Advances in Soft Computing

Research on damage detection of road surfaces has been an active area of research, but most studies have focused so far on the detection of the presence of damages. However, in real-world scenarios, road managers need to clearly understand the type of damage and its extent in order to take effective action in advance or to allocate the necessary resources. Moreover, currently there are few uniform and openly available road damage datasets, leading to a lack of a common benchmark for road damage detection. Such dataset could be used in a great variety of applications; herein, it is intended to serve as the acquisition component of a physical asset management tool which can aid governments agencies for planning purposes, or by infrastructure maintenance companies. In this paper, we make two contributions to address these issues. First, we present a large-scale road damage dataset, which includes a more balanced and representative set of damages. This dataset is composed of 18,034 road damage images captured with a smartphone, with 45,435 instances road surface damages. Second, we trained different types of object detection methods, both traditional (an LBP-cascaded classifier) and deep learning-based, specifically, MobileNet and RetinaNet, which are amenable for embedded and mobile and We compare the accuracy and inference time of all these models with others in the state of the art.

The Curriculum-Based Course Timetabling (CB-CTT) is a problem periodically solved in educational institutions, still, because of the diversity of conditions that define it within different educational contexts, selecting the solution approach that best suits the particular requirements of an instance is a complex task that can be properly formulated as an algorithm selection problem. In this paper, we analyze four selection mechanisms that could be used as algorithms recommendation models. From this analysis, it is concluded that the proposed regression approach exhibited the highest performance. Therefore, it could be applied for algorithm recommendation to solve CB-CTT instances.

The prediction of vessel maritime navigation has become an interesting topic in the last years, especially in areas of economical commercial exchange and security. Also, vessels monitoring requires better systems and techniques that help enterprises and governments to protect their interests. In specific, the prediction of vessels movements is important concerning safety and tracking. However, the applications of prediction techniques have a high cost of computational efficiency and low resource-saving. This article presents a sample method to select historical data on ship-specific routes to optimize the computational performance of the prediction of ship positions and route estimation in real-time. These historical navigation data can help us to estimate a complete path and perform vessel positions predictions through time. This method works in a vessel tracking system in order to save computational work when predictions or route estimations are in execution. The results obtained after testing the method are almost acceptable concerning route estimation with a precision of 74.98%, and with vessel positions predictions through time a 79% of accuracy.

Recognition of user intent from web queries is required by search engines to improve user experience by adapting search results to the user goals. In this paper we report findings of intent recognition from search queries, using two intent annotated benchmark datasets, ATIS and AOL web query dataset. Both these corpora have been automatically translated from English to Urdu. Through multiple experiments, we analyze and compare performance of four Deep Neural Network (DNN) based models and their architectures, i.e. CNN, LSTM, bi-directional LSTM, and CLSTM (CNN+LSTM). On ATIS dataset, CNN achieves 92.4% accuracy on binary classification. While on AOL dataset, BLSTM performs the best with 83.1% accuracy for 5% test sample proportion for 3 intent classes.

Polarity lexicons have largely been used for opinion mining, also known as sentiment analysis. Their high relevance has made them invaluable when it comes to classifying the opinions originated by digital platforms users. However, there is a great dependence on resources developed in other languages, in special English, to create a lexicon in Spanish. In this work we develop a polarity lexicon for Mexican Spanish from a thesaurus and two seed words sets inside it that we use to assign polarity values to words, from a frequentist treatment, and taking into account the assumption that synonymous words have similar polarities.

The most widely used method for mineral type classification from a rock thin section is done by the observation of optical properties of a mineral in a polarized microscope rotation stage. Several studies propose the application of digital image processing techniques and Neural Networks to automate this task. This study uses simpler and more scalable machine learning techniques, being nearest neighbor and decision tree, and adds new optical properties to be extracted from the digital images. Two datasets are used, one provided by Ferdowsi University of Mashhad, with 17 different minerals, and another built from scratch in the geology department of Federal University of Pelotas, containing 4 different minerals. The datasets are composed of mineral images captured under cross and plane polarized light from different rock thin sections. For each dataset used, we took a pair of images of the same mineral taken on different lights, extracted optical properties of color and texture, applied a machine learning algorithm and provided the results. At the end of this study, we demonstrate it is possible to achieve high accuracy as Neural Networks with more simple machine learning algorithms as the our dataset showed average results as high as 97% and Mashhad’s as high as 93%.

In this work, we report the results of our experiments on the task of distinguishing the semantics of verb-noun collocations in a Spanish corpus. This semantics was represented by four lexical functions of the Meaning-Text Theory. Each lexical function specifies a certain universal semantic concept found in any natural language. Knowledge of collocation and its semantic content is important for natural language processing, as collocation comprises the restrictions on how words can be used together. We experimented with a combination of GloVe word embeddings as a recent and extended algorithm for vector representation of words and a deep neural architecture, in order to recover most of the context of verb-noun collocations in a meaningful way which could discriminate among lexical functions. Our corpus was a collection of 1,131 Excelsior newspaper issues. As our results showed, the proposed deep neural architecture outperformed state-of-the-art supervised learning methods.

Emerging depth sensors and new interaction paradigms enable to create more immersive and intuitive Natural User Interfaces by recognizing body gestures. One of the vision-based devices that has received plenty of attention is the Leap Motion Controller (LMC). This device models the 3D position of hands and fingers and provides more than 50 features such as palm center and fingertips. In spite of the fact that the LMC provides such useful information of the hands, developers still have to deal with the hand gesture recognition problem. For this reason, several researchers approached this problem using well-known machine learning techniques used for gesture recognition such as SVM for static gestures and DTW for dynamic gestures. At this point, we propose an approach that applies a resampling technique based on fast Fourier Transform algorithm to feed a CNN+LSTM neural network in order to identify both static and dynamic gestures. As far as our knowledge, there is no full dataset based on the LMC that includes both types of gestures. Therefore, we also introduce the Hybrid Hand Gesture Recognition database, which consists of a large set of gestures generated with the LMC, including both type of gestures with different temporal sizes. Experimental results showed the robustness of our approach in terms of recognizing both type of gestures. Moreover, our approach outperforms other well-known algorithms of the gesture recognition field.

Machines that are able to recognize and predict human affective states or emotions have become increasingly desirable over the last three decades [10]. This can be attributed to their relevance to human endeavors accompanied by the ubiquity of computing devices and an increasing trend for technology to be ever more present and engrained in people’s daily lives. There have been advancements in AI applications that are able to detect a person’s affective states through the use of machine learning models. These advancements are mostly based on architectures that take as inputs facial expression image data instances or highly specific sensor data, such as ECG or skin conductivity readings. The problem with these approaches is that models are not being designed with capabilities to receive, and modularly add multiple sensory inputs, therefore failing to operate and deal with the differences that exist on how individuals experience emotions. The present publication proposes a methodology consisting of a continuous multi sensory data acquisition process, and the construction of a feed forward classification neural network with three, 200 neuron, hidden layers. Experimentation was carried out on six different subjects, collecting over 100 h of data points containing environmental and personal variables such as activity (accelerations), light exposure, temperature and humidity. A different model was trained for each one of the subjects for 60–1500 epochs, yielding individual prediction accuracies, on test sets, of 82(%)–95(%).

In the last years, innovative techniques like Transfer Learning have impacted strongly in Natural Language Processing, increasing massively the state-of-the-art in several challenging tasks. In particular, the Universal Language Model Fine-Tuning (ULMFiT) algorithm has proven to have an impressive performance on several English text classification tasks. In this paper, we aim at developing an algorithm for Spanish Sentiment Analysis of short texts that is comparable to the state-of-the-art. In order to do so, we have adapted the ULMFiT algorithm to this setting. Experimental results on benchmark datasets (InterTASS 2017 and InterTASS 2018) show how this simple transfer learning approach performs well when compared to fancy deep learning techniques.

Sentiment Analysis was developed to support individuals in the harsh task of obtaining significant information from large amounts of non-structured opinionated data sources, such as social networks and specialized reviews websites. A yet more challenging task is to point out which part of the target entity is addressed in the opinion. This task is called Aspect-Based Sentiment Analysis. The majority of work focuses on coping with English text in the literature, but other languages lack resources, tools, and techniques. This paper focuses on Aspect-Based Sentiment Analysis for Accommodation Services Reviews written in Brazilian Portuguese. Our proposed approach uses Convolution Neural Networks with inputs in Character-level. Results suggest that our approach outperforms lexicon-based and LSTM-based approaches, displaying state-of-the-art performance for binary Aspect-Based Sentiment Analysis.

Nowadays, the Mexican government is showing a great interest in decreasing the crime rate in Mexico. A way to carry out this task is to understand criminal behavior in each Mexico states by using an eXplainable Artificial Intelligence (XAI) model. In this paper, we propose to understand the criminal behavior of the Mexico city by using an XAI model jointly with our proposed feature representation based on the weather. Our experimental results show how our proposed feature representation allows for improving all tested classifiers. Also, we show that the XAI-based classifier improves other tested state-of-the-art classifiers.

Job Shop Scheduling problems have become popular because of their many industrial and practical applications. Among the many solving strategies for this problem, selection hyper-heuristics have attracted attention due to their promising results in this and similar optimization problems. A selection hyper-heuristic is a method that determines which heuristic to apply at given points of the problem throughout the solving process. Unfortunately, results from previous studies show that selection hyper-heuristics are not free from making wrong choices. Hence, this paper explores a novel way of improving selection hyper-heuristics by using neural networks that are trained with information from existing selection hyper-heuristics. These networks learn high-level patterns that result in improved performance concerning the hyper-heuristics they were generated from. At the end of the process, the neural networks work as hyper-heuristics that perform better than their original counterparts. The results presented in this paper confirm the idea that we can refine existing hyper-heuristics to the point of being able to defeat the best possible heuristic for each instance. For example, one of our experiments generated one hyper-heuristic that produced a schedule that reduced the makespan of the one obtained by a synthetic oracle by ten days.

Smart Mobility seeks to meet urban requirements within a city and solve the urban mobility problems, one of them is related with vehicular traffic. The anomalous vehicular traffic is an unexpected change in the day-to-day vehicular traffic caused by different reasons, such as an accident, an event, road works or a natural disaster. An early detection of anomalous vehicular traffic allows to alert drivers of the anomaly and can make better decisions during their journey. The current solutions for this problem are mainly focused on the development of new algorithms, without giving enough importance to the extraction of underlying information from vehicular traffic, and even more, when this is a univariate time series and it is not possible to obtain other context features that describes its behavior. To address this issue, we propose a methodology for temporary, spectral and aggregation features and an unsupervised learning model to detect anomalous vehicular traffic. The methodology was evaluated in a real vehicular traffic database. Experimental results show that by using spectral attributes the detection of anomalous vehicular traffic, the Isolation Forest obtains the best results.

Mission Critical Systems (MCS) require continuous operation since a failure might cause economic or human losses. Autonomous Decentralized Service Oriented Architecture (ADSOA) is a proposal to design and develop MCS in which the system functionality is divided into service units in order to provide functional reliability and load balancing; on the other hand, it offers high availability through distributed replicas. A fault detection and recovering technology has been proposed for ADSOA based on collective intelligence. In this technology, an operation service level degradation should be detected autonomously by the service units at a point in which the continuity of the service may be compromised. Once a failure of this type is detected, each service unit analyses the system’s state and collectively decide the strategy to recover itself. The recovery technology instructs those selected service units to be gradually cloned in order to get the operational service level.

Gentrification is a problem in big cities that confounds economic, political and population factors. Whenever it happens, people in the higher brackets of income replace people of low income. This replacement generates population displacement, which force people to change their lives radically.In this work, we use Classification Trees to generate an index, which will indicate the likelihood for a neighborhood to gentrify. This index uses many population variables that include things like age, education and transportation.This system can be used later to inform decisions regarding urban housing and transportation. We can prevent areas of the city of overflowing with private investment in lieu of public housing policy that allows people to stay in their places of living.We expect this work to be a stepping zone on working towards a generalization of gentrification effects in different cities in the world.

A Third Generation Artificial Neural Network plus a Parallel Multi-Core approach is presented. This approach is capable of efficiently tackle the problem of finding the shortest path between two nodes, for big cases with thousands of nodes. The efficient solution of the shortest path problem has applications in such important and current areas as robotics, telecommunications, operation research, game theory, computer networks, internet, industrial design, transport phenomena, design of electronic circuits and others, so it is a subject of great interest in the area of combinatorial optimization. Due to the parallel design of the Pulsed Neuronal Network presented here, it is possible speed up the solution using parallel multi-processors; this solution approach can be highly competitive, as observed from the good results obtained, even in cases with thousands of nodes.

Student attrition is one of the most important problems for any school, being it private or public.In public education, a high attrition rate reflects poorly in the school, as it is wasting public taxes on students that do not finish their majors. In private education, it means the school revenue decreases considerably. Much work has been done on predicting churn rates in the Telecommunication industry, in this work we use similar techniques to predict churn rates in education.We explore the data extensively and see the possible correlations between attrition and variables like entrance examination, place where the students are from and grades up to the point of abandonment of the major.

We describe in this paper the use of Modular Neural Networks (MNN) for pattern recognition with parallel processing using a cluster of computers with a master-slave topology. In this paper, we are proposing the use of MNN for face recognition with large databases to validate the efficiency of the proposed approach. Also, a parallel genetic algorithm for architecture optimization was used to achieve an optimal design of the MNN. The main idea of this paper is the use of parallel genetic algorithms to find the best architecture with large databases of faces, because when the database to be considered is large, the main problem is the processing time to train the MNN. Network parameters are adjusted by a combination of the training pattern set and the corresponding errors between the desired output and the actual network response. To control a learning process, a criterion is needed to decide the time for terminating the process.

In this paper we consider a class of iterative gradient projection methods for solving optimization problems with orthogonality constraints. The proposed method can be seen as a forward-backward gradient projection method which is an extension of a gradient method based on the Cayley transform. The proposal incorporates a self-adaptive scaling matrix and the Barzilai-Borwein step-sizes that accelerate the convergence of the method. In order to preserve feasibility, we adopt a projection operator based on the QR factorization. We demonstrate the efficiency of our procedure in several test problems including eigenvalue computations and sparse principal component analysis. Numerical comparisons show that our proposal is effective for solving these kind of problems and presents competitive results compared with some state-of-art methods.

Proximity or similarity searching is one of the most important tasks in artificial intelligence concerning multimedia databases. If there is a distance function to compare any two objects in a collection, then similarity can be modeled as a metric space. One of the most important techniques used for high dimensional data is the permutation-based algorithm, where the problem is mapped into another space (permutations space) where distances are much easier to compute, but solving similarity queries with the least number of distances computed is still a challenge. The approach in this work consists in using Locality-Sensitive Hashing (LSH). The experiments reported in this paper show that the proposed way to adapt LSH to the permutation based algorithm has a competitive tradeoff between recall and distances.

It is well known that Parallel Task Graphs (PTG) are modeled with Directed Acyclic Graphs (DAG Tasks). DAG tasks are scheduled in Heterogeneous Distributed Computing Systems (HDCS) for execution with different techniques which seek to reduce completion of each PTG. Proposed planning techniques generally only make use of the critical path in planning as an internal characteristic of the DAG Task, helping to optimize scheduling. In this study it is shown that analyzing other internal characteristics, such as layering and graph density aside from the critical path of DAG workflow tasks, before being scheduled in execution locations, can improve computer system performance, as well as optimize the use of their resources. For the above, the internal characteristics considered in this study of each DAG task are: the critical path, layering as well as graph density. The analyzed DAG tasks are synthetic loads produced with a graph generation algorithm as well as real application graphs. The findings obtained with the experiments performed show that the distribution estimation algorithm obtains better response times than the genetic algorithm.

Tuning conventional controllers could be a difficult task when experimental methodologies are implemented. Moreover, nowadays, Microgrids (MGs) require specific operation responses that could be achieved if the conventional controllers are correctly tuned. As a result, an optimization methodology that gets the correct parameters of conventional controller can improve the performance of the (MGs). This paper proposes the tuning of the conventional controllers used in a Grid Forming Inverters (GFMI) two voltage PID control loops, two current PID control loops, and the frequency PID controller. In a conventional control architecture of a GFMI. In GFMIs that act as voltage sources within a MG system, an incorrect tuning would harm the regulation of the dispatched voltage and frequency values to the linked electrical loads. Previously, optimization methods have been used for tuning conventional controllers, however, this is usually done in a grid-connected configuration. This work delimits the possible gain values to a desired controlled system response, by then optimizing over the controller requirements using genetic algorithms. In addition, a complete study of the tuning process under different genetic algorithm parameters (population and mutation) is presented.

With the advance of technology, the generation of massive amounts of information grows every day, generating complex problems difficult to manage in an efficient way. Therefore, researchers have studied and modeled the way in which natural biological systems react and behave in certain situations, allowing to developed algorithms that exhibit a capacity to learn and/or adapt to new situations, obtaining better results than traditional approaches. In this article we present a new variant of the Differential Evolution (DE) algorithm inspired by the concept of the Learnable Evolution Model (LEM) to enhance the search capability through a selection mechanism based on machine learning to create a set of rules that allows the inferring of new candidates in the population that emerge not only the random scan. The proposed algorithm is tested and validated on a set of 23 bechmark test functions and its performance is compared with other metaheuristics. Results indicate that the proposed DE+LEM is competitive with other metaheuristic.

When using the mixed-integer programming to model situations where the limit of the variables follows a box constraint, we find nonlinear problems. To solve this, linearization techniques of these disjunctive inequality constraints are typically used, including constants associated to the variable bounds called M-constants or big-M. Calculation of these constants is an open problem since their values affect the reliability of the optimal solution and convergence of the optimization algorithm. To solve this problem, this work proposes a new population-based metaheuristic optimization method, namely wound treatment optimization (WTO) for calculating the M-constant in a typical domain known as the fixed-charge transportation problem. WTO is inspired on the social wound treatment present in ants after raids. This method allows population diversity that allows to find near-optimal solutions. Experiments of the WTO method on the fixed-charge transportation problem validated its performance and efficiency to find tighten solutions of the M-constant that minimizes the objective function of the problem.

Deficient nutrition has caused high rates of overweight and obesity in the Mexican population, increasing the cases of people with diabetes and hypertension. In order to solve this, it is necessary to promote a change in the alimentation to reduce the rates of overweight and obesity. To achieve this, we propose a friendly solution to generate a change in the eating habits of the Mexicans by the generation of balance diets. Diet automation has been already created with different algorithms and applications in the past, but with a different purpose and objectives. Particularly, this work is focused on the design of balanced diets applying a Particle Swarm Optimization algorithm. The proposed methodology considers the physical characteristics of the user. To validate the accuracy of the proposed methodology several experiments were performed to asses if the proposal is capable of achieving the calorie goal in terms of the Harris-Benedict equation. The experimental results suggest that it is possible to generate diets using Particle Swarm Optimization algorithms with an error less than 10%.

In recent years, the interest to solve dynamic combinatorial optimization problems has increased. Metaheuristic algorithms have been used to find good solutions in a reasonably low time, in addition, the use of self-adaptive strategies has increased considerably because they have proved to be a good option to improve performance in these algorithms. In this research, a self-adaptive mechanism is developed to improve the performance of the genetic algorithm for dynamic combinatorial problems, using the strategy of genotype-phenotype mapping and probabilistic distributions. Results demonstrate the capability of the mechanism to help algorithms to adapt in dynamic environments.

Recent works have shown that it is possible to use information extracted from images to address the task of automatic gender identification. These proposals have validated their solutions using monolingual datasets, i.e., collections where images are shared by users having the same mother tongue. This paper aims to test the usefulness of images collected from users who do not share the same language. In principle, these users present cultural differences, which may be reflected in the images they share. However, a cross-cultural image-based approach would be very useful for languages where data is not available or scarce. The experiments presented demonstrate that characteristics obtained from the images, regardless of the users’ mother tongue, can be used for gender prediction. They mainly confirm the usefulness of a cross-cultural image-based approach, showing that culturally different individuals with equivalent profiles traits tend to share similar images.

Industrial processes such as manufacturing and machining parts, fault detection and quality control are some of the areas of study that encompass computational vision techniques, image processing and currently fuzzy logic. Particularly, the edge detection of objects in captured images is a technique widely used in industrial automated systems. In this work, we propose a technique for edge detection in digital images obtained from real pieces based on fuzzy logic. The fuzzy inference model works with 18 Mamdani type rules and was built with 8 input variables and one output variable. It is, the processing of the image was performed under the conditions of the lighting scenario, background and the color of the piece. The performance of the algorithm was evaluated on several images captured from different work environments and it was compared with traditional computer vision methods using gradient operators. The use of fuzzy logic in image processing expands the possibilities to solve a problem and provides more answers over the restrictions of classical methods.

Materials for aerospace industry such as aluminium alloys are of prime use in a variety of components. Performance and reliability of these components and structures mostly lie down on the fatigue resistance among other structural characteristics. Shot peening processing is widely employed to improve the fatigue properties. However, proper selection and control of peening factors (parameters) is needed to ensure that peening effects became beneficial rather than detrimental. The present study focuses on finding optimal peening parameters by considering multiple performance characteristics using grey fuzzy methodology. Adaptive neuro-fuzzy inference system (ANFIS) approach was used to investigate the effects of the input parameters, namely, shot type, coverage and incidence angle on the performance parameters, i.e. residual stresses, work hardening and stress concentrations. A confirmation test in terms of fatigue resistance was also carried out to validate the results from which and improvement was obtained.

Failure diagnosis and prevention are crucial areas of interest for the proposal of innovative methods and techniques that can help to increase the availability of industrial machinery and other complex systems. In this work we propose a Knowledge Discovery scheme, based on a Compensatory Fuzzy Logic (CFL), for failure detection and prevention. With an exploratory approach, the proposed methodology includes obtaining a characterization of operating conditions of a system, which can be useful for detecting harmful conditions. As a case of study we obtain data of operating conditions of a direct current (DC) motor. A set of fuzzy predicates are formulated and evaluated using the degrees of membership of the variables of the motor to adequate fuzzy membership functions. The truth values resulting of such evaluations are analyzed in view of the empiric knowledge of failures occurrence of DC motors. The main contribution of this work is to explore the possible advantages of using the compensatory fuzzy logic approach for fuzzy predicate evaluation for fault detection and prevention, which could be applied later to more complex systems.

In this article we present an approach for representing a resilient system which has the capability of absorb perturbations and overcome a disaster. A framework called KOSA is depicted, which is a world that contains a set of knowledge describing objectives, states and actions, linked by a set of rules. This link is expressed by a default theory. First, we define resilience as a relation among states and objectives. Secondly, from a given state, extensions are calculated, which provides information where to go to the future state. The connection, among two or more states creates different configurations that we call trajectories. These connections represent an evolution of the knowledge. Consequently, this reveals the existence of a resilient trajectory. Examples of piloting an airplane are concerned through this paper. Eventually, we present a discrete theoretical behavior of the complete model. Finally the notion of distance among extensions is introduced.

The popularity of some recent applications of AI has given rise to some concerns in society about the risks of AI. One response to these concerns has been the orientation of scientific and technological efforts towards a responsible development of AI. This stance has been articulated from different perspectives. One is to focus on the risks associated with the autonomy of artificial entities, and one way of making this focus operational is the “value-alignment problem” (VAP); namely, to study how this autonomy may become provably aligned with certain moral values. With this purpose in mind, we advocate the characterisation of a problem archetype to study how values may be imbued in autonomous artificially intelligent entities. The motivation is twofold, on one hand to decompose a complex problem to study simpler elements and, on the other, the successful precedents of this artifice in analogous contexts (e.g. chess for cognitive AI, RoboCup for intelligent robotics). We propose to use agent-based modelling of policy-making for this purpose because policy-making (i) constitutes a problem domain that is rich, accessible and evocative, (ii) one may claim that it is an essentially value-drive process and, (iii) it allows for a crisp differentiation of two complementary views of VAP: imbuing values in agents and imbuing values in the social system in order to foster value-aligned behaviour of the agents that act within the system. In this paper we elaborate the former argument, propose a characterisation of the archetype and identify research lines that may be systematically studied with this archetype.

The correct and continuous use of s3 products at home can be beneficial for the environment and at the same time could generate cost savings on bills. An automated home where the user has no interaction may be the most efficient and eco-friendly option, but it is not always the most comfortable option for the user. On the other hand, if the user interacts with the system as he pleases, the system may be wasting energy, so a middle point must be found. If the system learns about the user’s behavior and tries to shape it in order to make it eco-friendlier with the correct motivation, an engagement to this kind of behavior can be achieve. A first approach of the framework is presented, where a classification of the type of consumer is proposed depending on its personality to find his engagement on ecological behavior (EB). First, an artificial neural network is used to get the personality of the consumer. Then a Mamdani inference system is used with the result of the ANN to get an initial level of ecological behavior engagement.

We developed a virtual assistant that enables students to access interactive content adapted for an introductory undergraduate course on artificial intelligence. This chatbot is able to show answers to frequently asked questions in a hierarchical structured manner, leading students by either voice, text or tactile input to the content that better solves their questions and doubts. It was developed using Google Dialogflow as a simple way to generate and train a natural language model. Another convenience of this platform is its ability to collect usage data that is potentially useful for lecturers as learning indicators. The main purpose of this paper is to outline the methodology that guided our implementation so that it can be reproduced in different educational contexts and study chatbots as tools for learning. At the moment, several articles, news and blogs are writing about the potential, implementation and impact chatbots have in general contexts, however there is little to no literature proposing a methodology to reproduce them for educational purposes. In that respect, we developed four main categories as a generic structure of course content and focused on quick implementation, easy updating and generalization. The final product received a general approbation of the students due to its accessibility and well structured data.

Automatic speech recognition involves an understanding of what is being said. It can be audio-based, visual-based, or audio/visual-based according to the type of inputs. Modern speech recognition systems are based on machine learning techniques, such as deep learning. Deep learning systems improve their performance when more data are used to train them. Therefore, data has become one of the most valuable assets in the field of Artificial Intelligence. In this work, we present a methodology to create a database for audio/visual speech recognition. Due to the lack of Spanish datasets, we created a comprehensive Spanish-based speech recognition dataset. For this, we selected hundreds of YouTube videos, found the facial features, and aligned the voice beside text with millisecond accuracy using IBM speech-to-text technology. We split the data into three speaker face angles, where the frontal angle represents the simple case, and right-left angles represent harder cases. As a result, we obtained a dataset of more than 100 thousand samples consisting of a small video with its respective annotation. Our approach can be used to generate datasets on any language by merely selecting videos in the desired language. The database and the source code to create it are open-source.

The article describes general regularities of frequency dynamics of syntactic bigrams and the method used to analyse them. The work objective is to quantitatively estimate the typical rate of change in frequency of syntactic bigrams in English and Russian. Both changes in frequency of words contained in syntactic bigrams and changes in the co-occurrence of these words influence the total rate of changes in frequency of syntactic bigrams. Their contribution to the total rate of frequency changes was estimated using decomposition of the Kullback-Leibler symmetrized divergence. It was also determined to what extent frequencies of the syntactic bigrams respond to major social events. Data on frequencies of syntactic bigrams from the English and Russian sub-corpora of Google Books Ngram were used as a study material. It was found that the regularities of the syntactic bigram usage are similar in English and Russian. The proposed approach can be used in other fields of science.

In this paper the SPI system is presented. SPI is a software tool for planning under uncertainty based on learning Markov Decision Processes. A brief review of some similar tools as well as the scientific basis of factored representations and some of its variants are included. Among these variants are qualitative representations and hybrid qualitative-discrete representations that are the core of the software tool. The functional structure of SPI, which is composed of four main modules, is also described. These modules are: the compiler, the policy server, a format translator and a didactic simulator. The experimental results obtained when testing SPI in a robot navigation domain using different types of representations and different state partitions demonstrated its capability to reduce state spaces.

The pace of daily life causes new perspectives on the idea of new and old, human and technical. Within an era of doubts, hope and fear about technological progress, it’s necessary to understand how human condition has always been related and dependent of technique. It is precisely our relationship to technology that this text focuses on, but also on the many ways IA has influenced, or could transform, our idea of humanity. Through an abductive strategy, several arguments are presented to engage and disengage with both utopias and dystopias.

Nowadays, an eco-friendly way to satisfy the high-energy demand is by the exploitation of renewable sources. Wind energy is one of the viable sustainable sources. In particular, small-scale wind turbines are an attractive option for meeting the high demand for domestic energy consumption since exclude the installation problems of large-scale wind farms. However, appropriate wind resource, installation costs, and other factors must be taken into consideration as well. Therefore, a feasibility study for the setting up of this technology is required beforehand. This requires a decision-making problem involving complex conditions and a degree of uncertainty. It turns out that Bayesian Decision Networks are a suitable paradigm to deal with this task. In this work, we present the development of a decision-making method, built with Decision Bayesian Networks, to assess the use of small-scale wind turbines to meet the high-energy demand considering the available wind resource, installation costs, reduction in CO2 emissions and the achieved savings.

The development of ontologies associated with multi-agent systems provides mechanisms to model and correlate knowledge about the world to agent applications in various domains where simulations involve agents perform social exchanges or resource consumption. Norms that restrict and guide behavior often regulate the actions of agents, and therefore research on normative systems is necessary. Questions about the definition and construction of legal ontologies have been discussed in this research context. Legal ontologies have been proposed to formulate how laws can be modeled to formalize and manage law information in legal systems about regulation like traffic, taxes and other administrative rules. These ontologies have in their modeling formal laws with the purpose of providing information about permission, prohibition, obligation, rewards and punishments for these agents. Firstly, this article presents and discusses researches on ontologies applied in multi-agent systems, more specifically on legal ontologies. Next, we propose a model of legislative ontology related to the Brazilian domain of laws. The ontology of this model is provided by a web service and applied in multi-agent systems through a middleware. We used as a case study the Brazilian legislation that regulates fishing activity in an example scenario where the components simulate the restrictions for fishermen and government agents. Finally, we present how agent actions can be verified on our model and their applicability to the multi-agent systems.

Wind turbine generators are highly desirable to operate autonomously throughout the wind speed range below hurricane conditions. A key requirement to achieve this goal is to be able to control the wind turbine speed using a full-scope feedback control scheme. Currently, wind turbine speed is controlled by modulating the angular position of the rotor blades to catch the required amount of kinetic energy of the wind to produce the desired rotational speed. Typically, conventional PI controllers modulate the blades angular position only for wind speeds in the range from 12 m/s through 25 m/s. This paper introduces a fuzzy speed controller to control autonomously the turbine rotational speed in the whole wind speed range from 0 m/s throughout 30 m/s. After presenting several key concepts about small-scale wind turbines, the design of the fuzzy speed controller based on a TSK fuzzy system is introduced. In this regard, the proposed fuzzy speed controller intelligently extends the scope of control below nominal speed and above trip conditions.

The trend to accelerate the learning process in neural and fuzzy systems has led to the design of hardware implementations of different types of algorithms. In this paper we explore type-2 fuzzy logic systems acceleration, which can be applied to fuzzy logic control methods, signal processing, etc. Due to the three dimensional membership functions in the input of the system, different algorithms for the output processing stage have been developed. In order to have a fast response in type-2 fuzzy logic systems, in this paper we explore the Karnik-Mendel algorithms (KM), which are used to calculate the centroid at the output processing stage of the interval type-2 fuzzy system, through the application of iterative procedures. Because of the computation complexity of the iterative process, we propose a Hardware implementation of the KM algorithm using a High Level Synthesis tool, making possible to explore different types of implementation in order to obtain a significant reduction in computation time, and a reduction in hardware resources.

Since direct-current (DC) drives are commonly used electric drives, it is imperative to improve their operation under sudden torque-load disturbances. Several industrial applications work under torque-load changes that strongly affect the speed response of the motor, thus deteriorating the performance of the DC drive. On the other hand, discontinuous sliding-mode control (SMC) ensures robustness against disturbances and changes in parameters but has certain drawbacks, such as chattering. In this paper, a fuzzy-logic controller (FLC) based on artificial organic networks is proposed to adjust the control signal of the SMC. This control provides a smooth signal that reduces chatter. The Lyapunov stability of the DC motor driven by the proposed SMC with a fuzzy organic controller is tested and stability margins are computed. The proposed controller is validated via simulation results showing an excellent DC-drive performance. In fact, the fuzzy artificial organic controller can adjust the command signal to improve the transitory response of the DC drive. The proposed controller achieves a good performance for speed controllers using brushless DC motors.

Sight is the most critical sense because of its worth in human life; humans use it to guarantee their safeness, to move around, to identify persons, objects, among other activities. The eyes use two kinds of cells in visual perception, rods for luminosity and cones for color. Colorblindness is a mild disability that affects color perception in close to 10% of the world population. Partial solutions for the colorblind include glasses that increase the distance between colors, avoiding confusing regions when suffering mild colorblindness. Other alternatives include special symbols for labeling objects and text descriptions, but there is not a definitive solution. Alternatively, computer vision has developed some assistants for the colorblind based on color classification, including applications that highlight confusing regions or identify colors selected by the user. Recently, artificial intelligence, together with parallel computing, has become a good alternative in vision assistance, but there are several alternatives with different schemes for performing color classification, those include heuristically tuned human proposed equations, computer-generated equations, and Artificial Neural Networks (ANN’s), among others. In this paper, a labeling color assistant for the colorblind is developed using color classification with heuristically (GA and PSO algorithms) tuned proposed equations, genetic programming equations, and ANN’s. As a result of this research, is determined the best structure for color classification, based on the accuracy and processing time in a CUDA kernel, so that be possible a real-time labeling system for the colorblind with full high definition images.

Human limb movement sensing is crucial in different areas of science. In this paper, a method for sensing human limb movement and the subsequent reconstruction in a 3-D plane is described. The sensors used in this task are four Microsoft Kinect, which has depth and RGB cameras. Depth images are processed by artificial vision algorithms to delimit an area where the movements will be performed. In the other hand, RGB images are processed by a Convolutional Neural Network to acquire a series of specific points which correspond to the human body’s joints. A comparison of the proposed algorithm performance is also described. The equations that relate the information in two dimensions are obtained by processing the four sensors are used to generate a skeleton in 3-D.

Worldwide, traffic safety is a strong concern as traffic accidents are one of the leading causes of death. In this context, advanced driver assistance systems (ADAS) and autonomous vehicles are traffic management measures aimed at improving road safety and flow. Automatic detection and recognition of traffic signs are important for intelligent vehicles and ADAS systems. This work proposes a pipeline of digital image processing (DIP) techniques, machine learning (ML), and temporal coherence to perform the detection and recognition of Brazilian traffic signs in videos aiming an application for real-time systems to help in traffic safety and to reduce the number of fatal accidents. We are mainly interested in recognizing signs of speed limit group, no overtaking and obligatory passage, thus our detection considers the traffic sign with a circular shape and red border. For detection, the red color segmentation and the Hough transform are used to find circular regions that will be classified through the SVM algorithm in sign and not sign. For recognition of these signs, the support vector machines (SVM) are used. For speed limit signs the thresholding and contours are used to segment the digits for later classification. Our proposed method achieved an accuracy of 0.82 in detection, an increase of 18% in the number of recognized frames and 0.96 in the recognition stage using temporal coherence.

One of the most important skills for service robots is object manipulation, which is still a challenging task. Since object manipulation is a hard task, it is relevant to know if an object was successfully grasped, avoiding future wrong decisions. Object grasp confirmation is commonly solved by using robotic sensors (infrared, pressure, etc.), but, in many cases, these sensors are not available for all robots. In contrast, depth and RGB sensor are present in almost all service robots. In this work a novel computer vision based method oriented to hand object grasp confirmation is proposed, which uses a deep learning network trained with depth maps. In order to measure the performance of the proposed method, experiments were performed using a single-arm manipulator service robot for both known and unknown objects. Experimental results show that the proposed approach correctly identifies 99% of both classes (object grasped or not grasped) with known objects and $$92\%$$ with unknown objects. The grasping confirmation method was added to the Storing Groceries task, for RoboCup@Home competition, improving its time performance.

In this paper, we propose a monocular vision-based system that uses a MobileNet-SSD CNN for obstacle detection and collision avoidance in GPS-denied outdoor environments. This framework consists of two processes carried out simultaneously in a frame-to-frame basis: (1) an obstacle detector and classifier using a lightweight convolutional neural network with a UAV monocular onboard camera for real-time mobile systems; (2) a collision avoidance algorithm with a proportional controller responsible for the autonomous flight in GPS-denied outdoor environments. However, because object detection and classification are computationally intensive tasks, the processing is carried out off-board on a ground control station that receives online imagery and data of the UAV during the autonomous flight. The novel aspects in this work are related to the capacity of the system to detect and avoid obstacles in real-time with computationally low range hardware without GPU. We exploit public datasets meant for other purposes and carefully selected images to build a new lightweight dataset to train the CNN. Further, the output imagery data is used by a proportional controller that communicates back to the vehicle to evaluate a possible obstacle avoidance trajectory and execute it if necessary. We carried out evaluations and flights in real scenarios with multiple obstacles such as vehicles, people, bicycles, and trees for autonomous flights in GPS-denied outdoor environments with promising results.

In this paper, a method for automatic channel selection of EEG signals acquired during the execution of lower limb movements is presented; for this method the hip and knee joint torques are measured. The method is based on maximizing both the percentage of prototypes extracted and the relative dispersion of its respective torques using a genetic algorithm. The prototyping is made with clusterized KNN, a proposed modification of the K-nearest neighbors algorithm, and the dispersion is computed as the ratio of interquartile ranges (IQR) between original and resulting torques. Results show that frequent channels are consistent with those known to be activated during motor tasks and that additional channels, needed for extracting relevant information from the data, vary from subject to subject. Extracted data can be used as new inputs for later regression tasks and for further analysis in order to characterize neural processes.

In this work, a prototype of an autonomous topographic metrology system that uses a self-leveling Stewart platform system is presented, with the objective of evaluating the angular uncertainty of the azimuth plane adjustment process, which would disperse an associated collimation instrument. The self-leveling process of the prototype is achieved by means of two mutually independent four-bar kinematic chains that regulate the inclination of the platform on the “x” and “y” axes. The control of the prototype is based on the regulation of the motive source, of the kinematic chains, by means of a servomotor coupled to one of the fixed articulations and an accelerometer. The comparison of the angular error of adjustment of the calculated azimuth plane and that measured independently in an array of orthogonal toroid levels shows that the error changes as a function of the initial disturbed position and converges to a fixed value that depends on the accuracy of the source controller motor, the resolution in the range of the sensor and the alignment of the links and articulations of the kinematic chain.

Service robots have to face task diversity, large, uncertain and partially observable environments, which are inherent aspects of the domestic domain and increase the task planning problem’s complexity. Thus, in an attempt to overcome these challenges, in this work a task planning architecture for service robotics is proposed, which integrates a knowledge base approach with partially observable Markov decision processes (POMDP), and is constituted by three main components: (a) a knowledge base, (b) a POMDP construction module and (c) a task controller. Through a knowledge representation scheme, domain relevant information is exploited to define useful sub-regions in the planning search space. Once the search space is segmented, local POMDP policies are computed for each sub-region, then, a graph-based policy for the main task is built as a collection of these policies, for which the controller will determine the order in which they will be executed. Additionally, our architecture is able to integrate new functionalities as the robot is endowed with them. For evaluation purposes, a mobile robot navigation problem was used as case study to test our architecture, which shows the advantages of using domain specific knowledge in a task planning problem.

Navigation, localization, and mapping are challenging tasks that any mobile service robot needs to solve. Given that this type of robots generally navigate in 2D planar environments, a common and highly effective solution is laser-based mapping (SLAM) and navigation. Unfortunately, due to their incapability to detect obstacles outside of a single plane view, these algorithms are affected by irregular obstacles in the environment; even more when there are dynamic obstacles. To address this problem, we propose a method to integrate data from a 2D laser range finder (LRF) and an RGB-D camera. In this paper, our goal is to enrich a 2D grid-based map by extracting and processing a depth image from an RGB-D camera, fusing this with the information from the LRF. To test the algorithm, we set up five different scenarios in which pure laser navigation would be an ambitious task. Comparative results between pure LRF and LRF + RGB-D navigation are presented. In spite of the simplicity of the method, results show a significant improvement in the robot’s navigation, making it more robust in complex, dynamic environments.

This study presents an adaptive controller based on neuro-fuzzy networks and stochastic approximation techniques. The algorithm assumes that the mathematical model of the plant is unknown. An adaptive Fuzzy Rule Emulated Network (FREN) structure is implemented as the main controller. While, a modified version of the Simultaneous Perturbation Stochastic Approximation (SPSA) technique is added as the adaptation algorithm, which estimates the gradient of the plant with respect to the control effort. The proposed FREN+SPSA performance for position control is compared to conventional FREN and classical PID controllers. Experimental tests were performed on a cartesian robotic system, regulating the frequency of a DC motor to follow a desired trajectory. Experimental results show better performance of the proposed FREN+SPSA controller than the conventional FREN and PID controller.

This paper presents a path planning strategy for robotic manipulators based on genetic algorithms, dual quaternions and artificial potential field, designing a multi-objective function that allow trajectories be planned avoiding collisions in the workspace and singularity-free kinematic restrictions for manipulators as an optimization problem, satisfying position and orientation conditions. Its analysis is based on the problem of generating a trajectory followed by a sequence of coordinated movements capable of moving the manipulator to perform tasks in the workspace, the problem is not only generated these movements, but also implement strategies that define the path with tools that are easy to implement and avoid obstacles autonomously. Robot kinematics solved by dual quaternion can be used to combine translation with orientation on robotic manipulators in a systematic way, simplifying calculation operations compatible with conventional methods. The artificial potential field approach has been extended to collision avoidance for all manipulator links. A genetic algorithm is used to solve the problem, which the fitness of the problem can be measured by a multi-objective function that involves the distance between the initial and desired position/orientation, minimum joint displacement, dual quaternion configuration, the use of attraction potential to the goal and a repulsion potential to the obstacles and its own links. This method has been implemented in MatLab© for an ABB© IRB1600 robot. Collision avoidance demonstrations have been performed by simulating equipment and static objects in the robot’s workspace.

A three dimensional (3-D) digital image emerges as a straightforward extension of a two dimensional (2-D) digital image. A 3-D digital image can be obtained by digitizing the 3-D space in which one or more objects of interest can be contained. From each object in the digital image, several features describing their geometry and topology can be computed. One of these features is the Euler number. An alternative method to compute the Euler number of a 3-D digital object (image) in terms of a codification of the vertices of the object voxels is described. The set of formal propositions baseline of the proposal operation are provided, demonstrated and numerically validated with simple objects. Examples with images of different complexity show the applicability of the proposal. The proposed method emerges as an extension of the proposal introduced for the 2-D case in [21] and as alternative of the formulation well described in [22].

This paper presents the design and analysis of different controller schemes for a three wheeled omnidirectional robot, that is, a robot which can be driven by the control of three independent velocity variables: over $$ x $$ and $$ y $$ linear directions and a rotational direction to perform complex motions. Conventional Proportional, Proportional-Integral, Fuzzy and Fuzzy PI controllers are designed based on the kinematic behavior of the holonomic mobile omnidirectional robot. Velocities of the Robot are estimated by using an odometer module. Also, the implementation performed in Simulink tool of Matlab of the different schemes is presented. The control objective for these designs is to drive the two linear command velocities of the robot and to do it goes from an initial position to a final position without obstacles. Finally, a comparative analysis of the studied controllers in time and distance is performed for determining the performance using a Robotino simulation tool of Festo, which provides an uncertainty environment for the robot control.

Bio-inspired robots have chosen to propose novel developments aiming to inhabit and interact complex and dynamic environments. Bio-inspired four-legged robots, typically inspired on animal locomotion, provide advantages on mobility, obstacle avoidance, energy efficiency and others. Balancing is a major challenge when legged robots require to move over uncertain and sharp terrains. It becomes of particular importance to solve other locomotion tasks such as walking, running or jumping. In this paper, we present a preliminary study on the modeling and control balance design of a bio-inspired four-legged robot for standing on its aligned legs in a straight line. The proposed robot is loosely inspired on the bio-mechanics of the chameleon. Thus, a mathematical modeling, simulation, intelligent control strategy, prototype implementation and preliminary results of control balance in our robot are presented and discussed.

The ability to know the pose of a drone in a race track is a challenging task in Autonomous Drone Racing. However, to estimate the pose in real-time and at high-speed could be fundamental to lead an agile flight aiming to beat a human in a drone race. In this work, we present the architecture of a CNN to automatically estimates the drone’s pose relative to a gate in a race track. Due to the challenge in ADR, various proposals have been developed to address the problem of autonomous navigation, including those works where a global localisation approach has been used. Despite there are well-known solutions for global localisation such as visual odometry or visual SLAM, these methods may become expensive to be computed onboard. Motivated by the latter, we propose a CNN architecture based on the Posenet network, a work-oriented to perform camera relocalisation in real-time. Our contribution relies on the fact that we have modified and re-trained the Posenet network to adapt it to the context of relative localisation w.r.t. a gate in the track. The ultimate goal is to use our proposed localisation approach to tackle the autonomous navigation problem. We report a maximum speed of up to 100 fps in a low budget computer. Furthermore, seeking to test our approach in realistic scenarios, we have carried out experiments with small gates of 1 m of diameter under different light conditions.