Advances in Computational Intelligence

The paper presents comparison of few chosen approaches to recognition of autism on the basis of gene expression microarray. The important point in this task is selection of genes of the highest class discriminative ability. To solve the problem we have applied many selection methods, which are based on different principles. The limited set of genes in each method are selected for further analysis. In this paper we will compare the genetic algorithm and random forest in the role of final gene selection. The most important genes selected by each method are used as the input attributes to the support vector machine and random forest classifiers, cooperating in an ensemble. The final result of classification is generated by the random forest, performing the role of fusion system for an ensemble.

Feature selection is often required as a preliminary step for many machine learning problems. However, most of the existing methods only work in a centralized fashion, i.e. using the whole dataset at once. In this paper we propose a new methodology for distributing the feature selection process by samples which maintains the class distribution. Subsequently, it performs a merging procedure which updates the final feature subset according to the theoretical complexity of these features, by using data complexity measures. In this way, we provide a framework for distributed feature selection independent of the classifier and that can be used with any feature selection algorithm. The effectiveness of our proposal is tested on six representative datasets. The experimental results show that the execution time is considerably shortened whereas the performance is maintained compared to a previous distributed approach and the standard algorithms applied to the non-partitioned datasets.

In the last few years, ensemble learning has been the focus of much attention mainly in classification tasks, based on the assumption that combining the output of multiple experts is better than the output of any single expert. This idea of ensemble learning can be adapted for feature selection, in which different feature selection algorithms act as different experts. In this paper we propose an ensemble for feature selection based on combining rankings of features, trying to overcome the problem of selecting an appropriate ranker method for each problem at hand. The results of the individual rankings are combined with SVM Rank, and the adequacy of the ensemble was subsequently tested using SVM as classifier. Results on five UCI datasets showed that the use of the proposed ensemble gives better or comparable performance than the feature selection methods individually.

The combination of visual and textual information in a CBR system is a promising concept to overcome the limitations of existing medical CBR systems, which are mainly focused on evaluation of new and existing cases with data mining, clustering techniques or statistical analysis of patient’s health condition parameters. The advantage of our proposed medical CBR system, called DePicT, is the knowledge based recommendation, which utilizes case-based reasoning through analyzing image and text from patient health records. DePicT can find a solution regarding patient’s problem description even with partly missing information. It uses image interpretation parameters and profiles of word associations in the feature selection and case matching process to find similar cases for recommendation.

In this paper we evaluate the use of the machine learning algorithms Support Vector Machines (SVM), K-Nearest Neighbors (KNN), Classification and Regression Trees (CART) and Naive Bayes (NB) to identify non spontaneous saccades in clinical electrooculography tests. Our approach tries to solve problems like the use of manually established thresholds present in classical methods like identification by velocity threshold (I-VT) or identification by dispersion threshold (I-DT). We propose a modification to an adaptive threshold estimation algorithm for detecting signal impulses without the need of any user input. Also, a set of features were selected to take advantage of intrinsic characteristics of clinical electrooculography tests. The models were evaluated with signals recorded to subjects affected by Spinocerebellar Ataxia type 2 (SCA2). Results obtained by the algorithm show accuracies over 97%, recalls over 97% and precisions over 91% for the four models evaluated.

Time-series segmentation can be approached by combining a clustering technique and genetic algorithm (GA) with the purpose of automatically finding segments and patterns of a time series. This is an interesting data mining field, but its application to the optimal segmentation of financial time series is a very challenging task, so accurate algorithms are needed. In this sense, GAs are relatively poor at finding the precise optimum solution in the region where the algorithm converges. Thus, this work presents a hybrid GA algorithm including a local search method, aimed to improve the quality of the final solution. The local search algorithm is based on maximizing a likelihood ratio, assuming normality for the series and the subseries in which the original one is segmented. A real-world time series in the Spanish Stock Market field was used to test this methodology.

This paper proposes a nonlinear ordinal logistic regression method based on the hybridization of a linear model and radial basis function (RBF) neural network models for ordinal regression. The process for obtaining the coefficients is carried out in several steps. In the first step we use an evolutionary algorithm to determine the structure of the RBF neural network model, in a second step we transform the initial feature space (covariate space) adding the nonlinear transformations of the input variables given by the RBFs of the best individual in the final generation of the evolutionary algorithm. Finally, we apply an ordinal logistic regression in the new feature space. This methodology is tested using 8 benchmark problems from the UCI repository. The hybrid model outperforms both the linear and the nonlinear part obtaining a good compromise between them and better results in terms of accuracy and ordinal classification error.

This paper tackles marine energy prediction from the classification point of view, by previously discretising the real objective variable into a set of consecutive categories or ranges. Given that the range of energy flux is enough to obtain an approximation of the amount of energy produced, the purpose of this discretisation is to simplify the prediction task. A special kind of autoregressive models are considered, where the category to be predicted depends on both the previous values of energy flux and a set of meteorological variables estimated by numerical models. Apart from this, this paper introduces two different ways of adjusting the order of the autoregressive models, one based on nested cross-validation and the other one based on a dynamic window. The results show that these kind of models are able to predict the time series in an acceptable way, and that the dynamic window procedure leads to the best accuracy without needing the additional computational cost of adjusting the order of the model.

This research aims to advance blinking detection in the context of work activity. Rather than patients having to attend a clinic, blinking videos can be acquired in a work environment, and further automatically analyzed. Therefore, this paper presents a methodology to perform the automatic detection of eye blink using consumer videos acquired with low-cost web cameras. This methodology includes the detection of the face and eyes of the recorded person, and then it analyzes the low-level features of the eye region to create a quantitative vector. Finally, this vector is classified into one of the two categories considered —open and closed eyes— by using machine learning algorithms. The effectiveness of the proposed methodology was demonstrated since it provides unbiased results with classification errors under 5%.

Convolutional Neural Networks have systematically shown good performance in Computer Vision and in Handwritten Text Recognition tasks. This paper proposes the use of these models for document image binarization. The main idea is to classify each pixel of the image into foreground and background from a sliding window centered at the pixel to be classified. An experimental analysis on the effect of sensitive parameters and some working topologies are proposed using two different corpora, of very different properties: DIBCO and Santgall.

Least Square Support Vector Machines (LSSVMs) are an alternative to SVMs because the training process of LSSVM classifiers only requires to solve a linear equation system instead of solving a quadratic programming optimization problem. Nevertheless, the absence of sparseness in the solution (i.e. the Lagrange multipliers vector) obtained is a significant drawback which must be overcome. This work presents a new approach to building Sparse Least Square Support Vector Machines with fixed-size of support vectors for classification tasks. Our proposal named FSGAS-LSSVM relies on a binary-encoding single-objective genetic algorithms, in which the standard reproduction and mutation operators must be modified. The main idea is to leave a few support vectors out of the solution without affecting the classifier’s accuracy and even improving it. In our proposal, GAs are used to select a suitable fixed-size set of support vectors by removing non-relevant patterns or those ones, which can be corrupted with noise and thus prevent classifiers to achieve higher accuracies.

The use of ensemble methods for pattern classification have gained attention in recent years mainly due to its improvements on classification rates. This paper evaluates ensemble learning methods using the Minimal Learning Machines (MLM), a recently proposed supervised learning algorithm. Additionally, we introduce an alternative output estimation procedure to reduce the complexity of the standard MLM. The proposed methods are evaluated on real datasets and compared to several state-of-the-art classification algorithms.

This paper presents an extension of the well-known Extreme Learning Machines (ELMs). The main goal is to provide probabilities as outputs for Multiclass Classification problems. Such information is more useful in practice than traditional crisp classification outputs. In summary, Gaussian Mixture Models are used as post-processing of ELMs. In that context, the proposed global methodology is keeping the advantages of ELMs (low computational time and state of the art performances) and the ability of Gaussian Mixture Models to deal with probabilities. The methodology is tested on 3 toy examples and 3 real datasets. As a result, the global performances of ELMs are slightly improved and the probability outputs are seen to be accurate and useful in practice.

As machine learning becomes more popular in all fields, its use is well known in finance and economics. The growing number of people using models to predict the market’s behaviour can modify the market itself so it is more predictable. In this context, the key element is to find out which variables are used to build the model in a macroeconomic environment. This paper presents an application of kernel methods to predict the EUR/USD relationship performing variable selection. The results show how after applying a proper variable selection, very accurate predictions can be achieved and smaller historical data is needed to train the model.

The retina continually adapts its kinetics, average response and sensitivity to the conditions of the environment. Retinal neurons adapt essentially to the mean light intensity and its temporal fluctuations over the mean, also called temporal contrast. Contrast adaptation has two distinct temporal expressions with fast and slow components. Here, we present a configurable retina simulation environment that accurately reproduces both contrast components. A contrast increase in the visual input accelerates kinetics of the filter, reduces sensitivity and depolarizes the membrane potential. Slow adaptation does not affect the temporal response but produces a progressive hyperpolarization of membrane potential. The implemented model for contrast adaptation provides a neural basis of each retinal stage, from photoreceptors up to ganglion cells, to explain the observed retina behavior. Both forms of contrast adaptation, fast and slow, are captured by a combined model of shunting feedback of bipolar cells and short-term plasticity (STP) at the bipolar-to-ganglion synapse. Biological accuracy of the model is evaluated by comparison of the measured neural response with the simulated response fitted to published physiological data. One problem with the simulated model is finding its optimal parameter settings, since the model response is described by a complex system of different retina stages with linear, nonlinear and feedback connections. We propose to use a multiobjective genetic optimization to automatically search the parameter space and easily find a feasible configuration solution.

Hardware implementations of Artificial Neural Networks (ANNs) allow to exploit the inherent parallelism of these architectures. Nevertheless, ANN hardware implementation requires a large amount of hardware resources. Recently, Reservoir computing (RC) has arisen as an advantageous technique to implement Recurrent Neural Networks RNNs). In this work, we present an efficient approach to implement RC systems. The proposed methodology employs probabilistic logic to reduce the hardware area required to implement the arithmetic operations present in neural networks and conventional binary logic for the nonlinear activation function. We show the functionality and low hardware resources used by the proposed methodology.

Recent advances in FPGA technology have permitted the implementation of neurocomputational models, making them an interesting alternative to standard PCs in order to speed up the computations involved taking advantage of the intrinsic FPGA parallelism. In this work, we analyse and compare the FPGA implementation of two neural network learning algorithms: the standard Back-Propagation algorithm and C-Mantec, a constructive neural network algorithm that generates compact one hidden layer architectures. One of the main differences between both algorithms is the fact that while Back-Propagation needs a predefined architecture, C-Mantec constructs its network while learning the input patterns. Several aspects of the FPGA implementation of both algorithms are analysed, focusing in features like logic and memory resources needed, transfer function implementation, computation time, etc. Advantages and disadvantages of both methods are discussed in the context of their application to benchmark problems.

About 90% of breast cancers do not cause or are capable of producing death if detected at an early stage and treated properly. Indeed, it is still not known a specific cause for the illness. It may be not only a beginning, but also a set of associations that will determine the onset of the disease. Undeniably, there are some factors that seem to be associated with the boosted risk of the malady. Pondering the present study, different breast cancer risk assessment models where considered. It is our intention to develop a hybrid decision support system under a formal framework based on Logic Programming for knowledge representation and reasoning, complemented with an approach to computing centered on Artificial Neural Networks, to evaluate the risk of developing breast cancer and the respective Degree-of-Confidence that one has on such a happening.

Transmission line protective relaying is an essential feature of a reliable power system operation. Fast detecting, isolating, locating and repairing of the different faults are critical in maintaining a reliable power system operation. On the other hand, classification of the different fault types plays very significant role in digital distance protection of the transmission line. Accurate and fast fault classification can prevent from more damages in the power system. In this paper, an approach is presented to classify the fault in a double-circuit transmission line based on the adaptive Neuro- Fuzzy Inference System (ANFIS) using three phase current samples of only one terminal. This method is independent of effects of variation of fault inception angle, fault location, fault resistance and load angle. MATLAB/Simulink is used to produce fault signals. The proposed method is tested by simulating different scenarios on a given transmission line model. The simulation results denote that the proposed approach for fault identification is able to classify all the faults on the parallel transmission line within half cycle after the inception of fault.

A modelling procedure based on hybrid configuration composed of artificial neural networks, differential evolution and clonal selection algorithms is developed and applied in this work. The neural network represents the model of the system, while the differential evolution and clonal selection algorithms perform a simultaneous topological and parametric optimization of the model. The results indicated that the combination of the two optimizers produces better results compared with each of them working separately. As case study, styrene polymerization, a complex process which is difficult to model when taking into consideration all the internal interactions, was chosen. Neural networks, designed in an optimal form, proved to be adequate tools for modelling this system.

This paper addresses classification of 3D point cloud data from natural environments based on voxels. The proposed model uses multi-layer perceptrons to classify voxels based on a statistic geometric analysis of the spatial distribution of inner points. Geometric features such as tubular structures or flat surfaces are identified regardless of their orientation, which is useful for unstructured or natural environments. Furthermore, the combination of voxels and neural networks pursues faster computation than alternative strategies. The model has been successfully tested with 3D laser scans from natural environments.

This article presents a new framework for fitting measured scientific data to a simple empirical formula by introducing an additional linear neuron to the standard Gaussian kernel radial basis function (RBF) neural networks. The proposed method is first used to evaluate two benchmark datasets (Preschool boy and titanium heat) and then is applied to fit a set of stopping power data (MeV energetic carbon projectiles in elemental target materials C, Al, Si, Ti, Ni, Cu, Ag and Au) from high energy physics experiments. Without increasing computational complexity, the proposed approach significantly improves accuracy of fitting. Based on this type RBF neural network, a simple 6-parameter empirical formula is developed for various potential applications in curve fitting and nonlinear regression problems.

Rule extraction from neural networks represents a difficult research problem, which is NP-hard. In this work we show how a special Multi Layer Perceptron architecture denoted as DIMLP can be used to extract rules from ensembles of DIMLPs and Quantized Support Vector Machines (QSVMs). The key idea for rule extraction is that the locations of discriminative hyperplanes are known, precisely. Based on ten repetitions of stratified 10-fold cross validation trials and with the use of default learning parameters we generated symbolic rules from five datasets. The obtained results compared favorably with respect to another state of the art technique applied to Support Vector Machines.

The recognition of human activity has been extensively investigated in the last decades. Typically, wearable sensors are used to register body motion signals that are analyzed by following a set of signal processing and machine learning steps to recognize the activity performed by the user. One of the most important steps refers to the signal segmentation, which is mainly performed through windowing approaches. In fact, it has been proved that the choice of window size directly conditions the performance of the recognition system. Thus, instead of limiting to a specific window configuration, this work proposes the use of multiple recognition systems operating on multiple window sizes. The suggested model employs a weighted decision fusion mechanism to fairly leverage the potential yielded by each recognition system based on the target activity set. This novel technique is benchmarked on a well-known activity recognition dataset. The obtained results show a significant improvement in terms of performance with respect to common systems operating on a single window size.

Wearable activity recognition has attracted very much attention in the recent years. Although many contributions have been provided so far, most solutions are developed to operate on predefined settings and fixed sensor setups. Real-world activity recognition applications and users demand more flexible sensor configurations, which may deal with potential adverse situations such as defective or missing sensors. A novel method to intelligently select the best replacement for an anomalous or nonrecoverable sensor is presented in this work. The proposed method builds on an ontology defined to neatly describe wearable sensors and their main properties, such as measured magnitude, location and internal characteristics. SPARQL queries are used to retrieve the ontological sensor descriptions for the selection of the best sensor replacement. The on-body location proximity of the sensors is considered during the sensor search process to determine the most adequate alternative.

This work presents and compare six short-term forecasting methods for hourly aggregated solar generation. The methods forecast one day ahead hourly values of Spanish solar generation. Three of the models are based on MLP network and the other three are based on NARX. The two different types of NN use to forecast the same NWP data, comprising solar radiation, solar irradiation and the cloudiness index weighted with the installed solar power for the whole country. In addition of the NWP data the models are fed with the aggregated solar energy generation in hourly step given by the System Operator.

The results of the two types of NN are compared and discussed in the conclusions as much as the error variability along the day hours. The results obtained by the six methods are evaluated, concluding that the most accurate result is the one given by the developed NARX irradiance forecast method; achieving the lowest one day-ahead Mean Average Daily Error of 16.64%.

Public speaking is a non-trivial task since it is affected by how nonverbal behaviors are expressed. Practicing to deliver the appropriate expressions is difficult while they are mostly given subconsciously. This paper presents our empirical study on the nonverbal behaviors of presenters. Such information was used as the ground truth to develop an intelligent tutoring system. The system can capture bodily characteristics of presenters via a depth camera, interpret this information in order to assess the quality of the presentation, and then give feedbacks to users. Feedbacks are delivered immediately through a virtual conference room, in which the reactions of the simulated avatars can be controlled based on the performance of presenters.

The rapid growth social networks have led many companies to use mobile payment systems as business sales tools. As these platforms have an increasing acceptance among the consumers, the main goal of this research is to analyze the individuals’ use intention of these systems in a social network environment. The problem of variable selection arises in this context as key to understand user’s behaviour. This paper compares several non-parametric criteria to perform variable selection and combines them in a multiobjective manner showing a good performance in the experiments carried out and validated by experts.

Aggregating the preference of multiple experts is a very old problem which remains without an absolute solution. This assertion is supported by the Arrow’s theorem: there is no aggregation method that simultaneously satisfies three fairness criteria (non-dictatorship, independence of irrelevant alternatives and Pareto efficiency). However, it is possible to find a solution having minimal distance to the consensus, although it involves a NP-hard problem even for only a few experts. This paper presents a model based on Ant Colony Optimization for facing this problem when input data are incomplete. It means that our model should build a complete ordering from partial rankings. Besides, we introduce a measure to determine the distance between items. It provides a more complete picture of the aggregated solution. In order to illustrate our contributions we use a real problem concerning Employer Branding issues in Belgium.

In this research paper, a complex valued generalization of associative memory synthesized by Hopfield is considered and it is proved that it is impossible to synthesize such a neural network with desired unitary stable states when the dimension of the network (number of neurons) is odd. The linear algebraic structure of such a neural network is discussed. Using Sylvester construction of Hadamard matrix of suitable dimension, an algorithm to synthesize such a complex Hopfield neural network is discussed. Also, it is discussed how to synthesize real / complex valued associative memories with desired energy landscape (i.e. desired stable states and desired energy values of associated quadratic energy function).

The incremental learning is a method to compose an associate memory using a chaotic neural network and provides larger capacity than correlative learning in compensation for a large amount of computation. A chaotic neuron has spatio-temporal sum in it and the temporal sum makes the learning stable to input noise. When there is no noise in input, the neuron may not need temporal sum. In this paper, to reduce the computations, a simplified network without temporal sum are introduced and investigated through the computer simulations comparing with the network as in the past. It turns out that the simplified network is able to learn input patterns quickly with the learning parameter varying.

Many probabilistic models are frequently used for natural growth-patterns modelling and their forecasting such as the diffusion processes. The maximum likelihood estimation of the parameters of a diffusion process requires a system of equations that, for some cases, has no explicit solution to be solved. Facing that situation, we can approximate the solution using an optimization method. In this paper we compare five optimization methods: an Iterative Method, an algorithm based on Newton-Raphson solver, a Variable Neighbourhood Search method, a Simulated Annealing algorithm and an Evolutionary Algorithm. We generate four data sets following a Gompertz-lognormal diffusion process using different noise level. The methods are applied with these data sets for estimating the parameters which are present into the diffusion process. Results show that bio-inspired methods gain suitable solutions for the problem every time, even when the noise level increase. On the other hand, some analytical methods as Newton-Raphson or the Iterative Method do not always solve the problem whether their scores depend on the starting point for initial solution or the noise level hinders the resolution of the problem. In these cases, the bio-inspired algorithms remain as a suitable and reliable approach.

In this article, we present a general framework for estimation of Artificial Neural Networks (ANN) parameters using the Generalized Method of Moments (GMM), as an alternative to the conventional

Quasi

Maximum Likelihood (QML). We used a simple generalization for nonlinear models of the usual orthogonality conditions from linear regression in addition to the moment conditions that replicate the QML estimation. Consequently the resultant models are overidentified. Monte Carlo simulations suggested that GMM can outperform QML in cases with small samples or elevated noise.

In this paper a novel hybrid ensemble method aiming at the improvement of models accuracy in regression tasks is presented. The proposed ensemble is composed by a

strong

learner trained exploiting data belonging to the whole training dataset and a set of specialised

weak

learners trained by using data coming from limited regions of the input space determined by means of a Self Organising Map based clustering. In the simulation phase, the strong and weak learners operate alternatively according to their punctual self-estimated reliabilities so as to handle each specific sample by means of the most promising learner. The method has been tested both on literature and real world datasets achieving satisfactory results.

Protecting the lagoon area from the wave attack is one of the primary challenges in coastal engineering. Due to the scarcity of rubble and also to achieve economy, new types of breakwaters are being used in place of conventional rubble mound breakwaters. Emerged Perforated Quarter Circle Breakwaters (EPQCB) are artificial concrete breakwaters consisting of a curved perforated face fronting the waves with a vertical wall on rear side and a base slab resting on a low rubble mound base. The perforated curved front face has advantages like energy dissipation and good stability with less material as it is hollow inside. The estimation of hydrodynamic performance characteristics of EPQCB by physical model studies is complex, expensive and time consuming. Hence, computational intelligence (CI) methods are adopted for the evaluation of the performance characteristics like reflection, dissipation, transmission, run-up, rundown etc. A number of CI methods like Artificial Neural Network (ANN), Fuzzy logic, and hybrids such as ANFIS, ANN-PCO (particle swarm optimization), ANN-ACO etc., are available and are being used. The paper presents the work carried out to predict the dependent output variable of loss coefficient (K

l

) beyond the range of values of one of the input variables i.e., wave period (T) adopted in present work, using the input data on variables of wave height (H), wave period (T), structure height (h

s

), water depth (d), radius of the breakwater (R), spacing of perforations (S) and diameter of perforations (D) using ANFIS. For this purpose, both the conventional method of data segregation and also a new method called ‘beyond data range’ method are used for both training the ANFIS models and also to predict the dependent variable. Further, the input data was fed to the models in both dimensional and non-dimensional form in order to understand the effect of using non-dimensional data in place of dimensional parametric data. The performance of ANFIS models for all the four cases mentioned above was studied and it was found that prediction using conventional method with non-dimensional parameters performed better than other three methods. ANFIS models can be used to predict the performance characteristic K

l

of EPQCB beyond the input data range of wave period T.

Least Squares Support Vector Regression (LS-SVR) is a powerful kernel-based learning tool for regression problems. Nonlinear system identification is one of such problems where we aim at capturing the behavior in time of a dynamical system by building a black-box model from the measured input-output time series. Besides the difficulties involved in the specification a suitable model itself, most real-world systems are subject to the presence of outliers in the observations. Hence, robust methods that can handle outliers suitably are desirable. In this regard, despite the existence of a few previous works on robustifying the LS-SVR for regression applications with outliers, its use for dynamical system identification has not been fully evaluated yet. Bearing this in mind, in this paper we assess the performances of two existing robust LS-SVR variants, namely WLS-SVR and RLS-SVR, in nonlinear system identification tasks containing outliers. These robust approaches are compared with standard LS-SVR in experiments with three artificial datasets, whose outputs are contaminated with different amounts of outliers, and a real-world benchmarking dataset. The obtained results for infinite step ahead prediction confirm that the robust LS-SVR variants consistently outperforms the standard LS-SVR algorithm.

The morphological gradient is a widely used edge detector for grey-level images in many applications. In this communication, we generalize the definition of the morphological gradient of the fuzzy mathematical morphology based on uninorms. Concretely, instead of defining the morphological gradient from the usual definitions of fuzzy dilation and erosion, where the minimum and the maximum are used, we define it from the generalized fuzzy dilation and erosion, where we consider a general t-norm and t-conorm, respectively. Once the generalized morphological gradient is defined, we determine which t-norm and t-conorm have to be considered in order to obtain a high performance edge detector. Some t-norms and their dual t-conorms are taken into account and the experimental results conclude that the t-norms of the Schweizer-Sklar family generate a morphological gradient which outperforms notably the classical morphological gradient based on uninorms.

The correction of the vignetting effect in digital images is a key pre-processing step in several computer vision applications. In this paper, some corrections and improvements to the image vignetting correction algorithm based on the minimization of the log-intensity entropy of the image are proposed. In particular, the new algorithm is able to deal with images with a vignetting that is not in the center of the image through the search of the optical center of the image. The experimental results show that this new version outperforms notably the original algorithm both from the qualitative and the quantitative point of view. The quantitative measures are obtained using an image database with images to which artificial vignetting has been added.

Methods of introducing diversity into ensemble learning predictors for regression problems are presented. Two methods are proposed in this paper, one involving pruning and the other a hybrid approach. In these ensemble learning approaches, diversity is introduced while simultaneously training, as part of the same learning process. Here not all members of the ensemble are trained in the same manner, but selectively trained, resulting in a diverse selection of ensemble members that have strengths in different parts of the training set. The result is that the prediction accuracy and generalization ability of the trained ensemble is enhanced. Pruning and hybrid heuristics attempt to combine accurate yet complementary members; therefore these methods enhance the performance by dynamically modifying the pruned aggregation through distributing the ensemble member selection over the entire dataset. A comparison is drawn with Negative Correlation Learning and a static ensemble pruning approach used in regression to highlight the performance improvement yielded by the dynamic methods. Experimental comparison is made using Multiple Layer Perceptron predictors on benchmark datasets, and on a signal calibration application.

In a previous work we showed that hardlimit multilayer neural networks have more computational power than sigmoidal multilayer neural networks [1]. In 1962 Minsky and Papert showed the limitations of a single perceptron which can only solve linearly separable classification problems and since at that time there was no algorithm to find the weights of a multilayer hardlimit perceptron research on neural networks stagnated until the early eighties with the invention of the Backpropagation algorithm [2]. Nevertheless since the sixties there have arisen some proposals of algorithms to implement logical functions with threshold elements or hardlimit neurons that could have been adapted to classification problems with multilayer hardlimit perceptrons and this way the stagnation of research on neural networks could have been avoided. Although the problem of training a hardlimit neural network is NP-Complete, our algorithm based on mathematical programming, a mixed integer linear model (MILP), takes few seconds to train the two input XOR function and a simple logical function of three variables with two minterms. Since any linearly separable logical function can be implemented by a perceptron with integer weights, varying them between -1 and 1 we found all the 10 possible solutions for the implementation of the two input XOR function and all the 14 and 18 possible solutions for the implementation of two logical functions of three variables, respectively, with a two layer architecture, with two neurons in the first layer. We describe our MILP model and show why it consumes a lot of computational resources, even a small hardlimit neural network translates into a MILP model greater than 1G, implying the use of a more powerful computer than a common 32 bits PC. We consider the reduction of computational resources as the near future work main objective to improve our novel MILP model and we will also try a nonlinear version of our algorithm based on a MINLP model that will consume less memory.

Software agents are increasingly used to search for experts, recommend resources, assess opinions, and other similar tasks in the context of social networks, which requires to have accurate information that describes the features of the members of the network. Unfortunately, many member profiles are incomplete, which has motivated many authors to work on automatic member labelling, that is, on techniques that can infer the null features of a member from his or her neighbourhood. Current proposals are based on local or global approaches; the former compute predictors from local neighbourhoods, whereas the latter analyse social networks as a whole. Their main problem is that they tend to be inefficient and their effectiveness degrades significantly as the percentage of null labels increases. In this paper, we present Katz, which is a novel hybrid proposal to solve the member labelling problem using neural networks. Our experiments prove that it outperforms other proposals in the literature in terms of both effectiveness and efficiency.

Some optimization problems arise when X-ray diffraction profiles are used to determine the microcrystalline characteristics of materials, like the detection of diffraction peaks and the deconvolution process necessary to obtain the pure diffraction profile. After applying the genetic algorithms to solve satisfactorily the first problem, in this work we propose two evolutionary algorithms to solve the deconvolution problem. This optimization problem targets the objective of obtaining the profile that contains the microstructural characteristics of a material from the experimental data and instrumental effects. This is a complex problem, ill-conditioned, since not only there are many possible solutions, but also some of them lack physical sense. In order to avoid such circumstance, the regularization techniques are used, where the optimization of some of their parameters by means of intelligent computing permits to obtain the optimal solutions of the problem.

The present document presents/displays a model of Neuronal Networks Artificial RNA for the prognosis of the rate of nominal change in Colombia, including flow orders and the differential of the interest rates like variables of entrance to the model. Additionally methodological conclusions from the traditional treatment of the series of time were extracted.

The Finger-to-nose test (FNT) is an accepted neurological evaluation to study the coordination conditions. In this work, a methodology for the analysis of data from FNT is proposed, aimed at assessing the evolution of the condition of Spinocerebellar Ataxia type 2 (SCA2) patients. First of all, test results obtained from both patients and healthy individuals are processed through principal component analysis in order to reduce data dimensionality. Next, data were grouped in order to determine classes of typical responses. The Mean Shift algorithm was used to perform an unsupervised clustering with no previous assumption on the number of clusters, whereas the

$$k$$

-means method provided an independent validation on the optimal cluster number. Experimental results showed the highest internal evaluation for distribution into three clusters, which could be identified as the responses of healthy subjects, SCA2 patients with medium incoordination level, and patients with severe incoordination. A membership function is defined, which allows to establish the subjects’ condition based on the classification of their responses. The results support that these protocols and the implemented clustering procedure can be used to accurately evaluate the incoordination stages of healthy subjects and SCA2 patients, thus offering a method to assess the impact of therapies and the progression of incoordination.

Ad-hoc mobile wireless network is characterized by a very dynamic environment. However, the major obstacle to be resolved is to sustain the links of continuity and improved routing performance. In this paper, we propose a predictor based Neuro-fuzzy for the prediction of mobility. It predicts the trajectory of an ad-hoc mobile node in order to improve routing performance by reducing overhead and the number of broken connections. It allows estimating the stability of paths in Ad-Hoc mobile wireless networks. Using an Adaptive Neuro-Fuzzy Inference System (ANFIS) to predict the trajectory of an ad-hoc mobile, we demonstrate the effectiveness of the proposed predictor by testing it on a time series prediction problems.

The complexity of the fuzzy classification models in Digital Forensics is considered to be one of the most significant aspects that influence a decision making process. We focus on criteria for an optimal SOM size and amount of rules to be derived that results in accurate and interpretable model. In this paper, we proposed a new method for the SOM size determination based on the data exploratory analysis. Experiments showed that the proposed method gives an accuracy on the Android malware detection up to 92% while decreasing the number of recommended rules from 189 to 24 in comparison to Vesanto method for an optimal SOM size. This is an important step for automated training of Neuro-Fuzzy that will result in human-understandable model that will be used in Digital Forensics process.

While Support Vector Regression, SVR, is one of the algorithms of choice in modeling problems, construction of its error intervals seems to have received less attention. On the other hand, general noise cost functions for SVR have been recently proposed. Taking this into account, this paper describes a direct approach to build error intervals for different choices of residual distributions. We also discuss how to fit these noise models and estimate their parameters, proceeding then to give a comparison between intervals obtained using this method. under different ways to estimate SVR parameters as well as the intervals obtained by employing a full SVR Bayesian framework. The proposed approach is shown on a synthetic problem to provide better accuracy when models fitted coincide with the noise injected into the problem. Finally, we apply it to wind energy forecasting, exploiting predicted energy magnitudes to define intervals with different widths.

In this paper, we combine several tools used in text-mining in order to study both the lexicon and the semantic structure of a set of medieval texts. On the one hand, the study of occurrences (Principal Component Analysis, Topic Models, Self-Organizing Maps, Hierarchical Cluster Analysis) allows a wide scope of tools to extract and display information from big data. On the other hand, the study of co-occurrences (words belonging to a sentence, a paragraph) allows to keep track of the structure of each text, but is more tedious to handle and often leads to messy visualizations. Here we use the SOM algorithm to reduce the size of the data (clustering, removal of fickle information) while preserving the semantic structure ; then we can rely on classical but slower algorithms (HCA, graph representation) to purpose data visualization.

A saccade is an ocular movement that is characterized by speed and precision. The velocity profile of this movement is used to extract the maximum speed value, that is one of the most important features of the saccade. A gamma function was used by other authors to describe the waveform shape of the velocity profile. However, this function does not present an optimal profile description in records of patients suffering from Spinocerebellar Ataxia type 2. In order to find a function that better describes the velocity profile, this contribution compares the fittings of several functions through visual and numerical analysis. Results showed a better performance of the partial sums of gaussian series and a gaussian fit function.

Taking high resolution photos with mobile devices anytime anywhere is becoming increasingly common. Therefore, images of all kinds of text documents are recorded. This work presents esCam, an application for Android platform, whose goal is to preprocess the images of those text documents, in particular, perspective correction and image cleaning and enhancing. What truly differentiates our application is that esCam focuses on treatment of text that may appear in the image, using neural networks. These preprocessing steps are needed to make easier the digitalization and also to benefit subsequent steps such as document analysis and text recognition.

Personal computers and smartphones in their standard form are, in general, inaccessible to people with reduced mobility. It has been necessary to design alternative interfaces and peripherals that allow them to use the technology comfortably and effectively, a difficult task given the heterogeneity of the physical and cognitive profiles. We will demostrate a solution for computer access and alternative communication for people with disabilities using three of the most effective access methods: switch-based input, head tracking and eye tracking, and custom designed software.

One of the major risks of disability is a loss of autonomy that , in extreme, may lead to institutionalization. Lack of human resources for caregiving has led to designing robots to assist people in need. Assistive robotics are meant to help people cope with Activities of Daily Living (ADL). Most ADL are heavily affected by issues related to ambulation [

1

], so much effort in assistive robots has focused on robotic wheelchairs, rollators, walkers and even canes. These devices typically provide monitorization, physical support and help to cope with hazardous and/or complex situations. However, it is of key importance to provide just the right amount of help to people with disabilities. According to clinicians, an excess of assistance may lead to frustration and/or loss of residual skills. Lack of assistance, however, may lead to unacceptable risks and/or failure to accomplish the desired task. Hence, help must be adapted to each specific user.