Search

Recognizing human activities seamlessly and ubiquitously is now closer than ever given the myriad of sensors readily deployed on and around users. However, the training of recognition systems continues to be both time and resource-consuming, as datasets must be collected ad-hoc for each specific sensor setup a person may encounter in their daily life. This work presents an alternate approach based on transfer learning to opportunistically train new unseen or target sensor systems from existing or source sensor systems. The approach uses system identification techniques to learn a mapping function that automatically translates the signals from the source sensor domain to the target sensor domain, and vice versa. This can be done for sensor signals of the same or cross modality. Two transfer models are proposed to translate recognition systems based on either activity templates or activity models, depending on the characteristics of both source and target sensor systems. The proposed transfer methods are evaluated in a human–computer interaction scenario, where the transfer is performed in between wearable sensors placed at different body locations, and in between wearable sensors and an ambient depth camera sensor. Results show that a good transfer is possible with just a few seconds of data, irrespective of the direction of the transfer and for similar and cross sensor modalities.

Recent years have witnessed a widespread in the use of interval type-2 fuzzy logic systems (IT2 FLSs) in real-world applications. It has been shown recently that interval type-2 fuzzy sets (IT2 FSs) are more general than interval-valued fuzzy sets (IV FSs) [1] . Hence, there is a need to explore the capabilities of the more general forms of IT2 FSs (beyond IV FSs) and the applications areas they will be more suitable for. In addition, there is a need to develop the theory of the general forms of IT2 FLSs (gfIT2 FLSs), which employ IT2 FSs that are not equivalent to IV FSs and can have nonconvex secondary membership functions (MFs). Although these systems could be considered within the scope of general type-2 FLSs (GT2 FLSs), the practical implementation of GT2 FLSs has traditionally required the secondary MFs to be convex and normal type-1 fuzzy sets (T1 FSs). In addition, the type-reduction operation still presents a challenge for GT2 FLSs because of its computational complexity. In this paper, we present a complete framework for a type-2 FLS that uses the most recent perception of IT2 FSs (the so called general forms of interval type-2 fuzzy sets, gfIT2 FSs), whose secondary grades can be nonconvex T1 FSs. This framework includes new equations for the meet and join operations of gfIT2 FSs, as well as a new type reduction procedure for the type-2 FLS involving gfIT2 FSs. In addition, we present the type-2 FLS operation for singleton and nonsingleton fuzzification. We will introduce the various operations employed within a gfIT2 FLSs, from fuzzification (including singleton and nonsingleton) to inference, type-reduction, and defuzzification. We will also present two examples in which these gfIT2 FSs arise naturally when modeling sonar sensors input noise and the antecedents/consequents from a survey including different users.

State of the art classifiers split Alzheimer’s disease progression into a limited number of stages and use a comparatively small database. For the best treatment, it is desirable to have the highest resolution from the progression of the disease. This paper proposes a reliable deep convolutional neural network for the classification of six different Alzheimer’s disease stages based on Magnetic Resonance Imaging (MRI). The peculiarity of this paper is the introduction of a new, sixth, disease stage, and the large amount of data that has been taken into account. Additionally, not only the testing accuracy is analyzed, but also the robustness of the classifier to have feedback on how certain the neural network makes its predictions.

Lower limb injuries, especially those related to the knee joint, are some of the most common and severe injuries among sport practitioners. Consequently, a growing interest in the identification of subjects with high risk of injury has emerged during last years. One of the most commonly used injury risk factor is the measurement of joint angles during the execution of dynamic movements. To that end, techniques such as human motion capture and video analysis have been widely used. However, traditional procedures to measure joint angles present certain limitations, which makes this practice not practical in common clinical settings. This work presents SPIRA, a novel 2D video analysis system directed to support practitioners during the evaluation of joint angles in functional tests. The system employs an infrared camera to track retro-reflective markers attached to the patient’s body joints and provide a real-time measurement of the joint angles in a cost-and-time-effective way. The information gathered by the sensor is processed and managed through a computer application that guides the expert during the execution of the tests and expedites the analysis of the results. In order to show the potential of the SPIRA system, a case study has been conducted, performing the analysis with the both the proposed system and a gold-standard in 2D offline video analysis. The results (ICC( $$\rho$$ ) = 0.996) reveal a good agreement between both tools and prove the reliability of SPIRA.

This work reviews the challenge of power consumption prediction, approaching both short-term and long-term prediction problems using neural networks. A number of improvements are introduced for both problems using two types of neural nets. For short-term prediction, a modified LSTM network based on direct prediction of four hours horizon is presented, also an alternative model based on Convolutional Neural Networks is also introduced. Different improvements in the short-term prediction by the use of external inputs, and the concatenation of the LSTM sub-net outputs for the prediction, among others, are shown. Finally, long-term forecasting is considered and a modified LSTM model is proposed and trained for that purpose, achieving notable improvements with respect to non-dedicated models.

Ponencia de la conferencia "15th International Work-Conference on Artificial Neural Networks, IWANN 2019; Gran Canaria; Spain; 12 June 2019 through 14 June 2019" The automatic recognition of physical activities typically involves various signal processing and machine learning steps used to transform raw sensor data into activity labels. One crucial step has to do with the segmentation or windowing of the sensor data stream, as it has clear implications on the eventual accuracy level of the activity recogniser. While prior studies have proposed specific window sizes to generally achieve good recognition results, in this work we explore the potential of fusing multiple equally-sized subwindows to improve such recognition capabilities. We tested our approach for eight different subwindow sizes on a widely-used activity recognition dataset. The results show that the recognition performance can be increased up to 15% when using the fusion of equally-sized subwindows compared to using a classical single window.

Heart disease is currently one of the leading causes of death in developed countries. The electrocardiogram is an important source of information for identifying these conditions, therefore, becomes necessary to seek an advanced system of diagnosis based on these signals. In this paper we used samples of electrocardiograms of MIT-related database with ten types of pathologies and a rate corresponding to normal (healthy patient), which are processed and used for extraction from its two branches of a wide range of features. Next, various techniques have been applied to feature selection based on genetic algorithms, principal component analysis and mutual information. To carry out the task of intelligent classification, 3 different scenarios have been considered. These techniques allow us to achieve greater efficiency in the classification methods used, namely support vector machines (SVM) and decision trees (DT) to perform a comparative analysis between them. Finally, during the development of this contribution, the use of very non-invasive devices (2 channel ECG) was analyzed, we could practically classify them as wearable, which would not need interaction by the user, and whose energy consumption is very small to extend the average life of the user been on it.

Recent years have witnessed a huge progress in the automatic identification of individual primitives of human behavior, such as activities or locations. However, the complex nature of human behavior demands more abstract contextual information for its analysis. This work presents an ontology-based method that combines low-level primitives of behavior, namely activity, locations and emotions, unprecedented to date, to intelligently derive more meaningful high-level context information. The paper contributes with a new open ontology describing both low-level and high-level context information, as well as their relationships. Furthermore, a framework building on the developed ontology and reasoning models is presented and evaluated. The proposed method proves to be robust while identifying high-level contexts even in the event of erroneously-detected low-level contexts. Despite reasonable inference times being obtained for a relevant set of users and instances, additional work is required to scale to long-term scenarios with a large number of users.

The rapid growth of social networks has 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 data used were collected after performing an experiment using the Zong Payment tool. The paper includes a comparative analysis of several machine learning methods applied to the variable selection problem considering strategic, sociodemographic, and behavioral variables. The methods were applied isolatedly and, in a novel approach, using multiobjective optimization criteria. The results obtained by these algorithms were validated by an expert committee, who ranked them on the basis of their quality. The proposed methodology obtained the best results in terms of optimization as well as the highest scores from the experts.

The evaluation of multiple sequence alignments (MSAs) is still an open task in bioinformatics. Current MSA scores do not agree about how alignments must be accurately evaluated. Consequently, it is not trivial to know the quality of MSAs when reference alignments are not provided. Recent scores tend to use more complex evaluations adding supplementary biological features. In this work, a set of novel regression approaches are proposed for the MSA evaluation, comparing several supervised learning and mathematical methodologies. Therefore, the following models specifically designed for regression are applied: regression trees, a bootstrap aggregation of regression trees (bagging trees), least-squares support vector machines (LS-SVMs) and Gaussian processes. These algorithms consider a heterogeneous set of biological features together with other standard MSA scores in order to predict the quality of alignments. The most relevant features are then applied to build novel score schemes for the evaluation of alignments. The proposed algorithms are validated by using the BAliBASE benchmark. Additionally, an statistical ANOVA test is performed to study the relevance of these scores considering three alignment factors. According to the obtained results, the four regression models provide accurate evaluations, even outperforming other standard scores such as BLOSUM, PAM or STRIKE.

The recognition of human activity has been deeply explored during the recent years. However, most proposed solutions are mainly devised to operate in ideal conditions, thus not addressing crucial real-world issues. One of the most prominent challenges refers to common sensor technological anomalies. Sensor faults and failures introduce variations in the measured sensor data with respect to the equivalent observations in ideal conditions. As a consequence, predefined recognition systems may potentially fail to identify actions in the anomalous sensor data. This paper presents a novel model devised to cope with the effects introduced by sensor technological anomalies. The model builds on the knowledge gained from multi-sensor configurations, through asymmetrically weighting the decisions provided at both activity and sensor levels. Insertion and rejection weighting metrics are particularly used to eventually yield a unique recognized activity. For the sake of comparison, the tolerance to sensor faults and failures of standard activity recognition systems and the new proposed model are evaluated. The results prove classic activity-aware systems to be incapable of recognition under the effects of sensor technological anomalies, while the proposed model demonstrates to be robust against both sensor faults and failures.

Nowadays, the uncovering of new functional relationships between proteins is one of the major goals of biological studies. For this task, the integration of evidences from heterogeneous data sources by means of machine learning methodologies has been demonstrated to be an effective way of providing a complete genome-wide functional network and more accurate inferences of new functional associations. This work presents a new framework to be used in Artificial Neural Networks (ANNs) for the task of predicting functional relationships between proteins through the integration of evidences from heterogeneous data sources. The developing of such new methodology is motivated by the problems that arise when applying ANNs to this kind of problems, namely, the computational cost of ANN optimization process due to the nature of data (large number of instances and high dimensionality). The method selects smaller representative/non-random subsets from the original data set selected for ANN optimization process, resulting in a reduction of the number of data to be trained and, consequently, the computational cost. Moreover, the fact that the subsets are not only smaller, but also representative from the original one, (i) prevents the repetition of the optimization process several times with different random subsets of data, which is commonly used to get a reliable and fair evaluation of ANN's prediction accuracy, and (ii) benefits the learning procedure in the sense of a reduction of the overfitting problem, improving, this way, the prediction ability.

Singleton interval type-2 fuzzy logic systems (FLSs) have been widely applied in several real-world applications, where it was shown that the singleton interval type-2 FLSs outperform their singleton type-1 counterparts in applications with high uncertainty levels. However, one of the main criticisms of singleton interval type-2 FLSs is the fact that they outperform singleton type-1 FLSs solely based on their use of extra degrees of freedom (extra parameters) and that type-1 FLSs with a sufficiently large number of parameters may provide the same performance as interval type-2 FLSs. In addition, most works on type-2 FLSs only compare their results with singleton type-1 FLSs but fail to consider nonsingleton type-1 systems. In this paper, we aim to directly address and investigate this criticism. In order to do so, we will perform a comparative study between optimized singleton type-1, nonsingleton type-1, and singleton interval type-2 FLSs under the presence of noise. We will also present a multiobjective evolutionary algorithm (MOEA) for the optimization of singleton type-1, nonsingleton type-1, and singleton interval type-2 fuzzy systems for function approximation problems. The MOEA will aim to satisfy two objectives to maximize the accuracy of the FLS and minimize the number of rules in the FLS, thus improving its interpretability. Furthermore, we will present a methodology to obtain “optimal” consequents for the FLSs. Hence, this paper has two main contributions: First, it provides a common methodology to learn the three types of FLSs (i.e., singleton type-1, nonsingleton type-1, and singleton interval type-2 FLSs) from data samples. The second contribution is the creation of a common framework for the comparison of type-1 and type-2 FLSs that allows us to address the aforementioned criticism. We provide details of a series of experiments and include statistical analysis showing that the type-2 FLS is able to handle higher levels of noise than its nonsingleton and singleton type-1 counterparts.

An enormous effort has been made during the recent years towards the recognition of human activity based on wearable sensors. Despite the wide variety of proposed systems, most existing solutions have in common to solely operate on predefined settings and constrained sensor setups. Real-world activity recognition applications and users rather demand more flexible sensor configurations dealing with potential adverse situations such as defective or missing sensors. In order to provide interoperability and reconfigurability, heterogeneous sensors used in wearable activity recognition systems must be fairly abstracted from the actual underlying network infrastructure. This work presents MIMU-Wear, an extensible ontology that comprehensively describes wearable sensor platforms consisting of mainstream magnetic and inertial measurement units (MIMUs). MIMU-Wear describes the capabilities of MIMUs such as their measurement properties and the characteristics of wearable sensor platforms including their on-body location. A novel method to select an adequate replacement for a given anomalous or nonrecoverable sensor is also presented in this work. The proposed sensor selection method is based on the MIMU-Wear Ontology and builds on a set of heuristic rules to infer the candidate replacement sensors under different conditions. Then, queries are iteratively posed to select the most appropriate MIMU sensor for the replacement of the defective one. An exemplary application scenario is used to illustrate some of the potential of MIMU-Wear for supporting seamless operation of wearable activity recognition systems.

Joints angles are some of the most common measurements for the evaluation of lower limb injury risk, specially of lower limb joints. The 2D projections of these angles, as the Frontal Plane Projection Angle (FPPA), are widely used as an estimation of the angle value. Traditional procedures to measure 2D angles imply huge time investments, primarily when evaluating multiple subjects. This work presents a novel 2D video analysis system directed to capture the joint angles in a cost-and-time-effective way. It employs Kinect V2 depth sensor to track retro-reflective markers attached to the patient’s joints to provide an automatic estimation of the desired angles. The information registered by the sensor is processed and managed by a computer application that expedites the analysis of the results. The reliability of the system has been studied against traditional procedures obtaining excellent results. This system is aimed to be the starting point of an autonomous injury prediction system based on machine learning techniques.

Recently, it has been shown that interval type-2 fuzzy sets (IT2FSs) are more general than interval-valued fuzzy sets (IVFSs), and some of these IT2FSs can actually be non-convex. Although these IT2FSs could be considered within the general type-2 fuzzy sets’ (GT2FSs) scope, this latter have always been studied and developed under certain conditions considering the convexity and normality of their secondary grades. In recent works the operations of intersection and union for GT2FSs have been extended to include non-convex secondary grades. Hence, there is a need to develop the theory for those general forms of interval type-2 fuzzy logic systems (gfIT2FLSs) which use IT2FSs that are not equivalent to IVFSs and can have non-convex secondary grades. In this chapter, we will present the mathematical tools to define the inference engine for singleton gfIT2FLSs. This work aims to introduce the basic structure of such singleton gfIT2FLSs, paying special attention to those blocks presenting significant differences with the already well known type-2 FLSs which employ IT2FSs which are equivalent to IVFSs (we will term IVFLSs).

The online evolution and learning of fuzzy systems is highly important when dealing with changing environments over time. This capability is especially relevant in the field of control, due to the special characteristics of control problems. In this field, techniques capable of developing controllers with a minimum amount of prior knowledge about the plants to be controlled are desired. Furthermore, these controllers should provide a reduced number of interpretable rules. This paper presents a new Online Self-Evolving Neuro Fuzzy controller based on the Taylor Series Neuro Fuzzy (TaSe-NF) model. Under the assumption of no prior knowledge about the differential equations that define the plant to be controlled, this methodology is capable of incrementally evolving the structure of the controller and adapting its parameters online, while controlling the plant. The new methodology uses a scattered distribution of the fuzzy rules, thereby reducing the number of rules in the fuzzy controller. Moreover, the use of the TaSe-NF model to represent the antecedents of the rules enhances the interpretability of the obtained rules. The proposed evolving fuzzy controller is composed of two main blocks: On the one hand, the online local learning of the rule consequents tackles the task of providing a proper control at the present moment. On the other hand, a structure self-evolution method analyzes the error surface to determine which cluster/rule suffers the worst performance and therefore, needs to be further split. Simulation results are presented to illustrate the capabilities of this new online self-evolving controller.

The analysis of daily living human behavior has proven to be of key importance to prevent unhealthy habits. The diversity of activities and the individuals' particular execution style determine that several sources of information are normally required. One of the main issues is to optimally combine them to guarantee performance, scalability and robustness. In this paper we present a fusion classification methodology which takes into account the potential of the individual decisions yielded at both activity and sensor classification levels. Particularly tested on a wearable sensors based system, the method reinforces the idea that some parts of the body (i.e., sensors) may be specially informative for the recognition of each particular activity, thus supporting the ranking of the decisions provided by each associated sensor decision entity. Our method systematically outperforms the results obtained by traditional multiclass models which otherwise may require a high-dimensional feature space to acquire a similar performance. The comparison with other activity-recognition fusion approaches also demonstrates our model scales significantly better for small sensor networks.

The benefits arising from proactive conduct and subject-specialized healthcare have driven e-health and e-monitoring into the forefront of research, in which the recognition of motion, postures and physical exercise is one of the main subjects. We propose here a multidisciplinary method for the recognition of physical activity with the emphasis on feature extraction and selection processes, which are considered to be the most critical stages in identifying the main unknown activity discriminant elements. Efficient feature selection processes are particularly necessary when dealing with huge training datasets in a multidimensional space, where conventional feature selection procedures based on wrapper methods or 'branch and bound' are highly expensive in computational terms. We propose an alternative filter method using a feature quality group ranking via a couple of two statistical criteria. Satisfactory results are achieved in both laboratory and semi-naturalistic activity living datasets for real problems using several classification models, thus proving that any body sensor location can be suitable to define a simple one-feature-based recognition system, with particularly remarkable accuracy and applicability in the case of the wrist.

In modern proteomics, prediction of protein-protein interactions (PPIs) is a key research line, as these interactions take part in most essential biological processes. In this paper, a new approach is proposed to PPI data classification based on the extraction of genomic and proteomic information from well-known databases and the incorporation of semantic measures. This approach is carried out through the application of data mining techniques and provides very accurate models with high levels of sensitivity and specificity in the classification of PPIs. The well-known support vector machine paradigm is used to learn the models, which will also return a new confidence score which may help expert researchers to filter out and validate new external PPIs. One of the most-widely analyzed organisms, yeast, will be studied. We processed a very high-confidence dataset by extracting up to 26 specific features obtained from the chosen databases, half of them calculated using two new similarity measures proposed in this paper. Then, by applying a filter-wrapper algorithm for feature selection, we obtained a final set composed of the eight most relevant features for predicting PPIs, which was validated by a ROC analysis. The prediction capability of the support vector machine model using these eight features was tested through the evaluation of the predictions obtained in a set of external experimental, computational, and literature-collected datasets.

The majority of clustering methods that have been proposed in the literature are focused on classification problems, and more specifically on pattern recognition problems. In this paper, we provide an enhanced clustering function approximation (ECFA) methodology that is especially suited for function approximation problems, which calculates the error committed in every cluster using the real output of the radial basis function neural network (RBFN), and not just an approximate value of that output, trying to concentrate more clusters in those input regions where the approximation error is bigger, thus attempting to homogenize the contribution to the error of every cluster. ECFA migrates clusters to the zones of the input space where the approximation error is bigger, thus trying to homogenously distribute the total distortion in every cluster, producing a better share-out of clusters for the data input space. We will see that ECFA outperforms other clustering techniques, such as the FCM algorithm or the CFA algorithm not only with respect to the final approximation error but also with respect to the execution time.

Protein-protein interactions (PPIs) play a crucial role in cellular processes. In the present work, a new approach is proposed to construct a PPI predictor training a support vector machine model through a mutual information filter-wrapper parallel feature selection algorithm and an iterative and hierarchical clustering to select a relevance negative training set. By means of a selected suboptimum set of features, the constructed support vector machine model is able to classify PPIs with high accuracy in any positive and negative datasets.

Radial basis function networks are traditionally known as local approximation networks as they are composed by a number of elements which, individually, mainly take care of the approximation about a specific area of the input space. Then, the joint global output of the network is obtained as a linear combination of the individual elements' output. However, in the network optimization, the performance of the global model is normally the only objective to optimize. This might cause a deficient local modelling of the input space, thus partially losing the local character of this type of models. This work presents a modified radial basis function network that maintains the approximation capabilities of the local sub-models whereas the model is globally optimized. This property is obtained thanks to a special partitioning of the input space, that leads to a direct global–local optimization. A learning methodology adapted to the proposed model is used in the simulations, consisting of a clustering algorithm for the initialization of the centers and a local search technique. In the experiments, the proposed model shows satisfactory local and global modelling capabilities both in artificial and real applications.

Protein–protein interaction (PPI) prediction is one of the main goals in the current Proteomics. This work presents a method for prediction of protein–protein interactions through a classification technique known as support vector machines. The dataset considered is a set of positive and negative examples taken from a high reliability source, from which we extracted a set of genomic features, proposing a similarity measure. From this dataset we extracted 26 proteomics/genomics features using well-known databases and datasets. Feature selection was performed to obtain the most relevant variables through a modified method derived from other feature selection methods for classification. Using the selected subset of features, we constructed a support vector classifier that obtains values of specificity and sensitivity higher than 90% in prediction of PPIs, and also providing a confidence score in interaction prediction of each pair of proteins.

In this study, a novel fuzzy controller, which is able to self-design from scratch, while working online, is proposed. The controller does not use the information regarding the differential equations that govern the plant's behavior or any of their bounds. The algorithm presented is able to determine the most-adequate topology for the fuzzy controller based on the data obtained during the system's normal operation. Therefore, the controller can start operating with an empty set of fuzzy rules and needs no offline training. The proposed methodology comprises two phases: adaptation of the consequents for every selected topology and online addition of new membership functions (MFs). Some of the main advantages of this method are its robustness under changes on the plant's dynamics, good performance in noisy situations, and the ability to perform variable selection among a group of candidate variables. Unlike other online methods, the modification of the topology is based on the analysis of the whole operating region of the plant, thus providing higher robustness. Several simulation examples are used to show these features.

Optimization of the local sub-models when using neuro-fuzzy systems to model input/output data can be spoiled during the global optimization process, therefore special care has to be taken to avoid this loss. Some approaches involving grid input space partitioning that optimize local sub-models whilst still obtaining a good global system performance have appeared. However, this is a more complex task for radial basis function networks, scatter-partitioning fuzzy systems and neuro-fuzzy models in general. This work presents a modified neuro-fuzzy model, known as the TaSe-NF model, which is a special case of a scatter-partitioned Takagi--Sugeno--Kang fuzzy system or normalized radial basis function neural network. The TaSe-NF model maintains the optimization properties of the local sub-models (fuzzy rules or RBF nodes) when the model is globally optimized thanks to the modified calculation of the final normalized activation of the rules, which provides an appropriate input space partitioning to achieve these objectives. To show the characteristics of the proposed model, a learning methodology for this model, which consists of a clustering algorithm especially suited to function approximation problems, a local search technique and a membership function merging approach, with the objective of improving the transparency, i.e. ease of interpretability, of the extracted fuzzy rule set is also proposed.

Color change modelling is an essential problem in a wide range of colorimetric applications. Specifically, the approximation to the color expected after a physical chemical or natural color is essential in industry and in color science in general. This work proposes a modified neuro-fuzzy approach as a solution for a color change modelling problem. Neuro-fuzzy systems are well known methods for data modelling; their main advantage is their ability to provide an accurate solution from which an interpretable set of rules that can be extracted and used by the scientists. However these models have the problem that the global approximation optimization can lead to a deficient interpretation of the rules extracted from the model. This work proposes a modified neuro-fuzzy model that performs a simultaneous global and local modelling; this property is reached thanks to a special partitioning of the input space in the system. Specifically, the proposed methodology will be applied to a very important problem from the clinical and odontologic point of view as it is the modelling of the tooth color variation after a bleaching process. The availability of tools that help to predict these changes, from the initial chromaticity of the tooth, can solve the problem of lack of information on the expected tooth color after a specific treatment and help the dentist in the decision making on the most appropriate protocol for this treatment and in providing adequate information to the patient.

This paper presents a fast and new deterministic model selection methodology for incremental radial basis function neural network (RBFNN) construction in time series prediction problems. The development of such special designed methodology is motivated by the problems that arise when using a K-fold cross-validation-based model selection methodology for this paradigm: its random nature and the subjective decision for a proper value of K, resulting in large bias for low values and high variance and computational cost for high values. Taking into account these drawbacks, the proposed model selection approach is a combined algorithm that takes advantage of two balanced and representative training and validation sets for their use in RBFNN initialization, optimization and network model evaluation. This way, the model prediction accuracy is improved, getting small variance and bias, reducing the computation time spent in selecting the model and avoiding random and computationally expensive model selection methodologies based on K-fold cross-validation procedures.

The Nickel reduction process is a complex task where many dynamic optimization problems arises that, nowadays, requires a human operator to take decisions based on his experience and intuition. In order to help the operator to optimize the reduction process in terms of maximum amount of mineral extracted and minimum energy consumption, a control system integrated by several modules is being designed. One of the modules has the task of predicting how much petroleum will be burned in the ovens where the raw material is processed. This paper proposes an algorithm to design Radial Basis Function Neural Networks that will be able to predict accurately the amount of petroleum given a set of input parameters. The algorithm is also able of identifying the most relevant input parameters for the network so the dimensionality reduction problem is ameliorated. Hence, this paper, as it will be shown in the experiments section is able to apply the synergy of different Soft Computing techniques to the industrial process obtaining satisfactory results.

Least squares support vector machines (LSSVM) with Gaussian kernel represent the most used of the kernel methods existing in the literature for regression and time series prediction. These models have a good behaviour for these types of problems due to their generalization capabilities and their smooth interpolation, but they are very dependent on the feature selection performed and their computational cost notably increases with the number of training samples. Time series prediction can be tackled as a regression problem by constructing a set of input/output data from the series; this traditional approach and the use of typical recursive or direct strategies present serious drawbacks in long-term prediction. This paper presents an alternative based on the settings of specific-to-problem kernels to be applied to time series prediction focusing on large scale prediction. A simple methodology for kernel creation based on the periodicities in time series data is proposed. An alternative to LSSVM models with lower computational cost, the Kernel Weighted K-Nearest Neighbours (KWKNN) is described for function approximation. A parallel version of KWKNN is also presented to deal with large data sets.

The problem of selecting the patterns to be learned by any model is usually not considered by the time of designing the concrete model but as a preprocessing step. Information theory provides a robust theoretical framework for performing input variable selection thanks to the concept of mutual information. Recently the computation of the mutual information for regression tasks has been proposed so this paper presents a new application of the concept of mutual information not to select the variables but to decide which prototypes should belong to the training data set in regression problems. The proposed methodology consists in deciding if a prototype should belong to or not to the training set using as criteria the estimation of the mutual information between the variables. The novelty of the approach is to focus in prototype selection for regression problems instead of classification as the majority of the literature deals only with the last one. Other element that distinguishes this work from others is that it is not proposed as an outlier detector but as an algorithm that determines the best subset of input vectors by the time of building a model to approximate it. As the experiment section shows, this new method is able to identify a high percentage of the real data set when it is applied to highly distorted data sets.

The visible/near-infrared spectrum consists of overtones and combination bands of the fundamental molecular absorptions found in the visible and near-infrared region. The analysis of the spectrum might be difficult because overlapping vibrational bands may appear nonspecific and poorly resolved. Nevertheless, the information it could be retrieved from the analysis of the spectrum might be very useful for the food industry producers, consumers, and food distributors because the meat could be classified based on the spectrum in several aspects such as the quality, tenderness, and kind of meat. This paper applies Mutual Information theory and several classification models (Radial Basis Function Neural Networks and Support Vector Machines) in order to determine the breed of pork meat (Iberian or White) using only as input the infrared spectrum. First, the more relevant wavelengths from the spectrum will be chosen, then, those wavelengths will be the input data to design the classifiers. As the experiments will show, the proposed techniques, when applied with a correct design methodology are capable of obtaining quality results for this specific problem.

The design of radial basis function neural networks (RBFNNs) still remains as a difficult task when they are applied to classification or to regression problems. The difficulty arises when the parameters that define an RBFNN have to be set, these are: the number of RBFs, the position of their centers and the length of their radii. Another issue that has to be faced when applying these models to real world applications is to select the variables that the RBFNN will use as inputs. The literature presents several methodologies to perform these two tasks separately, however, due to the intrinsic parallelism of the genetic algorithms, a parallel implementation will allow the algorithm proposed in this paper to evolve solutions for both problems at the same time. The parallelization of the algorithm not only consists in the evolution of the two problems but in the specialization of the crossover and mutation operators in order to evolve the different elements to be optimized when designing RBFNNs. The subjacent genetic algorithm is the non-sorting dominated genetic algorithm II (NSGA-II) that helps to keep a balance between the size of the network and its approximation accuracy in order to avoid overfitted networks. Another of the novelties of the proposed algorithm is the incorporation of local search algorithms in three stages of the algorithm: initialization of the population, evolution of the individuals and final optimization of the Pareto front. The initialization of the individuals is performed hybridizing clustering techniques with the mutual information (MI) theory to select the input variables. As the experiments will show, the synergy of the different paradigms and techniques combined by the presented algorithm allow to obtain very accurate models using the most significant input variables.

The auditory steady-state response is an EEG potential elicited by the repetitive presentation of auditory stimuli. Researchers have found contradictory results about the influence of cognitive tasks, such as the selective attention, over this potential. It has been proved that selective attention is able to modulate cortex originated steady-state responses, such as the visual. This fact has been widely used to develop brain-computer interfaces. However for complete locked-in patients, such as those in an advanced state of Amyotrophic lateral sclerosis, visual stimuli are not longer suitable, hence the need of another type of stimulus, generally auditory, for both stimulation and feedback. In this paper we present a study based on artificial neural networks that evidences the effects of selective attention over auditory steady-state responses and the use in brain-computer interfaces is discussed.

Traditionally, the autoregressive moving average (ARMA) model has been one of the most widely used linear models in time series prediction. Recent research activities in forecasting with artificial neural networks (ANNs) suggest that ANNs can be a promising alternative to the traditional ARMA structure. These linear models and ANNs are often compared with mixed conclusions in terms of the superiority in forecasting performance. In this paper we propose a hybridization of intelligent techniques such as ANNs, fuzzy systems and evolutionary algorithms, so that the final hybrid ARIMA-ANN model could outperform the prediction accuracy of those models when used separately. More specifically, we propose the use of fuzzy rules to elicit the order of the ARMA or ARIMA model, without the intervention of a human expert, and the use of a hybrid ARIMA-ANN model that combines the advantages of the easy-to-use and relatively easy-to-tune ARIMA models, and the computational power of ANNs.

The challenge of predicting future values of a time series covers a variety of disciplines. The fundamental problem of selecting the order and identifying the time varying parameters of an autoregressive moving average model (ARMA) concerns many important fields of interest such as linear prediction, system identification and spectral analysis. Recent research activities in forecasting with artificial neural networks (ANNs) suggest that ANNs can be a promising alternative to the traditional ARMA structure. These linear models and ANNs are often compared with mixed conclusions in terms of the superiority in forecasting performance. This study was designed: (a) to investigate a hybrid methodology that combines ANN and ARMA models; (b) to resolve one of the most important problems in time series using ARMA structure and Box-Jenkins methodology: the identification of the model. In this paper, we present a new procedure to predict time series using paradigms such as: fuzzy systems, neural networks and evolutionary algorithms. Our goal is to obtain an expert system based on paradigms of artificial intelligence, so that the linear model can be identified automatically, without the need of human expert participation. The obtained linear model will be combined with ANN, making up an hybrid system that could outperform the forecasting result.

Clustering algorithms have been successfully applied in several disciplines. One of those applications is the initialization of radial basis function (RBF) centers composing a neural network, designed to solve functional approximation problems. The Clustering for Function Approximation (CFA) algorithm was presented as a new clustering technique that provides better results than other clustering algorithms that were traditionally used to initialize RBF centers. Even though CFA improves performance against other clustering algorithms, it has some flaws that can be improved. Within those flaws, it can be mentioned the way the partition of the input data is done, the complex migration process, the algorithm's speed, the existence of some parameters that have to be set in order to obtain good solutions, and the convergence is not guaranteed. In this paper, it is proposed an improved version of this algorithm that solves the problems that its predecessor have using fuzzy logic successfully. In the experiments section, it will be shown how the new algorithm performs better than its predecessor and how important is to make a correct initialization of the RBF centers to obtain small approximation errors.

There exists a wide range of paradigms, and a high number of different methodologies that are applied to the problem of time series prediction. Most of them are presented as a modified function approximation problem using input/output data, in which the input data are expanded using values of the series at previous steps. Thus, the model obtained normally predicts the value of the series at a time (t+h) using previous time steps (t-@t"1),(t-@t"2),...,(t-@t"n). Nevertheless, learning a model for long term time series prediction might be seen as a more complicated task, since it might use its own outputs as inputs for long term prediction (recursive prediction). This paper presents the utility of two different methodologies, the TaSe fuzzy TSK model and the least-squares SVMs, to solve the problem of long term time series prediction using recursive prediction. This work also introduces some techniques that upgrade the performance of those advanced one-step-ahead models (and in general of any one-step-ahead model), where they are used recursively for long term time series prediction.

The function approximation problem has been tackled many times in the literature by using radial basis function neural networks (RBFNNs). In the design of these neural networks there are several stages where, the most critical stage is the initialization of the centers of each RBF since the rest of the steps to design the RBFNN strongly depend on it. The improved clustering for function approximation (ICFA) algorithm was recently introduced and proved successful for the function approximation problem. In the ICFA algorithm, a fuzzy partition of the input data is performed but, a fuzzy partition can behave inadequately in noise conditions. Possibilistic and mixed approaches, combining fuzzy and possibilistic partitions, were developed in order to improve the performance of a fuzzy partition. In this paper, a study of the influence of replacing the fuzzy partition used in the ICFA algorithm with the possibilistic and the fuzzy-possibilistic partitions will be done. A comparative analysis of each kind of partition will be performed in order to see if the possibilistic approach can improve the performance of the ICFA algorithm both in normal and in noise conditions. The results will show how the employment of a mixed approach combining fuzzy and possibilistic approach can lead to improve the results when designing RBFNNs.

This paper presents a novel learning methodology for multigrid-based fuzzy system (MGFS), and its application to the CATS time series prediction benchmark. The MGFS model keeps the advantages of the traditional grid-based fuzzy systems (GBFS), and overcomes the problem inherent to all GBFSs when dealing with high dimensional input data. Thus the MGFS model keeps interpretability, low computational cost and high generalization. A novel architecture selection algorithm for MGFSs that allows performing input variable selection is proposed. It identifies the sub-optimal architecture, according to a provided data set of input/output data. The architecture selection algorithm is completed with a structure identification procedure, used to obtain the optimal input space partitioning of the different sub-grids of the model. The complete algorithm is used to obtain the MGFS models for the CATS series prediction problem, solved using a direct prediction-based approach.

The use of Radial Basis Function Neural Networks (RBFNNs) to solve functional approximation problems has been addressed many times in the literature. When designing an RBFNN to approximate a function, the first step consists of the initialization of the centers of the RBFs. This initialization task is very important because the rest of the steps are based on the positions of the centers. Many clustering techniques have been applied for this purpose achieving good results although they were constrained to the clustering problem. The next step of the design of an RBFNN, which is also very important, is the initialization of the radii for each RBF. There are few heuristics that are used for this problem and none of them use the information provided by the output of the function, but only the centers or the input vectors positions are considered. In this paper, a new algorithm to initialize the centers and the radii of an RBFNN is proposed. This algorithm uses the perspective of activation grades for each neuron, placing the centers according to the output of the target function. The radii are initialized using the center's positions and their activation grades so the calculation of the radii also uses the information provided by the output of the target function. As the experiments show, the performance of the new algorithm outperforms other algorithms previously used for this problem.

Typically, Takagi-Sugeno-Kang (TSK) fuzzy rules have been used as a powerful tool for function approximation problems, since they have the capability of explaining complex relations among variables using rule consequents that are functions of the input variables. But they present the great drawback of the lack of interpretability, which makes them not to be so suitable for a wide range of problems where interpretability of the obtained model is a fundamental key. In this paper, we present a novel approach that extends the work by Bikdash (IEEE Trans. Fuzzy Systems 7 (6) (1999) 686-696), in order to obtain an interpretable and accurate model for function approximation from a set of I/O data samples, which make use of the Taylor Series Expansion of a function around a point to approximate the function using a low number of rules. Our approach also provides an automatic methodology for obtaining the optimum structure of our Taylor series-based (TaSe) fuzzy system as well as its pseudo-optimal rule-parameters (both antecedents and consequents).

A novel approach to achieve real-time global learning in fuzzy controllers is proposed. Both the rule consequents and the membership functions defined in the premises of the fuzzy rules are tuned using a one-step algorithm, which is capable of controlling nonlinear plants with no prior offline training. Direct control is achieved by means of two auxiliary systems: The first one is responsible for adapting the consequents of the main controller's rules to minimize the error arising at the plant output, while the second auxiliary system compiles real input-output data obtained from the plant. The system then learns in real time from these data taking into account, not the current state of the plant but rather the global identification performed. Simulation results show that this approach leads to an enhanced control policy thanks to the global learning performed, avoiding overfitting.

This paper presents a multiobjective evolutionary algorithm to optimize radial basis function neural networks (RBFNNs) in order to approach target functions from a set of input-output pairs. The procedure allows the application of heuristics to improve the solution of the problem at hand by including some new genetic operators in the evolutionary process. These new operators are based on two well-known matrix transformations: singular value decomposition (SVD) and orthogonal least squares (OLS), which have been used to define new mutation operators that produce local or global modifications in the radial basis functions (RBFs) of the networks (the individuals in the population in the evolutionary procedure). After analyzing the efficiency of the different operators, we have shown that the global mutation operators yield an improved procedure to adjust the parameters of the RBFNNs.

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.

The identification of a model is one of the key issues in the field of fuzzy system modeling and function approximation theory. There are numerous approaches to the issue of parameter optimization within a fixed fuzzy system structure but no reliable method to obtain the optimal topology of the fuzzy system from a set of input-output data. This paper presents a reliable method to obtain the structure of a complete rule-based fuzzy system for a specific approximation accuracy of the training data, i.e., it can decide which input variables must be taken into account in the fuzzy system and how many membership functions (MFs) are needed in every selected input variable in order to reach the approximation target with the minimum number of parameters.

In this paper, a novel approach is presented to fine tune a direct fuzzy controller based on very limited information on the nonlinear plant to be controlled. Without any off-line pretraining, the algorithm achieves very high control performance through a two-stage algorithm. In the first stage, coarse tuning of the fuzzy rules (both rule consequents and membership functions of the premises) is accomplished using the sign of the dependency of the plant output with respect to the control signal and an overall analysis of the main operating regions. In stage two, fine tuning of the fuzzy rules is achieved based on the controller output error using a gradient-based method. The enhanced features of the proposed algorithm are demonstrated by various simulation examples.