The knapsack problem is among one the most well-known and studied optimization problems. Its potential applications make it a good model for a real-life problem. In this work, the Knapsack Problem with Forfeits will be addressed. In this variant, pairs of conflicting items in the solution imply a penalty. Given a set of items and a conflict graph, the objective is to find a collection of items that respect the knapsack's capacity and maximizes the total value of the items minus the penalties. This problem can be applied from the organization of the workforce on the shop floor to investment decision problems. This work proposes the construction of a new method for the problem using well-established tools, the ILS meta-heuristic, and the VND local search heuristic, and there is also the possibility of verifying other heuristics and meta-heuristics. Preliminary results have already been obtained, showing that the new method surpasses others approaches. It can be expected that this work offers an indication of the path that can be taken to refine the resolution of the problem, presenting analyzes and comparisons between methods applied to the problem.

Vehicles have more and more built-in sensors, such sensors are interconnected in an internal network called CAN and their values can be accessed through the OBD-II interface. This allows a large amount of data from different variables from the act of driving.
Several works have proposed applications that benefit from the availability of these data. Most applications fall into one of the following problems: classification, grouping, prediction. In general, the works perform the following process: data extraction, cleaning and transformation data, training model, evaluation model.
In this work it is proposed to extract measures from Information Theory to add data to pro- cess. With this, it is intended to have new values to add or replace the data pre-processing and then evaluate the performance of the model. For evaluation, it is intended to analyze two applications: driver identification (classification problem) and pollutant gas emission prediction (prediction problem).
Preliminary results were obtained for driver identification. An experiment was done for a small and a large amount of data. For the small amount of data, the process according to the literature proved to be superior in most classification algorithms in relation to the proposed process. For the large amount of data, the difference between the processes became small, however with the process according to the literature being slightly superior with the exception of the Naive Bayes algorithm (which performed better with the proposed process, but had lower accuracy than the other classifiers).
The next step is to compare the proposed process and the one commonly used in the litera- ture for a prediction problem: prediction of polluting gas emissions.

Brazil’s and the world’s energy matrices are in an energy generation decentralization process due to the Distributed Generation (DG) expansion resulting in the energy generated through Photovoltaic Panels (PP) installed on the roofs of residences in urban areas. This change has made the energy quality monitoring for the Low Voltage Distribution Networks (LVDN) one of the Distribution System Operators (DSO) biggest challenges, who are only aware of the energy quality levels through the customers’ complaints to the distributors. There are few projects implementing systems for this purpose, possibly due to the high investment required for this technology. Therefore, this research proposes a cyber-physical system for monitoring the direction of electric current flow and voltage imbalances in an RDBT using the street lighting infrastructure. The proposed system will provide the DSO’s with real-time information on the power quality in the RDBT through a low-cost data acquisition model using the street lighting infrastructure through Telemanagement Relays (TR). The system was empirically validated by monitoring a street where photovoltaic generation systems stochastically connected to RDBT inject power. The results show that the proposed system can identify the direction of the current in the RDBT and estimate the voltage unbalance with an average absolute error of 0.096% compared to the professional power analyzer. The results of this project can make relevant contributions to the challenge in power quality monitoring; for example, DSO’s can use the proposed system to access real-time information relative to the RDBT through a low-cost data acquisition model. It is expected that the results of this project would bring relevant contributions to the challenge of monitoring the levels of power quality, thus collaborating for a more orderly expansion of the distribution of energy production in Brazil.

The technological evolution of recent years, accelerated by the Covid-19 pandemic, has been responsible for a rapid and continuous paradigm shift in education, with the adoption of multimedia learning environments. During the learning process in these environments, the absorption of content by students decreases as the volume of information transmitted to them increases. In other words, there is a cognitive overload in one or both visual and verbal channels. Currently, there is a scarcity of studies that use Computer Vision and Data Science tools for the analysis of Cognitive Load. Tools of this nature would enable automated analysis of large volumes of videos and, consequently, the evaluation and generation of multimedia content that optimize student learning. In this work, a pilot study was conducted with a sample of 13 students from the School of Medicine at the Universidad de Atacama, Chile. Thus, a methodology was developed to extract and investigate correlations between visual characteristics of the students' faces and cognitive load. A database of facial videos of students watching multimedia-enhanced lectures on their computer screens was used. This video database was initially organized and preprocessed, followed by the application of Deep Learning models to extract visual points of interest from the face in each frame. Micro-actions were previously annotated by the researcher, and the resulting data were evaluated to identify relevant patterns related to cognitive load using unsupervised and supervised machine learning algorithms. In addition to addressing the main investigation of this research, the results of this study include a proof of concept for analyzing the correlation of facial expressions with individual exam scores, for further analysis of cognitive load in multimedia learning environments.

The prediction of cardiovascular events in patients diagnosed with Metabolic Syndrome (MS)
is a topic of great relevance to the field of Health in general and essential for Endocrinology.
This dissertation aims to recommend a Machine Learning (ML) model for estimating the risks
of cardiovascular events in patients with MS, exploring the markers of the Framingham Risk
Score (FRS) and MS. Methodologically, we used a logistic regression model and analyses with
decision tree, random forest, gradient boosting, support vector machine, and k-nearest neighbors
to test our central hypothesis that biomarkers (variables related to MS) have a positive, strong,
and significant impact on cardiovascular events in patients with MS. Technically, the research
was conducted through experiments performed in different scenarios. In the first scenario, an
algorithm was developed to assess cardiovascular risk in patients with and without MS. In
subsequent scenarios, patients with and without MS were analyzed, considering the markers of
MS and FRS as dependent variables, while the Metabolic Syndrome condition was adopted as an
independent variable. In the fifth scenario, an analysis was performed to select the most suitable
regression and classification model for predicting cardiovascular risk in a combined dataset of
heart diseases. In the sixth scenario, the developed model was based on the Framingham Global
Risk Score (FRS), incorporating the markers of MS into the experiments. Data were obtained
from the National Center for Health Statistics (NHANES) repository, a combined dataset of heart
diseases from the UCI Machine Learning Repository and the Kaggle platform. In summary, the
main findings of this dissertation are as follows: (1) In the first scenario, a percentage difference
of 81.74% was observed in the mean CVD risk between populations with and without Metabolic
Syndrome, showing a significant increase in cardiovascular risk in the population with MS; (2)
In subsequent scenarios, the Random Forest (RF) model stood out, achieving high accuracy
in all variable combinations, especially when combining the markers of MS with the gender
marker; (3) In the fifth scenario, the RF model was identified as the most suitable, highlighting
the importance of variables related to MS in predicting cardiovascular risk and emphasizing the
need for improvements in models for better identification of positive cases; (4) Both the model
with three MS markers and the model with five MS markers combined with the Framingham
Risk Score (MS + FRS) showed considerable performance, with significant correlations and
accuracies (0.80 and 0.84, respectively). These simpler combinations of variables can be an
interesting approach as they provide relevant information for cardiovascular risk prediction in a
less invasive manner, avoiding the need for more complex tests.

Purpuse: To elaborate a new biometric formula based on a database of the Brazilian population and comparison with the performance of other 6 formulas existing in the scientific community. Location: Participants were evaluated and underwent surgery at an ophthalmological hospital in Brasília, DF, Brazil. Design: A retrospective clinical study to compare the performance of biometric formulas. Methodology: The performances of machine learning models and Barrett Universal II, Haigis, Hoffer Q, Holladay 1, Kane and SRK/T formulas were analyzed by comparing the median absolute error and the percentage of eyes with less than 0, 5 and 1.0D of absolute error. Results: 1526 single eyes were analyzed, where the Barrett Universal II, Haigis, Hoffer Q, Holladay 1, Kane and SRK/T formulas and the MLP and SVR models obtained, respectively, the following median absolute error values: 0.393, 0.426, 0.427, 0.387, 0.409, 0.410, 0.405 and 0.370 D. The performances of the SVR model and the Holladay 1 and Barrett Universal II formulas were also superior in all studied subgroups. Conclusion: The proposed computational model-based formula obtained superior results compared to other formulas for calculating intraocular lens power in this population and in all studied biometric subgroups.

This dissertation is composed of three studies that aim to explore the influence of gamification,
which refers to the integration of game elements and mechanics in non-game contexts, such as
educational environments, on student performance and engagement and the impact of gender
stereotypes. , correlating them with the Grit personality trait. This trait is characterized by a
combination of passion and perseverance towards long-term goals, which encompasses qualities
such as resilience, determination and the ability to maintain effort and interest in the face of
challenges and setbacks. Given this, three studies were conducted, the first study is a systematic
review of the literature that investigates the correlation between Grit and gamified systems,
suggesting that gamification positively influences student engagement and performance. The
second study focuses on gender differences in stereotypical gamified tutoring systems, suggesting
that Grit alone may not be enough to reduce these disparities, indicating in the results that
gender stereotypes can negatively affect performance and engagement for women with low levels
of Grit. The third study is a qualitative analysis that explores participants’ emotional perceptions
in a stereotyped gamified environment, identifying negative feelings of lack of control. Overall,
the findings highlight the need for equitable and carefully designed gamified environments to
avoid negative effects on student performance and engagement. Future research should consider
a holistic approach, investigating complex correlations between personality traits, engagement
methods and academic performance, and the impact of gender stereotyping. Future studies
should also guide the development of teaching and learning methodologies.

The use of Information Technologies is increasingly evident in educational environments. Online educational teaching platforms make it possible to help students and teachers in the formative mission of teaching and learning in various areas of knowledge. Among the educational platforms focused on e-learning, we have the Massive Open Online Course - MOOC, which allow an adaptive educational environment for users with the support of automated resources for recommendations, carried out by artificial intelligence techniques, with the objective of learning of students according to their usage profiles. In this sense, researchers have been increasingly interested in providing teachers with strategies to accompany students more efficiently in sending feedback in the context of MOOCs in order to use the inherently human capacities of teachers to adjust the feedbacks according to the needs of students, taking into account the Sentiment analysis in posts. However, this work focuses on investigating the perception of teachers in relation to their cognitive effort and time spent in creating and adapting feedback recommendations in a simulated educational platform. This study will compare three groups of teachers, using one of the scenarios (manual, automated and semi-automated) of monitoring a simulated educational platform. The scenarios set will be used in a random experiment, where the participating teachers will evaluate through a form what is their perception of cognitive effort and time dedicated to the creation and adequacy of feedback recommendations and monitoring of students on the platform. In addition to evaluating the Technology Acceptance Model (TAM) of the proposed scenarios as a Leaning Analitics module for MOOCs environments. The results are expected to validate the hypotheses raised: that the use of artificial intelligence in the automated scenario influences the perception of teachers, leading them to present a lower workload than the manual scenario. When evaluating teachers according to the level of feelings of students in the posts, we look for indications that the perception of mental demand in the automated scenario does better when analyzing the manual scenario. Being important contributions to understand the perception of teachers when using educational platforms in their classes.

This work involves designing a physiological adaptive control law for a turbodynamic ventricular assist device (TVAD) using a lumped parameter time-varying model that describes the cardiovascular system. The TVAD is a rotary blood pump driven by an electrical motor. The system simulation also includes the adaptive feedback controller, which provides a physiologically correct cardiac output under different preload and afterload conditions. The cardiac output is estimated at each heartbeat, and the control objective is achieved by dynamically changing the motor speed controller's reference based on the systolic pressure error. TVADs provide support for blood circulation in patients with heart failure. Several control strategies have been developed over the years, emphasizing the physiological ones, which adapt their parameters to improve the patient's condition. In this work, a new strategy is proposed using a variable gain physiological controller to keep the cardiac output in a reference value under changes in both preload and afterload. Computational models are used to evaluate the performance of this control technique, which has shown better adaptability results than constant speed controllers and constant gain controllers.

Efficiently monitoring the ocean is essential to plan strategies that enable the health of the coast to be maintained. This is evident in environmental disaster scenarios, such as the oil spill reported in 2019 over a large stretch of the northeastern Brazilian coast, bringing environmental and socioeconomic consequences to the affected locations. The event showed the need to expand the national monitoring network, composed mainly of marine buoys, which are static or passive components in this task. This work proposes a framework for a persistent monitoring system with active sensing from autonomous collaborative vehicles involved in a maritime mission. This system manages simulations of a dispersion process and synchronizes agents on a mission, working on a patrol policy.

Objective: Automatic epithelium identification in Scheimpflug camera images.

Methods: A dataset of 279 normal corneas exams, obtained through the captures performed by a Scheimpflug camera, was analyzed. The proposed method has 4 steps: image acquisition; epithelium identification; separation of the curvatures of interest; and valuation and analysis of the epithelium.

Results: The Canny, Zerocross and Log algorithms were able to detect the epithelium in the total measure, and the lowest averages found in both thicknesses were with the log and zerocross methods with their variations. These measurements provided 79.74 μm, 79.85 μm, and 80.38 μm in the normal thickness and 65.91 μm, 66.08 μm, and 67.25 μm in thickness by Euclidean distance. However, the zerocross had the lowest number of defective images, and log had more than 50% of the base images with problems. In the central measure, the lowest averages found in both thicknesses were also with the log and zerocross methods with their variations, which had 75.50 μm, 75.58 μm and 75.75 μm in the normal thickness and 61.61 μm, 61.70 μm and 62.04 μm in the thickness by the Euclidean distance.

Conclusion: This work shows that with the images of the Scheimpflug camera and applying classic edge detection methods we were able to detect the epithelium in normal corneas. This shows that we were able to collect important information from the corneal layers and that this parameter can be added to any equipment that uses the Scheimpflug camera.

Safety and effectiveness are crucial quality attributes for insulin
infusion pump systems. Therefore, regulatory agencies require
the quality evaluation and approval of such systems before the
market to decrease the risk of harm, motivating the usage of a
formal Model-Based Approach (MBA) to improve quality.
Nevertheless, using a formal MBA increases costs and
development time because it requires expert knowledge and
thorough analyses of behaviors. We aim to assist the quality
evaluation of such systems in a cost-effective and time-efficient
manner, providing re-usable project artifacts by applying our
proposed approach (named MBA with CPN - MBA/CPN). We
defined a Coloured Petri nets MBA and a case study on a
commercial insulin infusion pump system to verify and validate a
reference model (as a component of MBA/CPN), describing
quality assessment scenarios. We also conducted an empirical
evaluation to verify the productivity and reusability of modelers
when using the reference model. Such a model is relevant to
reason about behaviors and quality evaluation of such
concurrent and complex systems. During the empirical
evaluation, using the reference model, 66.7% of the 12
interviewed modelers stated no effort, while 8.3%, stated low
effort, 16.7% medium effort, and 8.3% considerable effort.
Based on the modelers` knowledge, we implemented a web-
based application to assist them in re-using our proposed
approach, enabling simulation-based training. Although a
reduced number of modelers experimented with our approach,
such an evaluation provided insights to improve the MBA/CPN.
Given the empirical evaluation and the case study results,
MBA/CPN showed to be relevant to assess the quality of insulin
infusion pump systems.

In the recent years, companies were forced to adapt to the new guidelines and strategies to prevent and reduce transmission of COVID-19 in the workplace.
One of the main challenges in this adaptation is to effectively manage the workday schedule in order to reduce social contact.
This work presents a comprehensive optimization framework for automatically planning employee (staff) schedules during pandemic events.
Our framework uses integer linear programming for defining a set of general constraints that can be used to represent several types of distancing restrictions and different objective functions.
To use the framework, a company must simply instantiate a subset of these constraints with an objective function (according to its priorities).
We tested our scheduling framework in three different real-life companies, and the results show that our approach is able to improve the number of in-person workers by 15\% while attending the companies social distance restrictions.

Artificial Intelligence technology has been instrumental in the role of health care in society. It has been used in several branches of medicine. One of its main applications is in the context of the prognosis of breast cancer disease. Cancer is considered to be the second leading cause of death from disease in the world. In this context, breast cancer stands out, which is considered the highest occurrence of cancer among women in the world. One of the main challenges in this scenario is to identify which are the most relevant characteristics in the development of this type of neoplasia by a patient. This type of filter is carried out using the characteristic selection methods. This work presents a hybrid model of selection of characteristics, which is used by the patient's clinicians, in order to make a prediction of breast cancer recurrence.

According to the World Health Organization, heart disease is the leading cause of death in the world. Thousands of people die every year due to complications caused by this type of disease. In this sense, cardiologists seek early diagnosis, preventing the disease from reaching the pathological phase (i.e. the stages where limitations arise and become increasingly greater for its carriers). In the context of tachycardia-type arrhythmias, patients need to adopt the continuous use of medication and change habits to reduce the impact caused by this type of disease, thus providing a healthier life. Therefore, this study proposes and evaluates a prognostic model for arrhythmias that combines information regarding the patient's lifestyle, comorbidities and general health status to predict morbidity and mortality in patients diagnosed with tachycardia. The proposed model uses as parameters the CHA2DS2-VASc score, the clinical data and the patient's habits that, together with the use of SVM (Support Vector Machines) and KNN (K nearest neighbors) is able to classify the degree of morbidity and estimate the life perspective. In this way, the physician has access to information that should help in the decision making about the steps that will be taken in the treatment of that patient, which may involve changes in medication and dosages, levels of rigidity with habits, and in critical cases surgical intervention, always aiming to extend the patient's longevity.

This work proposes an exact method for the Orientation Problem with Hotel Selection and Time Windows (OPHS-TW) and uses it in the context of Alagoas. In OPHS-TW, a set of vertices with scores and time windows are given, and a set of hotels. The objective is to determine a fixed number of connected trips that visit some vertices and maximize the sum of the collected scores. As far as we know, this is the first exact method for OPHS-TW. Our exact model was developed using Integer Linear Programming (ILP). Computational experiments performed on OPHS TW instances found in the literature show that our exact method for OPHS-TW is capable of proving several previously unknown optima. Our algorithm found 33 unknown solutions from the literature. Of these unknown solutions, 32 were proven to be optimal. In total, 357 solutions proved to be optimal.

Latin American metropolises have been facing serious traffic congestion problems as a result of rapid population growth, increasing vehicle numbers and inefficient public policies. Most cities do not have a real-time urban traffic control system to optimize the flow of vehicles. In this context, the use of urban traffic optimization models appears as a low cost alternative to evaluate several problems and promote possible improvements. Through the Urban Mobility Simulator (SUMO), this paper proposes a new urban traffic simulation model for a prominent road in the city of Maceió, Alagoas, Brazil. The chosen road was Fernandes Lima Avenue, as it represents one of the most important road corridors of the city. The proposed model allows understanding the behavior of the vehicle flow on the road, which has peculiarities such as the blue lane (exclusive for public transportation) and three sections with pedestrian traffic lights. The validation of the proposed model was done considering data obtained from real observations. The results indicate that the model is capable of providing estimates with errors smaller than 5% related to the volume of vehicle traffic at signalized intersections and smaller than 10% related to the total average travel time of vehicles traveling along the entire avenue. Finally, the analysis of the experimental results shows that it is possible to use the proposed model to apply possible interventions on the street, such as the removal of the pedestrian signals and the blue stripe, resulting in an increase in vehicle flow, reduction of travel time, fuel consumption, and carbon dioxide emissions.

The economy of a city or region is directly proportional to its public transport system efficiency. Planning of a public transport system depends on several factors such as transport modals, origin-destination demands, quality and reliability of this service, operational costs, among others. These features leads to very complex problems like network design and vehicles frequency setting. It is usual to use heuristics, like genetic algorithm, to solve such problems in real data cases since these has large instances. In this context, present work focus to propose an optimization methodology, using biased random-key genetic algorithms, for a public bus transport system considering two aspects: (i) service quality, by minimizing passengers waiting time; (ii) economic analysis to the concessionary company, by minimizing operational cost. Thus, this methodology is applied in a case study with Maceió bus transport data to propose two frequencies scenarios that performs over 10% better than actual one in both aforesaid aspects. Finally, futures jobs would be done in order to develop multi-objective optimizations to find compound solutions.

Education in Statistic, Technology, Engineering and Mathematics (STEM) is permeated by
a traditional cis-heteronormative culture that dictates who stays in these spaces and who
will evade these courses. There is the idea that heteronormativity is the default status quo
and everything that deviates from it is seen as minority groups in STEM. The literature
presented in this dissertation confirms these statements and highlights the importance of this
work, while also showing with theories and quasi-experiments these assumptions existing
in traditional STEM teaching environments, such as the classroom and in virtual teaching
environments. STEM in its current teaching and inclusion model continues to perpetuate
this non-inclusive tradition. The use of educational technologies is increasingly present
in the classroom, especially in a current scenario where remote education is so necessary.
In this sense, gamification emerges as a powerful alternative capable of providing virtual
teaching environments. Gamification can be understood as the use of game concepts in
non-game environments, it is the use of concepts, techniques, forms, methods, elements
and variables characteristic of games that can be used in the development of a teaching
environment. The use of gamification in educational environments can be an alternative
to enable effective distance learning. Gamification in virtual teaching environments can be
considered a powerful technique for students and teachers to interact with the teaching
environment. The concern with teaching and learning is not recent, in the literature it is easy
to find studies related to this concern. Gamification can replicate stereotypes traditionally
conventionalized in STEM, when gamification elements such as colors, phrases, avatars,
sound effects and others are applied correctly, it contributes positively to learning, while,
when designed without a contextual-cultural analysis, it can reaffirm stereotypes of traditional
environments in gamified environments. Gamified and stereotyped technologies can directly
impact student performance, interfering with learning and even their evolution in teaching. In
this context, this study presented and identified the effects of stereotypes in gamified teaching
environments and their relationship with the flow experience, performance expectation and
performance of socially minorized groups such as lesbians, gays, bisexuals, transvestites,
transgender men, queer, intersex, asexuals, pansexuals, more diversity (LGBTQIAP+) in
STEM courses. For this, a systematic review is presented followed by a meta-analysis, a
quasi-experiment and a qualitative study. In this way, therefore, this study contributes to
the understanding of the impact that sexual stereotypes in gamified environments have on
student performance. It is hoped that the results of this research can also contribute to a
better inclusion of sexual diversity.

The use of gamification has been widely studied in recent years, particularly when it is used
as an intervention to increase motivation and engagement in educational settings. Even with
evidence that gamification causes positive impacts, there are also studies that point to problems
in its application. In some cases, positive results are not achieved due to the influence of several
factors, one of which is the threat of gender stereotyping, which can impact the student’s level
of self-efficacy and performance in learning. This is because self-efficacy is the feeling of having
the ability to achieve goals. Thus, when a given task is proposed and the individual does not
believe he has the necessary capabilities to perform it, success in completing the task is weak.
This factor influences the learning performance and can cause negative impacts. For example,
the stereotype that men are more skilled in math can negatively influence some women, causing
them to have a very low perception of self-efficacy, causing problems in learning, even if the
activities are carried out in a gamified environment. Thus, in this dissertation, our objective is
to identify and explain how gender stereotypes impact on self-efficacy, on the flow experience (a
state of total immersion desired in educational settings) and on student learning performance in
gamified online educational environments. For this, systematic literature review (meta-analysis)
and quasi-experimental studies will be conducted. The systematic reviews provided an overview
of how gender stereotypes and gamification affect self-efficacy, flow experience and learning
performance. Quantitative and qualitative quasi-experimental studies will be conducted to
identify and explain the effects caused by gender stereotyping in gamified online educational
environments. Thus, we aim to provide evidence regarding whether gender stereotypes affect
the flow experience, whether they affect self-efficacy and whether they contribute to poor
student learning performance. As a result, we also hope to identify if there is a gender that
might be more affected by stereotype and understand the causes of these effects. This evidence
will contribute to the generation of guidelines, recommendations and practices that result in
adaptations and implementations of gamified online educational environments without the
threat of stereotyping, the creation of fairer environments that promote gender equity.

The share of renewable energies in electricity generation has been growing worldwide. Monitoring and acquiring data is essential to recognize the renewable resources available on-site, evaluate the efficiency of electrical conversion, detect failures and optimize electrical production. Commercial monitoring systems for the photovoltaic system are generally expensive and closed for modifications. This work proposes a low-cost real-time IoT system, for micro and mini photovoltaic generation systems, that can monitor DC voltage, DC current, AC power, and seven meteorological variables. The proposed system measures all the relevant meteorological variables, it measures PV generation variables directly from the plant (not from the inverter) it is implemented using open software, connects with the internet without cables, storages data locally and in the cloud, and uses Network Time Protocol (NTP) to synchronize the devices' clocks. To the best of our knowledge, no work reported in the literature presents these features altogether. Besides, experiments carried out with the proposed system showed good effectiveness and reliability. This system enables the use of fog and cloud computing in a photovoltaic system, and the creation of a time series measurements dataset, which enables the use of machine learning to create smart photovoltaic systems.

Clustering is a fundamental class of problems with numerous applications in many areas of knowledge, including: bioinformatics, computer vision, data mining, text mining, and web page grouping. Given a set of n objects, clustering aims to automatically group such objects in k groups, usually disjunct, called clusters or groupings using a pre-established similarity measure. Clustering problems in general have high computational complexity and involve a large amount of input data. Thus, the use of parallel architectures such as Graphics Processing Units (GPUs) is interesting alternatives to accelerate the clustering process. In this work, we conducted a comparative study of GPU-accelerated metaheuristics for grouping data. Three population meta-heuristics were implemented in the GPU: Particle Swarm Optimization (PSO), Differential Evolution (DE), and Scatter Search (SS). The implementation of these meta-heuristics was divided into two parts: the independent part of the problem and the dependent part of the problem. The independent part of the problem refers to the selection, replacement, and combination operators for each meta-heuristic, while the dependent part refers to the objective function. The independent part was implemented using the libCudaOptimize framework, and the dependent part was created by transforming the clustering problem into a global optimization problem subject to cash constraints. The proposed meta-heuristics were compared with the best current clustering algorithm considering the efficiency of the execution time and the quality of the solution. The results indicate that the GPU-based PSO (GPU-PSO) obtained the best results in comparison with the other GPU-based metaheuristics and the best method today. Also, our GPU implementation of the objective function obtained an average speedup of 175x over the sequential version. These results demonstrate that a GPU approach to the clustering problem is very promising.

The process of grouping data is known as clustering. In the literature, the clustering process, or grouping, has two variations: (i) if the number of clusters is predefined, this process is known as the Clustering Problem (CP) or k-Clustering Problem, and (ii) when the number of clusters is not defined, the process becomes known as the Automatic Clustering Problem (PCA). The Automatic Clustering Problem is classified as NP-difficult, which prevents the exact value of the clusters. The importance of having well-grouped data is crucial for
making more assertive decisions. The data grouping technique has applicability in the most diverse areas of knowledge, such as: engineering, administration, economics, biology, physics, among others, which present problems with the representation of mathematical models. Genetic algorithms represent a class
of algorithms that can be used to solve this type of problem, these algorithms are based on Darwinian evolution processes, selecting the best solutions within a population. The BRKGA (Bia-sed Random Key Genetic Algorithm), in translation Genetic Algorithm of Addicted Random Keys is presented as a variation of the genetic algorithms, in which the solutions of a problem are represented as vectors of real keys, generated randomly, in the continuous interval of [0.1). The fitness of a viable solution is determined by the decoder that maps this vector to a real value. This work proposes a multipopular BRKGA to identify the ideal number of declusters, in accordance with the silhouette index, a measure of grouping efficiency widely used in the literature. In the proposed algorithm, the solution space is partitioned so that each subpopulation the algorithm represents solutions with a cluster k number. Thus, for each subpopulation, an independent BRKGA is applied. Computational experiments were carried out in fifty-five instances of the literature. The proposed algorithm is compared with existing methods in the literature, showing superior results, leading to the understanding that the algorithm is promising.

Data scientists often develop machine learning models to solve a variety of problems in the industry and academy. To build these models, these professionals usually perform activities that are also performed in the traditional software development lifecycle, such as eliciting and implementing requirements. One might argue that data scientists could rely on the engineering of traditional software development to build machine learning models. However, machine learning development presents certain characteristics, which may raise challenges that lead to the need for adopting new practices. The literature lacks in characterizing this knowledge from the perspective of the data scientists. In this paper, we characterize challenges and practices addressing the engineering of machine learning models that deserve attention from the research community. To this end, we performed a qualitative study with eight data scientists across five different companies having different levels of experience in developing machine learning models. Our findings suggest that: (i) data processing and feature engineering are the most challenging stages in the development of machine learning models; (ii) it is essential synergy between data scientists and domain experts in most of these stages; and (iii) the development of machine learning models lacks the support of a well-engineered process.

We present with this work processes towards selection automation to support the Systematic Review of Literature (SLR) so that the time in the analysis of articles in the selection phase is optimized, allowing the reviewers to move to the extraction phase in a faster and smoother way, based on statistical views. The standard SLR methodology has three main phases: planning of review, conducting of review, and dissemination. Conducting review is the most daunting phase because almost all steps refer to manual reading and categorization. In this phase, authors of reviews verify which studies meet the inclusion criteria for the extraction phase by a manual reading of title and abstract. We use topic modeling to map and search for relationships not so evident to human discernment, added to the efficiency of the average time for article analysis in the selection phase, regardless of the number of articles, but rather the computational power to reduce the time of this phase from months to hours or minutes. This work proposes some strategies in topics for clustering the studies, generation of summary for clustered studies, and data graphics. For better data stratification, we already tested some propositions, and we have achieved quite good results. We obtain these results by exploring the BIBTEXdata. For clustering of studies, we transformed the title, abstract and keywords, into a wordcloud for each study, and grouped using a natural language processing technique called Sentence Boundary Detection for finding and segmenting meaningful individual sentences, studies with same sentences are put together, organized, and clustered by sentences frequency. We achieve the generation of summary for clustered studies using natural language generation. We perform a comparison of Markov Chain Generation with Recurrent Neural Network generation for quality assessment of the generated text. We obtain data graphics by exploring BIBTEXdata already available, and min- ing relations of semantic changes or author’s groups of collaboration. The results methodology follows the best practices for conducting and reporting reviews, thus solving a practical problem effectively with reproducible and repeatable results. The main goal of review automation is to save lives by accelerating the adoption of better healthcare standards. We present techniques for studies summarization, categorization in related topics, and bibliography statistics.

Lung cancer is the most common and deadly form of cancer, and its early diagnosis is de-cisive to the patient’s survival. Computed Tomography (CT) is the gold-standard imagingmodality for lung cancer management, but recent studies have shown the potential of Mag-netic Resonance Imaging (MRI) in lung cancer diagnosis and how combining multimodalitymedical images can yield better outcomes. In this study, we evaluated whether the combina-tion of CT and MRI scans from lung cancer patients can leverage a more precise malignancysuspiciousness classification. For such, we registered paired CT and MRI scans from 47 pa-tients, segmented the nodules in each modality, extracted radiomics features, and performedan experiment with an XGBoost classifier, evaluating models’ performance metrics across30 trials. The same experiment was performed with four sets of features: 1) CT-only; 2)MRI-only; 3) CT and MRI features; 4) CT/MRI fused images. Our results indicate that theimage fusion approach can yield significant AUC performance gains over the single modal-ities models, with an average AUC of0.794, but feature concatenation is not an adequatestrategy for dealing with multimodality data, as its average AUC of0.770didn’t indicate im-provement over the single modalities. Additionally, we observed that MRI, with an averageAUC of0.770, has shown significantly better performance than CT, with0.754, encourag-ing further studies in MRI as lung cancer management image modality. Finally, an analysison the importance of radiomics features reinforced the relevance of features that reflects onmorphological characteristics of a nodule, such as its dimension and roundness, as well astexture features that relate to the intratumoral environment, measuring its complexity andhomogeneity.

The recognition of dynamic signs of sign languages is a difficult task that is starting to become feasible with the use of deep neural networks. However, the absence of large annotated databases makes the training of these classifiers unfeasible. In this work, a database, entitled Skelibras, was built, containing $57760$ samples of $6572$ classes of dynamic signs of Brazilian Sign Language (Libras). Each sign contains sequences of skeletons (poses) of the body and hands, automatically extracted and synchronized from videos of the Corpus de Libras base. To extract and organize these annotated data consistently, a methodology capable of identifying and tracking the poses of each speaker is presented; synchronize subtitles and speakers present in conversations; and synchronize video data acquired from different points of view with subtitles. We performed experiments on variations of deep neural networks based on convolutional layers, dense layers, and LSTMs units to validate and provide preliminary results on the basis generated in this work, thus enabling future comparison with new dynamic signal recognition methods.

The Minimum Broadcast Time (MBT) is a well-known data dissemination problem whose goal is to find a broadcast scheme that minimizes the number of steps needed to execute the broadcast operation. The Weighted Minimum Broadcast Time (WMBT) is a generalization of the MBT, such that each operation has a cost. Both problems have many applications in distributed systems and swarms robots. This work proposes Biased Random-Key Genetic Algorithms (BRKGA) and a hybrid algorithm (BRKGA + Integer Linear Programming) for the MBT and WMBT. We carry out experiments with our BRKGA on instances commonly used in the literature, and also on massive synthetic instances (up to 1000 vertices), allowing us to cover many possibilities of real industry topologies. Our proposal is also compared with state-of-the-art exact methods and heuristics. Experimental results show that our algorithms are able to outperform the best-known heuristics for the MBT and WMBT, and also that they are a very good alternative for large instances that cannot be solved by current exact methods.

A presença de code smells indica uma má escolha de implementação e geralmente piora a qualidade do software. Portanto, a detecção de code smells é uma técnica simples para identificar oportunidades de refatoração em sistemas de software. Neste contexto, estudos anteriores desenvolveram ferramentas de detecção de code smells, fornecendo resultados diferentes. No entanto, tais ferramentas apresentam algumas limitações porque os code smells podem ser subjetivamente interpretados e detectados de maneiras diferentes. Para superar essas limitações, usamos diferentes técnicas de aprendizado de máquina para classificar os códigos prejudiciais. Um código prejudicial é um fator essencial a ser levado em consideração ao se reportar os resultados de detecção de code smells, uma vez que os códigos maliciosos contêm bugs, fornecendo a priorização dos esforços de refatoração. Em nosso experimento, aplicamos algoritmos de aprendizado de máquina diferentes em projetos de código aberto para detectar códigos prejudiciais que exploram diferentes recursos e técnicas.

O teste de software é uma atividade crucial para verificar a qualidade interna de um software. Durante os testes, os programadores frequentemente criam testes para o comportamento normal de uma determinada funcionalidade (por exemplo, um documento foi corretamente enviado para a nuvem?). Contudo, pouco se sabe se os programadores também criam testes para o comportamento excepcional (por exemplo, o que acontece se a rede falhar durante o envio do arquivo?). Para minimizar esta falha de conhecimento, neste artigo desenhamos e realizamos um estudo de método misto para entender como 417 projetos Java de código aberto estão testando o comportamento excepcional usando os frameworks JUnit e TestNG, e a biblioteca AssertJ. Verificamos que 254 (60,91%) projetos têm pelo menos um método de teste dedicado a testar o comportamento excepcional. Verificamos também que o número de métodos de teste de comportamento excepcional em relação ao número total de métodos de teste se situa entre 0% e 10% em 317 (76,02%) projetos. Além disso, 239 (57,31%) projetos testam apenas até 10% das exceções utilizadas no sistema em teste (System Under Test - SUT). Quando se trata de aplicações móveis, descobrimos que, em geral, os programadores prestam menos atenção aos testes de comportamento excepcionais quando comparados com os programadores de desktop/servidor e multiplataforma. Em geral, encontramos mais métodos de teste abrangendo exceções personalizadas (as criadas no próprio projeto) quando comparadas com as excepções padrão disponíveis no Java Development Kit (JDK) ou em bibliotecas de terceiros. Para triangular os resultados, realizamos uma pesquisa com 66 programadores dos projetos que estudamos. Em geral, os resultados da pesquisa confirmam os nossos resultados. Em particular, a maioria dos participantes concorda que os programadores negligenciam frequentemente testes de comportamento excepcionais. Como implicações, os nossos números podem ser importantes para alertar os desenvolvedores de que devem ser empregados mais esforços na criação de testes para o comportamento excepcional.

This work aims to assist in the identification of clothing attributes using a multimodal Deep
Learning strategy.We propose the use of images and unstructured textual descriptions to organize
clothing items catalogs. Such types of data are employed to train deep neural network architectures
in multi-class classification problems, which are then able to automatically recognize
attributes from these two types of data commonly found in e-commerce environments. Three
classes of architecture were experimented: variations of the VGG architecture for recognition
from images; architectures combining embedding, convolutional and recurrent layers for text
recognition; and hybrid architectures that combine elements from each of the previous architectures.
Using a database that we collected through a Web Crawler from a large e-commerce site,
we show in our experiments that hybrid architectures achieve a better result in the classification
task by combining both types of data. Our methodology makes it possible to feed clothing
recommendation systems, due to the possibility of compiling and structuring the catalog data,
in addition to allowing the detection of insufficient visual and textual descriptions for a given
attribute that can be improved, when unimodal classifiers fail to recognize this attribute.

O teste baseado em modelos (Model-Based Tesing - MBT) baseia-se em modelos do comportamento do sistema para gerar testes abstratos. Testadores reutilizam especificações formais (e.g., modelos de redes de Petri coloridas (Coloured Petri Nets - CPN)) para projetar testes para sistemas críticos de segurança. Neste trabalho, por uma revisão terciária, foi identificado um número considerável de revisões de literatura focadas na análise do uso de linguagens de especificação para realizar MBT. Entretanto, ainda existe uma lacuna de pesquisa em relação à análise de abordagens baseadas em CPN para geração de testes abstratos. Para preencher essa lacuna de pesquisa, neste trabalho, uma análise comparativa de abordagens para geração de testes abstratos com base em modelos CPN foi também realizada, por meio de uma revisão sistemática da literatura e um estudo de caso sobre sistemas médicos: eletrocardiografia e bomba de infusão de insulina. A partir da análise comparativa, são fornecidas informações para os testadores que precisam selecionar a abordagem de geração de testes abstratos mais adequada ao aplicar MBT usando CPN. Com os resultados obtidos, é possível identificar que a escolha depende do tamanho do espaço de estados do modelo CPN.

Este trabalho apresenta uma plataforma aberta de um robô para atividades de pesquisa e educacionais. A plataforma é definida como um robô omnidirecional, que é um dos primeiros passos no currículo de robótica quando se deseja controlar um robô, dado que este não impões restrições ao movimento num plano. Os robôs são projetados para serem de baixo custo e com acesso aberto ao software e hardware. A maior parte do projeto mecânico també é fácil de imprimir nas impressoras 3D comumente usadas. O robô é integrado com a plataforma ROS - Robot Operating Systems, amplamente utilizada na comunidade de robótica, e o modelo de simulação em ROS-Gazebo é fornecido para permitir uma passagem fácil do ambiente de simulação ao ambiente experimental. Experimentos reais com os robôs propostos são mostrados, assim como a plataforma de simulação com múltiplos robôs.

Code comprehension is crucial in software maintenance activities, though it can be hindered by misunderstandings and confusion patterns, namely, atoms of confusion. They are small pieces of code using specific programming language constructs, such as Conditional Operators and Comma Operators. A previous study showed that these atoms of confusion impact developers' performance, i.e., time and accuracy, and increase code misunderstandings. However, empirical knowledge of the impact of such atoms on code comprehension is still scarce, especially when it comes to analyzing that impact on developers' visual attention. The present study evaluates whether developers misunderstand the code in the presence of atoms of confusion with an eye tracker. For this purpose, we measure time, accuracy, and analyze the distribution of visual attention. We conduct a controlled experiment with 30 students and software practitioners. We ask the subjects to specify the output of three tasks with atoms and three without atoms randomly assigned using a Latin Square design. We use an eye-tracking camera to detect the visual attention of the participants while solving the tasks. From an aggregated perspective, we observed an increase by 43.02% in time and 36.8% in gaze transitions in code snippets with atoms. For accuracy, no statistically significant difference was observed. We also confirm that the regions that receive most of the eye attention were the regions with atoms. Our findings reinforce that atoms hinder developers’ performance and comprehension. So, developers should avoid writing code with them.

A multi-parameter monitoring system is usually applied to keep track of the clinical condition of patients in Intensive Care Units (ICUs). ICUs are continuous monitoring environments used to host patients in serious health conditions. These Systems of Systems (SoS), used in ICUs comprise of a set of Constituent Systems (CS) to measure parameters such as heart rate, respiratory frequency, and temperature. Due to the critical nature and relevance of ICUs, such SoS shall be as reliable as possible. This is specially true in emergency situations, as is the case of the COVID-19 outbreak, which resulted in the burden of health care systems. Multi-parameter monitoring systems in ICUs shall have high levels of reliability, given that failures can provide risks to the safety of patients. We performed relia- bility analysis and provided insights to assist the management of multi-parameter monitoring systems used in ICUs, also considering preventive maintenance. We elicited requirements by interviewing a professional with more than fifteen years of hospital experience in maintenance. In addition, we analyzed existing systems and literature reviews. Therefore, we modeled multi-parameter monitoring systems for ICUs using the CHESS-ML modeling language. Afterward, we conducted reliability analysis by applying the CHESS-SBA plugin state-based analysis, simulating different scenarios. Based on the reliability analysis, we identified that main power supply and the battery are the CS that present major negative impacts in the total reliability the entire system in failure situations. In emergency situations, reduced time ranges of preventive maintenance, when applied during a short period, showed to be promising strategies to increase quality of multi-parameter monitoring systems for ICUs.