpretability. (This paper is submitted for the CIAB2012 workshop).

Konstantinos Theofilatos, Christos Dimitrakopoulos, Maria Antoniou, Efstratios Georgopoulos, Stergios Papadimitriou, Spiros Likothanassis, Seferina Mavroudi

Biomedical Color Image Segmentation through Precise Seed Selection in Fuzzy Clustering

Biomedical color images play major role in medical diagnosis. Often a change of state is identified through minute variations in color at tiny parts. Fuzzy C-means (FCM) clustering is suitable for pixel classification to isolate those parts but its success is heavily dependent on the selection of seed clusters. This paper presents a simple but effective technique to generate seed clusters resembling the image features. The HSI color model is selected for near-zero correlation among components. The approach has been tested on several cell images having low contrast at adjacent parts. Results of segmentation show its effectiveness.

Byomkesh Mandal, Balaram Bhattacharyya

Backmatter

Weitere Informationen

Premium Partner

Neuer Inhalt

BranchenIndex Online

Die B2B-Firmensuche für Industrie und Wirtschaft: Kostenfrei in Firmenprofilen nach Lieferanten, Herstellern, Dienstleistern und Händlern recherchieren.

Whitepaper

- ANZEIGE -

Best Practices für die Mitarbeiter-Partizipation in der Produktentwicklung

Unternehmen haben das Innovationspotenzial der eigenen Mitarbeiter auch außerhalb der F&E-Abteilung erkannt. Viele Initiativen zur Partizipation scheitern in der Praxis jedoch häufig. Lesen Sie hier  - basierend auf einer qualitativ-explorativen Expertenstudie - mehr über die wesentlichen Problemfelder der mitarbeiterzentrierten Produktentwicklung und profitieren Sie von konkreten Handlungsempfehlungen aus der Praxis.
Jetzt gratis downloaden!

Bildnachweise