Hybrid systems, which combine statistical and machine learning (ML) techniques using residual (error forecasting) modeling, have been highlighted in the literature due to their accuracy and ability to forecast time series with different characteristics. In these architectures, a crucial task is the proper modeling of the residuals since they may present random fluctuations, complex nonlinear patterns, and heteroscedastic behavior. Hence, the selection, specification, and training of one ML model to forecast the residuals are costly and challenging tasks since issues, such as underfitting, overfitting, and misspecification, can lead to a system with low accuracy or even deteriorate the linear forecast of the time series. This article proposes a hybrid system, named dynamic residual forecasting (DReF), that employs a modified dynamic selection (DS) algorithm to decide: the most suitable ML model to forecast a pattern of the residual series and if it is a promising candidate to increase the accuracy of the time series forecast from the linear combination. Thus, the DReF aims to reduce the uncertainty of the ML model selection and avoid the deterioration of the time series forecast. Furthermore, the proposed system searches for the most suitable parameters of the DS algorithm for each data set. In this article, the proposed method uses a pool of five ML models widely adopted in the literature: multilayer perceptron, support vector regression, radial basis function, long short-term memory, and convolutional neural network. An experimental evaluation was conducted using ten well-known time series. The results show that the DReF obtains superior results for the majority of the data sets compared with single and hybrid models of the literature.

Deep Convolutional Neural Networks (CNNs) have reached the attention in the last decade due to their successful application to many computer vision domains. Several handcrafted architectures have been proposed in the literature, with increasing depth and millions of parameters. However, the optimal architecture size and parameters setup are dataset-dependent and challenging to find. For addressing this problem, this work proposes a Multi-Objective Grammatical Evolution framework to automatically generate suitable CNN architectures (layers and parameters) for a given classification problem. For this, a Context-free Grammar is developed, representing the search space of possible CNN architectures. The proposed method seeks to find suitable network architectures considering two objectives: accuracy and F1-score. We evaluated our method on CIFAR-10, and the results obtained show that our method generates simpler CNN architectures and overcomes the results achieved by larger (more complex) state-of-the-art CNN approaches and other grammars.

Selecting the best architecture for a Deep Neural Network (DNN) is a non-trivial task since there is a massive amount of possible configurations (layers and parameters) and great difficulty in how to choose them. To make this task more independent of human interaction, this work addresses the DNN architecture selection problem as a multi-objective optimization task with different criteria in a combinatorial context. For this, we defined a new way to represent the architecture of DNN (layer sequence) as a solution in the optimization process. The proposed method attempts to find the best composition and sequence of layers for the DNN architecture satisfying two criteria: accuracy and F1score. The method was evaluated for performance and compared to the exhaustive and random approaches and state-of-the-art DNN algorithms. The results obtained showed that the proposed method is capable of achieving results close to the optimum, and competitive when compared to those results reached by state of the art algorithms.

Dynamic predictor selection has been applied to time series context to improve the accuracy to forecast. A crucial step in dynamic selection methods if the definition of the region of competence, which is composed of the most similar patterns to a test pattern, because the predictor that attains the best performance in this region is selected to forecast this test pattern. The performance of dynamic selection methods depends on two main parameters, the size of the region of competence and the similarity measure (also called of distance measure). This work evaluates the influence of these parameters on six real-world time series to forecasting one step. In the experiments, Bagging is adopted to generate a pool of predictors, where the best predictor is selected per query pattern based on its performance on the region of competence. The results show that the choice of an appropriate distance measure, as well as the size of the region of competence, is mandatory to boost the performance of the prediction system. Moreover, the results reinforce the importance of using a dynamic selection approach to improve forecasting accuracy when compared to the monolithic models, also called of single models.

A desired characteristic of the automotive diagnostics systems is to make fault predictions to prevent unexpected car breakdowns, avoiding financial losses and physical damages to the drivers. Based on that, the objective of this work is to evaluate intelligent hybrid systems to forecast real-time information from three in-vehicle sensors: engine coolant temperature, air fuel ratio (AFR) internal combustion and automobile battery voltage. Numerical results showed that, in general, combining forecasters from the residual modeling is a promising approach in the context of automotive data. In addition, the alternative combination of nonlinear with linear models suggests a hopeful proposition that can be used in other applications.

In time series analysis and forecasting, machine learning (ML) models have been widely used due to their flexibility and accuracy. However, the tuning process of their parameters is a hard task, mainly when complex time series are addressed. So, it is difficult to guarantee the optimal adjustment of the ML model parameters. This paper proposes a recursive approach based on the Perturbation theory to correct the forecasting of ML models. From the initial forecasting given by an ML model, a new ML model is trained using the error series (the difference between the actual series and forecasting) of the first model to decrease the overall error of the system. This process can be recursively repeated until convergence or some stop criterion. The response of the perturbative approach is composed of the sum of the predictions (perturbations) of the ML models trained in each recursion. The proposed approach is investigated with four ML models: Support Vector Regression, Multilayer Perceptron, Long Short-Term Memory, and Radial Basis Function network. The evaluation is performed with an experimental investigation conducted on four time series: Canadian Lynx, Sunspot, Star Brightness, and S&P500 index. The results show that the perturbative approach improves significantly the accuracy of all evaluated ML models.

This paper describes a novel GMM-UBM based system that deals with the session noise variability problem. The system uses the Type-2 Fuzzy GMM framework by considering the speaker GMM parameters to be uncertain in an interval. The parameters intervals are estimated using a multicondition model training on noisy speeches that are synthesized from the speaker's utterances. Experiments were conducted using the MIT Device Speaker Verification Corpus with utterances having the lowest noise level as training data. The result shows an improvement in the EER of 24.11% for the proposed method compared to the GMM-UBM when evaluated over the noisiest utterances. This shows that the method reduces the effects of the session variability.

The stock market is the most important institution for global investments all around the world. Among the possibles analysis, the study and forecasting of ultra-high-frequency time series is an interesting and great challenge to econometric modeling and statistical analysis due its complex behaviour. This work proposes a hybrid intelligent system to forecast ultra-high-frequency stock prices. The intelligent system is composed of a genetic algorithm (GA) that seeks the best parameters (the number of nodes in the input and hidden layers, and the training algorithm) of an artificial neural network (ANN) of type multiLayer perceptron (MLP). The proposed method called Time-delay Added Evolutionary Forecasting (TAEF) is a data-driven approach, that performs a pos-processing, where the objective is to reduce the difference between the forecasting and the actual series. The experimental study is performed using ultra-high frequency time series (Amazon, APPLE, Google and Intel stock prices) and shows that the proposed approach overcomes classical techniques of the computational intelligence and statistics in light of six relevant evaluation measures.

In this paper, a simulation of intelligent agents is developed to recreate the environment of negotiation of stock markets. The focus is analyze the behavior of movement/ fluctuation of stock markets. This movement can be captured by a measure called volatility, which is the difference between two stock prices in distinct periods. It characterizes the sensibility of a market change in the world economy. The contributions of this work are three-fold: (i) a simulation of dynamics of stock markets based in intelligent agents; (ii) based in this simulation an analysis of the volatility dynamic of the simulated time series; (iii) after that, a investigation about the relationship between the volatility of the markets, distribution of gain/loss money of agents and the coefficient of the exponential function based on the ideal gas theory of Maxwell-Boltzmann. This information can be used, for example, to predict the future behavior of the markets.

The time series forecasting is an useful application for many areas of knowledge such as biology, economics, climatology, biology, among others. A very important step for time series prediction is the correct selection of the past observations (lags). This paper uses a new algorithm based in swarm of particles to feature selection on time series, the algorithm used was Frankenstein's Particle Swarm Optimization (FPSO). Many forms of filters and wrappers were proposed to feature selection, but these approaches have their limitations in relation to properties of the data set, such as size and whether they are linear or not. Optimization algorithms, such as FPSO, make no assumption about the data and converge faster. Hence, the FPSO may to find a good set of lags for time series forecasting and produce most accurate forecastings. Two prediction models were used: Multilayer Perceptron neural network (MLP) and Support Vector Regression (SVR). The results show that the approach improved previous results and that the forecasting using SVR produced best results, moreover its showed that the feature selection with FPSO was better than the features selection with original Particle Swarm Optimization.

In this paper it is introduced a new perturbative approach for time series forecasting. The model uses the error of the series, that is the difference between real value of the series and the output of a predictive method, to improve the series forecasting. The methodology proposed is inspired in the Perturbation Theory, that consists in a set of approximation schemes used to describe a complicated problem in terms of simpler ones. For an experimental investigation, this theory, is combined with the TAEF method, that has interesting results when compared with the literature. This combination is called P-TAEF (Perturbative TAEF). Its results over some time series are discussed and compared with previous results found in the literature. It was used several performance measures that showed the robustness of the perturbative approach.

The Perturbative Time-delay Added Evolutionary Forecasting (P-TAEF) method for time series prediction is inspired in the Perturbation Theory, a concept already commonly used in other areas of science (physics, mathematics, etc), and evolutionary computing. This methodology is shown and an experimental investigation is conducted with some relevant time series and the results achieved are discussed and compared with previous results found in the literature.

Forecasting systems have been widely used for decision making and one of its most promising approaches is based on Artificial Neural Networks (ANN). In this paper, a hybrid swarm system is presented for the time series forecasting problem, which consists of an intelligent hybrid model composed of an ANN combined with Particle Swarm Optimizer (PSO). The proposed method searches the relevant time lags for a correct characterization of the time series, as well as the number of processing units in the hidden layer, the training algorithm and the modeling of ANN. The proposed method shows an efficient procedure to adjust the ANN parameters through the use of a particle swarm optimization mechanism. An experimental analysis is conducted with the proposed method using six real world time series and the results are discussed according to five performance measures.

Artificial Neural Networks (ANN) have been widely used in order to solve the time series forecasting problem. One of its most promising approaches is the combination with other intelligence techniques, as genetic algorithms, evolutionary strategies, etc. The efficiency of these technics, if used correctly, can be very high. Unfortunately, in terms of fitness function, there is still some lacks of experimental (and theoretical) results to help the practitioners to use these technics in order to find better predictions. This paper proposes others fitness functions (instead of conventional MSE based) and presents an experimental investigation of eight different fitness functions for time series prediction based on five well known measures of statistical performance in the literature. Using a hybrid method for tuning of the ANN structure and parameters (a modified genetic Algorithm), an analysis of the final results effects are made according with four relevant time series. This work shows that small changes of the fitness function evaluation can lead to a significantly improved performance.