Search

This paper investigates a number of computational intelligence techniques in the detection of heart disease. Particularly, comparison of six well known classifiers for the well used Cleveland data is performed. Further, this paper highlights the potential of an expert judgment based (i.e., medical knowledge driven) feature selection process (termed as MFS), and compare against the generally employed computational intelligence based feature selection mechanism. Also, this article recognizes that the publicly available Cleveland data becomes imbalanced when considering binary classification. Performance of classifiers, and also the potential of MFS are investigated considering this imbalanced data issue. The experimental results demonstrate that the use of MFS noticeably improved the performance, especially in terms of accuracy, for most of the classifiers considered and for majority of the datasets (generated by converting the Cleveland dataset for binary classification). MFS combined with the computerized feature selection process (CFS) has also been investigated and showed encouraging results particularly for NaiveBayes, IBK and SMO. In summary, the medical knowledge based feature selection method has shown promise for use in heart disease diagnostics.

In this paper we combine the main concepts of estimation of distribution algorithms (EDAs) and immune algorithms (IAs) to be a hybrid algorithm called immune-estimation of distribution algorithms (IEDA). Both EDAs and IAs are extended from genetic algorithms (GAs). EDAs eliminate the genetic operation including crossover and mutation from the GAs and places more emphasis on the relation between gene loci. It adopts the distribution of selected individuals in search space and models the probability distributions to generate the next population. However, the primary gap of EDAs is lock of diversity between individuals. Hence, we introduce the IAs that is a new branch in computational intelligence. The main concepts of IAs are suppression and hypermutation that make the individuals be more diversity. Moreover, the primary gap of IAs is to pay no attention to the relation between individuals. Therefore, we combine the main concepts of two algorithms to improve the gaps each other. The classification risk of data mining is applied by this paper and compares the results between IEDA and general GAs in the experiments. We adopt the thyroid gland data set from UCI databases. Based on the obtained results, our research absolute is better than general GAs including accuracy, type I error and type II error. The results show not only the excellence of accuracy but also the robustness of the proposed algorithm. In this paper we have got high quality results which can be used as reference for hospital decision making and research workers.

Multi-agent modeling has emerged as a promising discipline for dealing with decision-making processes in distributed information system applications. One such application is the modeling of distributed manufacturing networks which can link up various manufacturing firms to form a “virtual” manufacturing consortium on a global basis. This paper presents a multi-agent model that can be deployed in a dispersed manufacturing network involving companies with different core competence. The multi-agent model plays the important role of monitoring the information flow and task allocation among the network companies by means of intelligent agents, which are software programs designed to accomplish specific tasks like “real” human agents with specialized skills. This proposed model is equipped with various computational intelligence technologies for various purposes, including rule-based reasoning for decomposing jobs into fundamental tasks and object-oriented technology for the design and creation of the intelligent agents, together with a genetic algorithm scheme to optimize the sequence of tasks to be carried out by various agents in order to achieve minimum overhead incurred during the operations across the dispersed network. In this paper, the detailed infrastructure of a multi-agent model and how it can be deployed to enhance the performance of a dispersed manufacturing network are presented.

This paper describes an approach using computational intelligence methods to form a hybrid model as a classification method for 2-lead cardiac arrhythmias. The hybridization of methods can increase the performance in a system and take advantage of the benefits offered by such techniques in solving complex problems. The interpretation of electrocardiograms is a useful task for physicians, but when it comes to reviewing more than 24 h of information, it becomes a laborious task for them. For this reason, the design a computational model that helps in such a task is very useful for the timely medical diagnosis. The hybrid model is build using artificial neural networks and fuzzy logic. Training and testing of the hybrid model was with the Massachusetts Institute of Technology and Beth Israel Hospital (MIT-BIH) arrhythmia database. The heartbeats are preprocessed to improve results of classification. Ten different classes of normal and arrhythmia signals for building the hybrid model are considered. We used two electrode signals or leads included in the MIT-BIH arrhythmia database, MLII and V1, V2, or V3 as second electrode signal. The hybrid model is composed by two basic module units, as described below. A basic module unit to perform the classification for each signal lead is used. Each basic module unit is composed of three different classifiers based on the following models: fuzzy KNN algorithm, multilayer perceptron with gradient descent and momentum (MLP-GDM), and multilayer perceptron with scaled conjugate gradient backpropagation (MLP-SCG). The outputs from the classifiers are combined using a fuzzy system for integration of results. We designed two fuzzy systems, Mamdani type-1 fuzzy system (type-1 FIS) and an interval type-2 fuzzy system (IT2FIS). The reason is to perform a comparison between type-1 FIS and IT2FIS in the hybrid model. We have obtained best results in the classification rate using IT2FIS instead of type-1 FIS in the basic units. Finally, a type-1 FIS is used to determine the global classification for the 2 basic units in hybrid model. We obtained a good classification rate in each basic module unit, 92.90% and 92.70% of classification rate for basic modules unit 1 and unit 2 respectively. Finally, we obtained a 93.80% when used type-1 FIS and 94.20% of classification rate used IT2FIS combining both basic module units. In the results presented, we improve the global classification in proposed hybrid model combining neural networks and fuzzy logic used both signal lead included in MIT-BIH arrhythmia database. The proposed hybrid model maybe extended to use multi-lead arrhythmia classification using other databases that contain 12 leads to be able to make a complete medical diagnosis.

With the recent development of computational intelligence (CI), data-driven models have gained growing interest to be applied in various scientific disciplines. This paper aims at proposing a hybrid CI-based data-driven framework as a complement for the physics-based models used in climatological prediction. The proposed framework, called FB-STEP, is based on a combination of fuzzy Bayesian strategy and multifractal analysis technique. The focus is to address three major research challenges in multivariate climatological prediction: (1) modeling complex spatio-temporal dependency among climatological variables, (2) dealing with non-linear, chaotic dynamics in climatic time series, and (3) reducing epistemic uncertainty in the data-driven prediction process. The present work not only explores Fuzzy-Bayesian modeling of spatio-temporal processes, but also presents an elegant approach of dealing with intrinsic chaos in time series, through a synergism between multifractal analysis and Bayesian inference mechanism. Similar concepts may also be successfully employed in developing expert or intelligent systems for wide range of applications, including reservoir-water dynamics modeling, flood monitoring, traffic flow modeling, chemical-mechanical process monitoring, and so on. Thus, the present research work carries a significant value not merely in the field of climate research, but also in the domains of AI and machine intelligence. The experimentation has been carried out to spatio-temporally extrapolate the climatic conditions of five different locations in India, with the help of historical data on temperature, humidity, precipitation rate, and soil moisture. A comparative study with popular linear and non-linear methods has validated the efficacy of the proposed data-driven approach for climatological prediction.

The increase in mortality rate due to life-threatening diseases has become an issue of concern in today’s world. Early detection and diagnosis of diseases thus becomes necessary to reduce the severity of their side effects. Computational intelligence techniques like rare association rule mining can be extensively used for the analysis of diseases. This paper introduces an efficient technique to identify the symptoms and risk factors for three adverse diseases: cardiovascular disease, hepatitis and breast cancer, in terms of rare association rules. Existing research on rare association rule mining is based on the notion that the entire data to be operated on is available at the onset of the mining process. The medical databases in practice may get modified over time due to the addition of new records or deletion of previous records. Moreover, the user may switch to a new threshold for generating the desired set of rare association rules when the database gets updated. A straightforward yet incompetent solution for generating the current set of rare association rules would be to re-execute the entire mining algorithm from scratch, for each modified bunch of data and updated threshold. The algorithm proposed in this study is capable of generating the new set of rare association rules from updated medical databases in a single database scan without re-executing the entire mining process. It can efficiently handle the cases of transaction insertion and deletion and also provides flexibility to the user to generate the new set of rare association rules when threshold is updated. Experimental analysis illustrates the significance of proposed approach over traditional approach of repeatedly mining the entire updated database.

Glaucoma is an asymptomatic disease caused by damage to the optic nerve due to elevated intraocular pressure. With early diagnosis, the chances of controlling progression are greater. Glaucoma is a major global health problem as a cause of blindness, second only to cataracts. This paper presents the development of a method for the automatic detection of glaucoma in retinal images using a deep learning approach together with the exploration of texture attributes through phylogenetic diversity indexes. The methodology employed is as follows: First, image acquisition is done from the RIM-ONE, DRIONS-DB, and DRISHTI-GS databases, followed by training the convolutional neural network for optical disk segmentation. After this segmentation, it is necessary to perform removal of the blood vessels after which feature extraction is applied to the images generated from the RGB channels and the gray levels. The extracted attributes are based only on the features of texture using phylogenetic diversity indexes. Classification was performed using an approach based on the convolutional neural network. The method proved to be promising, achieving accuracy, sensitivity, and specificity of 100%. The medical expertise developed by the proposed method proved to be robust and efficient for the automatic detection task of glaucoma through retinal image analysis, using image processing and computational intelligence techniques. We thus propose a specialist system through the construction of a complete CAD tool to integrate into real clinical environments, providing a second opinion on the diagnosis of glaucoma quickly, efficiently, and at low cost.

In this paper, a new strategy for dealing with time series exhibiting a seasonal pattern is proposed. The strategy is applied in the context of time series forecasting using kNN regression. The key idea is to forecast every different season using a different specialized kNN learner. Each learner is specialized because its training set only contains examples whose targets belong to the season that is able to forecast. This way, the forecast of a specialized kNN learner is an aggregation of target values of the same season, reducing the likelihood of misleading forecasts. Although the strategy is applied to kNN, we think that other computational intelligence approaches could take advantage of it.

Identifying students’ learning styles has several benefits such as making students aware of their strengths and weaknesses when it comes to learning and the possibility to personalize their learning environment to their learning styles. While there exist learning style questionnaires for identifying a student's learning style, such questionnaires have several disadvantages and therefore, research has been conducted on automatically identifying learning styles from students’ behavior in a learning environment. Current approaches to automatically identify learning styles have an average precision between 66% and 77%, which shows the need for improvements in order to use such automatic approaches reliably in learning environments. In this paper, four computational intelligence algorithms (artificial neural network, genetic algorithm, ant colony system and particle swarm optimization) have been investigated with respect to their potential to improve the precision of automatic learning style identification. Each algorithm was evaluated with data from 75 students. The artificial neural network shows the most promising results with an average precision of 80.7%, followed by particle swarm optimization with an average precision of 79.1%. Improving the precision of automatic learning style identification allows more students to benefit from more accurate information about their learning styles as well as more accurate personalization towards accommodating their learning styles in a learning environment. Furthermore, teachers can have a better understanding of their students and be able to provide more appropriate interventions.

Computational Intelligence (CI) can advance qualitative analysis of human motion.CI can provide logical validity for common sense coaching rules from motion data.Deviations from individual swing plane influence ball impact and flight pattern.Amateur golfers tend to slice/fade/pull over 83% producing 'Outside-In' ball impact with the driver.Motion data export solution was implemented via clipboard and text processing. Quantifying and validating descriptive heuristic rules that govern someone's skills and expertise have been a known philosophical quest since the early Greek philosophers. Inherent to sport coaching is the qualitative assessment of complex human motion patterns, relying on subjective and 'hard-to-quantify' criteria that can be subject to experts/coaches disagreement. This paper presents an application of Artificial Neural Networks (ANN) for the discovery of predictive power of swing plane heuristic rules influencing golf ball trajectories. The golf data set (531 samples from 14 golfers) utilised in the experiments was captured via a ubiquitous computing device embedded in the handle of a driver club. Out of multiple swing performance factors influencing ball trajectory, the selected subset of features for subspace modelling was linked only to the swing plane concept. Quantitative evidence supporting empirical coaching rules for swing plane assessment were obtained by supervised learning of ANN models. Optimised ANN models Radial Basis Function (RBF) and Support Vector Machine (SVM), were able to draw inference from captured swing data linking ball trajectories with variations of swing plane (with overall classification of 87%). The obtained swing plane computer model inference, data analysis and implemented concept of generic data export utility support kinesiology, golf coaching, inform club fitting, golf manufacturing technology and demonstrate new cross- and multi-disciplinary integration of sport science, augmented coaching, ubiquitous computing, computational intelligence and the applications of expert systems for growing availability of sport, injury prevention/rehabilitation and golf related data sets.

This paper is motivated by decommissioning studies in the field of oil and gas, which comprise a very large number of installations and are of interest to a large number of stakeholders. Generally, the problem gives rise to complicated multi-criteria decision aid tools that rely upon the costly evaluation of multiple criteria for every piece of equipment. We propose the use of machine learning techniques to reduce the number of criteria by feature selection, thereby reducing the number of required evaluations and producing a simplified decision aid tool with no sacrifice in performance. In addition, we also propose the use of machine learning to explore the patterns of the multi-criteria decision aid tool in a training set. Hence, we predict the outcome of the analysis for the remaining pieces of equipment, effectively replacing the multi-criteria analysis by the computational intelligence acquired from running it in the training set. Computational experiments illustrate the effectiveness of the proposed approach.

Cloud service composition builds new value-added services by combining existing single services. However, because of the exuberant growth of cloud services and the varying quality of service (QoS), discovering required services and creating a service composition with certain quality guarantees becomes a significant technical issue and attracts much concern. Computational intelligence techniques are considered to be effective in solving such problems, and researchers have made substantial efforts in this area. Nevertheless, to the best of our knowledge, there is not any systematic research about this issue with a particular focus on computing intelligence. Thus, the current study aims to create a panoramic view of QoS-aware cloud service composition from the perspective of computational intelligence. The objectives of this paper are to (1) investigate the relevant studies on this field; (2) make a comprehensive examination of the literature from different aspects: active researchers, research motivations, QoS parameters, algorithms, datasets, optimization strategies and could layers; (3) identify the areas which need further research. For this, a search protocol has been well defined, and 105 articles from 2009 to 2018 were selected. This study classified these articles into three groups, including non-heuristic, heuristic and meta-heuristic, and then examined the research works from several aspects. The results indicate that reducing response time is the most important motivation for researchers, and meta-heuristic algorithms, especially genetic algorithms, are the most widely used computational intelligence techniques. Besides, the most widely used QoS attributes and datasets are also revealed. Additionally, the study points out there are still some research challenges in this area, such as QoS evaluation in a dynamic environment and inter-service correlations. In general, this study classified and compared the existing computational intelligence techniques; analyzed the research status and identified future research directions. It can provide a basis for both researchers and practitioners who are interested in this area. More significantly, in the field of expert and intelligent systems, this study can assist in the design and development of expert and intelligent systems in enterprises, it can efficiently assist enterprises in business decisions and risk reduction.

The paper explores the usefulness of hybridizing two distinct nature inspired computational intelligence techniques viz., Artificial Neural Networks (ANN) and Genetic Algorithms (GA) for modeling slump of Ready Mix Concrete (RMC) based on its design mix constituents viz., cement, fly ash, sand, coarse aggregates, admixture and water-binder ratio. The methodology utilizes the universal function approximation ability of ANN for imbibing the subtle relationships between the input and output variables and the stochastic search ability of GA for evolving the initial optimal weights and biases of the ANN to minimize the probability of neural network getting trapped at local minima and slowly converging to global optimum. The performance of hybrid model (ANN-GA) was compared with commonly used back-propagation neural network (BPNN) using six different statistical parameters. The study showed that by hybridizing ANN with GA, the convergence speed of ANN and its accuracy of prediction can be improved. The trained hybrid model can be used for predicting slump of concrete for a given concrete design mix in quick time without performing multiple trials with different design mix proportions.

A methodology based on GFS and OVO in the framework of IDS is proposed.Linguistic labels enable smoother borderline, and allows higher interpretability.Divide-and-conquer learning scheme, improves precision for rare attack events.Several metrics of performance show the goodness of this approach on KDDCUP'99.Our results excels the state-of-the-art of GFS for IDS and C4.5 decision tree. Security policies of information systems and networks are designed for maintaining the integrity of both the confidentiality and availability of the data for their trusted users. However, a number of malicious users analyze the vulnerabilities of these systems in order to gain unauthorized access or to compromise the quality of service. For this reason, Intrusion Detection Systems have been designed in order to monitor the system and trigger alerts whenever they found a suspicious event.Optimal Intrusion Detection Systems are those that achieve a high attack detection rate together with a small number of false alarms. However, cyber attacks present many different characteristics which make them hard to be properly identified by simple statistical methods. According to this fact, Data Mining techniques, and especially those based in Computational Intelligence, have been used for implementing robust and accuracy Intrusion Detection Systems.In this paper, we consider the use of Genetic Fuzzy Systems within a pairwise learning framework for the development of such a system. The advantages of using this approach are twofold: first, the use of fuzzy sets, and especially linguistic labels, enables a smoother borderline between the concepts, and allows a higher interpretability of the rule set. Second, the divide-and-conquer learning scheme, in which we contrast all possible pair of classes with aims, improves the precision for the rare attack events, as it obtains a better separability between a "normal activity" and the different attack types.The goodness of our methodology is supported by means of a complete experimental study, in which we contrast the quality of our results versus the state-of-the-art of Genetic Fuzzy Systems for intrusion detection and the C4.5 decision tree.

In car racing, blocking refers to maneuvers that can prevent, disturb or completely block an overtaking action by an incoming car. In this paper, we present an advanced overtaking behavior that is able to deal with opponents implementing blocking strategies of various difficulty level. The behavior we developed has been integrated in an existing fuzzy-based architecture for driving simulated cars and tested using The Open Car Racing Simulator (TORCS). We compared a driver implementing our overtaking strategy against four of the bots available in the TORCS distribution and Simplix, a state-of-the-art driver which won several competitions. The comparison was carried out against opponents implementing three blocking strategies of increasing difficulty and two different scenarios: (i) a basic scenario with one opponent on a straight stretch to overtake as quickly as possible; (ii) an advanced scenario involving a race on a non-trivial track against several opponents. The results from the basic scenario show that our strategy can always overtake the opponent car; in particular, our strategy is slightly more risky than the other ones and may result in a little damage, however, all the other controllers show a more careful and safe policy that often prevents them to complete an overtaking maneuver. When racing against several opponents on complex tracks, our strategy results in the best trade-off between the time spent being blocked by an opponent ahead and the number of overtaking maneuvers completed.

Accurate time series forecasting is a key issue to support individual and or- ganizational decision making. In this paper, we introduce novel methods for multi-step seasonal time series forecasting. All the presented methods stem from computational intelligence techniques: evolutionary artificial neu- ral networks, support vector machines and genuine linguistic fuzzy rules. Performance of the suggested methods is experimentally justified on sea- sonal time series from distinct domains on three forecasting horizons. The most important contribution is the introduction of a new hybrid combination using linguistic fuzzy rules and the other computational intelligence methods. This hybrid combination presents competitive forecasts, when compared with the popular ARIMA method. Moreover, such hybrid model is more easy to interpret by decision-makers when modeling trended series., The research was supported by the European Regional Development Fund in the IT4Innovations Centre of Excellence project (CZ.1.05/1.1.00/02.0070). Furthermore, we gratefully acknowledge partial support of the project KON- TAKT II - LH12229 of MSˇMT CˇR.

The objective of this paper was to perform a comparative analysis of the computational intelligence algorithms to identify breast cancer in its early stages. Two types of data representations were considered: microarray based and medical imaging based. In contrast to previous researches, this research also considered the imbalanced nature of these data. It was observed that the SMO algorithm performed better for the majority of the test data, especially for microarray based data when accuracy was used as performance measure. Considering the imbalanced characteristic of the data, the Naive Bayes algorithm was seen to perform highly in terms of true positive rate (TPR). Regarding the influence of SMOTE, a well-known imbalanced data classification technique, it was observed that there was a notable performance improvement for J48, while the performance of SMO remained comparable for the majority of the datasets. Overall, the results indicated SMO as the most potential candidate for the microarray and image dataset considered in this research.

In this paper, two CI techniques, namely, single multiplicative neuron (SMN) model and adaptive neuro-fuzzy inference system (ANFIS), have been proposed for time series prediction. A variation of particle swarm optimization (PSO) with co-operative sub-swarms, called COPSO, has been used for estimation of SMN model parameters leading to COPSO-SMN. The prediction effectiveness of COPSO-SMN and ANFIS has been illustrated using commonly used nonlinear, non-stationary and chaotic benchmark datasets of Mackey-Glass, Box-Jenkins and biomedical signals of electroencephalogram (EEG). The training and test performances of both hybrid CI techniques have been compared for these datasets.

A novel hybrid optimization algorithm combining computational intelligence techniques is presented to solve the multifactor highway passenger volume prediction problem. In this paper, we can get and discretize a reduced decision table, which implies that the number of evaluation criteria such as travel quantity, fixed-asset investment, railway mileage, and waterway passenger volume are reduced with no information loss through rough set theory (RST) method. Particle swarm optimization (PSO) algorithm based on the random global optimization is inducted into the network training. The PSO algorithm is used for glancing study in order to confirm the initial values, and then the back propagation neural network (BPNN) is used for given accuracy to found the PSO-BPNN model. And this reduced information is used to form a classification rule set, which is regarded as an appropriate input parameter to training PSO-BPNN model. The RST-PSO-BPNN model is obtained to forecast highway passenger volume. The rules developed by RST analysis show the best prediction accuracy if a case matches any one of the rules. The keystone of this hybrid optimization algorithm is using rules developed by RST for an object that matches any one of the rules and the PSO-BPNN model for one that does not match any of them. The effectiveness of our optimization algorithm was verified by experiments comparing the traditional gray model method. For the experiment, highway passenger volumes of China during the period 1995-2009 were selected, and for the validation, the novel hybrid optimization algorithm is reliable.

This paper presents a methodology that enables fault detection in dynamic systems based on recent immune theory. The fault detection is a challenging problem due to increasing complexity of processes and agility necessary to avoid malfunction or accidents. The fault detection central challenge is determining the difference between normal and potential harmful activities at dynamic systems. A promising solution is emerging in the form of Artificial Immune Systems (AIS). The Danger Model (DM) proposes that the immune system reacts not against self or non-self but by threats generated into the organism: the danger signals. DM-based fault detection system proposes a new formulation for a fault detection system. A DM-inspired methodology is applied to a fault detection benchmark provided by DAMADICS to compare its relative performance to others algorithms. The results show that the strategy developed is promising for incipient and abrupt fault detection in dynamic systems.

The process of combining multiple computational intelligence techniques to build a single hybrid model has become increasingly popular. As reported in the literature, the performance indices of these hybrid models have proved to be better than the individual components when used alone. Hybrid models are extremely useful for reservoir characterization in petroleum engineering, which requires high-accuracy predictions for efficient exploration and management of oil and gas resources. In this paper, we have utilized the capabilities of data mining and computational intelligence in the prediction of porosity and permeability, two important petroleum reservoir characteristics, based on the hybridization of Fuzzy Logic, Support Vector Machines, and Functional Networks, using several real-life well-logs. Two hybrid models have been built. In both, Functional Networks were used to select the best of the predictor variables for training directly from input data by using its functional approximation capability with least square fitting algorithm. In the first model (FFS), the selected predictor variables were passed to Type-2 Fuzzy Logic System to handle uncertainties and extract inference rules, while Support Vector Machines made the final predictions. In the second, the selected predictor variables were passed to Support Vector Machines for training by transforming them to a higher dimensional space, and then to Type-2 Fuzzy Logic to handle uncertainties, extract inference rules and make final predictions. The simulation results show that the hybrid models perform better than the individual techniques when used alone for the same datasets with their higher correlation coefficients. In terms of execution time, the hybrid models took less time for both training and testing than the Type-2 Fuzzy Logic, but more time than Functional Networks and Support Vector Machines. This could be the price for having a better and more robust model. The hybrid models also performed better than a combination of two of the individual components, Type-2 Fuzzy Logic and Support Vector Machines, in terms of higher correlation coefficients as well as lower execution times. This is due to the effective role of Functional Networks, as a best-variable selector in the hybrid models.

This paper proposes a novel computational intelligence technique, based on the sociological concept of human group formation, with the aim to acquire a better solution to classification problems. The key concept of the human group formation is about the behavior of in-group members that try to unite with their own group as much as possible, and at the same time maintain social distance from the out-group members. This study compares the performance of the proposed model with that of fuzzy ARTMAP, radial basis function network, and learning vector quantization. Experimental results demonstrate the potential of the proposed approach in offering an efficient and effective solution to the problem.

Online banking and e-commerce have been experiencing rapid growth over the past few years and show tremendous promise of growth even in the future. This has made it easier for fraudsters to indulge in new and abstruse ways of committing credit card fraud over the Internet. This paper focuses on real-time fraud detection and presents a new and innovative approach in understanding spending patterns to decipher potential fraud cases. It makes use of self-organization map to decipher, filter and analyze customer behavior for detection of fraud.

Artificial neural networks and fuzzy systems, have gradually established themselves as a popular tool in approximating complicated nonlinear systems and time series forecasting. This paper investigates the hypothesis that the nonlinear mathematical models of multilayer perceptron and radial basis function neural networks and the Takagi-Sugeno (TS) fuzzy system are able to provide a more accurate out-of-sample forecast than the traditional auto regressive moving average (ARMA) and ARMA generalized auto regressive conditional heteroskedasticity (ARMA-GARCH) linear models. Using series of Brazilian exchange rate (R$/US$) returns with 15min, 60min and 120min, daily and weekly basis, the one-step-ahead forecast performance is compared. Results indicate that forecast performance is strongly related to the series' frequency and the forecasting evaluation shows that nonlinear models perform better than their linear counterparts. In the trade strategy based on forecasts, nonlinear models achieve higher returns when compared to a buy-and-hold strategy and to the linear models.