Progress in Artificial Intelligence

A three-dimensional computational model based on finite elements was developed to evaluate the thermal behaviour of composite slabs with steel deck exposed to a standard fire. The resulting numerical temperatures are then used to obtain a new analytical method, which is an alternative to the simplified method provided by the standard, to accurately determine the temperatures at the reinforcing bars (rebar). The fitting of the analytical model to the numerical data was done by solving a linear least squares problem using the singular value decomposition. The resulting formula fits very well the numerical data, allowing to make predictions of the temperature in the rebar with an approximation error equal to zero and an estimating error at least 77% lower than that obtained with the proposal included in the standard.

In a warehouse environment, tasks appear dynamically. Consequently, a task management system that matches them with the workforce too early (e.g., weeks in advance) is necessarily sub-optimal. Also, the rapidly increasing size of the action space of such a system consists of a significant problem for traditional schedulers. Reinforcement learning, however, is suited to deal with issues requiring making sequential decisions towards a long-term, often remote, goal. In this work, we set ourselves on a problem that presents itself with a hierarchical structure: the task-scheduling, by a centralised agent, in a dynamic warehouse multi-agent environment and the execution of one such schedule, by decentralised agents with only partial observability thereof. We propose to use deep reinforcement learning to solve both the high-level scheduling problem and the low-level multi-agent problem of schedule execution. The topic and contribution is relevant to both reinforcement learning and operations research scientific communities and is directed towards future real-world industrial applications.

The São Paulo Court of Justice has the highest number of lawsuits of all courts. The lawsuits are composed of raster-scanned documents enclosed in unstructured volumes, of which some are unreadable document images. Natural Language Processing techniques fail to extract from some of these documents due to the low quality of images. This article proposes a methodology to automatize the retrieval of document images from lawsuit databases based on the contents of the images. We developed a hybrid algorithm for feature extraction from document images and used a distance metric to retrieve similar images. The TJSP’s database was used to validate our proposal, resulting in a system that allows finding similar images with an accuracy above eighty percent.

Natural Language Processing techniques usually fail to classify low quality lawsuit document images produced by a flatbed scanner or fax machine or even captured by mobile devices, such as smartphones or tablets. As the courts of justice have many lawsuits, the manual detection of classification errors is unfeasible, favouring fraud, such as using the same payment receipt for more than one fee. An alternative to classifying low-quality document images is visual-based methods, which extract features from the images. This article proposes classification models for lawsuit document image processing using transfer learning to train Convolutional Neural Networks most quickly and obtain good results even in smaller databases. We validated our proposal using a TJSP dataset composed of 2,136 unrecognized document images by Natural Language Processing techniques and reached an accuracy above 80% in the proposed models.

Precedents constitute the starting point of judges’ reasoning in national legal systems. Precedents are also an essential input for case-based reasoning (CBR) methodologies. Although considerable research has been done on CBR applied to legal practice, the precedent retrieval techniques are a relatively new and unexplored field of AI & Law. Only a few works have tested or developed methods for identifying such previous similar cases. This work uses text mining (TM), natural language processing (NLP), and data visualization methods to provide a semi-automated rapid literature review and identify how justice courts and legal practitioners may use AI to retrieve similar cases. Based on Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA), automation techniques were used to expedite the literature review. In this study, we confirmed the feasibility of automation tools for expediting literature reviews and provided an overview of the current research state on legal precedents retrieval.

The European Draft Regulation on Artificial Intelligence was presented in April 2021 with ambitious aims: to be a far-reaching regulation aimed to guarantee the highest level of protection for ‘Union values, fundamental rights and principles’, and at the same time promote innovation. However, several possible drawbacks are likely to jeopardize these ambitious purposes: the risk-based approach, grounded on different levels of risks, is imprecise; the conformity assessment has loopholes and might not be as protective as originally envisaged; some requirements are difficult to meet; the contextualization of this regulation within the European legal framework gives raise to overlaps and potential conflicts; the rights of the ones affected by this technology are not properly safeguarded; the European innovation boost might suffer a major setback. Instead of a solution for artificial intelligence, the European Union might have created a new problem.

Autonomous vehicles have profound implications for laws governing police, searches and seizures, and privacy. Complicating matters, manufacturers are developing these vehicles at varying rates. Each level of vehicle automation, in turn, poses unique issues for law enforcement. Semi-autonomous (Levels 2 and 3) vehicles make it extremely difficult for police to distinguish between dangerous distracted driving and safe use of a vehicle’s autonomous capabilities. Fully autonomous (Level 4 and 5) vehicles solve this problem but create a new one: the ability of criminals to use these vehicles to break the law with a low risk of detection. How and whether we solve these legal and law enforcement issues depends on the willingness of nations to adapt legal doctrines. This article explores the implications of autonomous vehicle stops and six possible solutions including: (1) restrictions on visibility obstructions, (2) restrictions on the use and purchase of fully autonomous vehicles, (3) laws requiring that users provide implied consent for suspicion-less traffic stops and searches, (4) creation of government checkpoints or pull-offs requiring autonomous vehicles to submit to brief stops and dog sniffs, (5) surveillance of data generated by these vehicles, and (6) opting to do nothing and allowing the coming changes to recalibrate the existing balance between law enforcement and citizens.

Political bill comments published in digital media may reveal the issuer’s stances. Through this, we can identify and group the polarity of these public opinions. The automatic stance detection task involves viewing the text and the target topic. Due to the diversity and emergence of new bills, the challenge approached is to estimate the polarity of a new topic. Thus, this paper evaluates cross-target stance detection with many-to-one approaches in a collected Portuguese dataset of the political pool from the Brazilian Chamber of Deputies website. We proposed a new corpus for the bills’ opinion domain and tested it in several models, where we achieved the best result with the mBERT model in classification with the joint input topic and comment method. We verify that the mBERT model successfully handled cross-target tasks with this corpus among the tested algorithms.

The present study analyzes the role of algorithms as an integral part of artificial intelligence (AI) and checks whether it is possible to interfere in the construction of the algorithmic system to generate benefits for platforms and gig workers or if the so-called complexity of the algorithms is an absolute impediment factor. to fulfill that ideal. The research methodology will be based on a qualitative approach, through bibliographic research in law, sociology, and programming. The research delves into the field of research on the relationship between algorithms and work aimed at creating algorithms oriented to providing decent worker-centered digital work and, for that reason, contributes to the emerging literature on algorithmic work.

The admissibility of AI systems that focus on determining the measure of punishment must be analyzed in light of ECHR and ECtHR jurisprudence. We cannot live the AI evolution in a passive way and is a matter of time before it is adopted in the criminal justice system. The following paper focuses on the respect for fundamental rights as a filter of such instruments. We highlight the right to a fair trial (article 6), the principle of legality (article 7) and the prohibition of discrimination (article 14). Predictability can justify the adoption of predictive tools, ensuring fairer decision. On the other hand, explainability is an essential requirement that has been developed by explainable artificial intelligence. There are several AI models that must be adopted depending on domain and intended purpose. Only a multidisciplinary approach can ensure the compatibility of such instruments with ECHR. Thus, a confrontation between legal and engineering concepts is essential so that we can design tools that are more efficient, fairer and trustable.

This work aims to systematize the knowledge on emerging Intelligent Information Retrieval (IIR) practices in scenarios whose context is similar to the field of tax law. It is a part of a project that covers the emerging techniques of IIR and its applicability to the tax law domain. Furthermore, it presents an overview of different approaches for representing legal data and exposes the challenging task of providing quality insights to support decision-making in a dedicated legal environment. It also offers an overview of the related background and prior research referring to the techniques for information retrieval in legal documents, establishing the current state-of-the-art, and identifying its main drawbacks. A summary of the most appropriate technologies and research approaches of the technologies that apply artificial intelligence technology to help legal tasks is also depicted.

The region of interest (RoI) identification has a significant potential for yielding information about relevant histological features and is imperative to improve the effectiveness of digital pathology in clinical practice. The typical RoI is the stratified squamous epithelium (SSE) that appears on relatively small image areas. Hence, taking the entire image for classification adds noise caused by irrelevant background, making classification networks biased towards the background fragments. This paper proposes a novel approach for epithelium RoI identification based on automatic bounding boxes (bb) construction and SSE extraction and compares it with state-of-the-art histology RoI localization and detection techniques. Further classification of the extracted epithelial fragments based on DenseNet made it possible to effectively identify the SSE RoI in cervical histology images (CHI). The design brings significant improvement to the identification of diagnostically significant regions. For this research, we created two CHI datasets, the CHI-I containing 171 color images of the cervical histology microscopy and CHI-II containing 1049 extracted fragments of microscopy, which are the most considerable publicly available SSE datasets.

Audio classification using breath and cough samples has recently emerged as a low-cost, non-invasive, and accessible COVID-19 screening method. However, a comprehensive survey shows that no application has been approved for official use at the time of writing, due to the stringent reliability and accuracy requirements of the critical healthcare setting. To support the development of Machine Learning classification models, we performed an extensive comparative investigation and ranking of 15 audio features, including less well-known ones. The results were verified on two independent COVID-19 sound datasets. By using the identified top-performing features, we have increased COVID-19 classification accuracy by up to 17% on the Cambridge dataset and up to 10% on the Coswara dataset compared to the original baseline accuracies without our feature ranking.

The diagnosis of stroke, a leading cause of death in the world, using computed tomography (CT) scans, makes it possible to assess the severity of the incident and to determine the type and location of the lesion. The fact that the brain has two hemispheres with a high level of anatomical similarity, exhibiting significant symmetry, has led to extensive research based on the assumption that a decrease in symmetry is directly related to the presence of pathologies. This work is focused on the analysis of the symmetry (or lack of it) of the two brain hemispheres, and on the use of this information for the classification of computed tomography brain scans of stroke patients. The objective is to contribute to a more precise diagnosis of brain lesions caused by ischemic stroke events. To perform this task, we used a siamese network architecture that receives a double two-dimensional image of a CT slice (the original and a mirrored version) and a label that reflects the existence or not of a visible stroke event. The network then extracts the relevant features and can be used to classify brain-CT slices taking into account their perceived symmetry. The best performing network exhibits an average accuracy and F1-score of 72%, when applied to CT slices of previously unseen patients, significantly outperforming two state-of-the-art convolutional network architectures, which were used as baselines. When applied to slices chosen randomly, that may or may not be from the same patient, the network exhibits an accuracy of 97%, but this performance is due in part to overfitting, as the system is able to learn specific features of each patient brain.

In recent years, hospital overcrowding has become a crucial aspect to take into consideration in inpatient management, which may negatively affect the quality of service provided to the patient. Inpatient management aims, through efficient planning, to maximise the availability of beds and conditions for the patient, considering cost rationalisation. In this way, this research has allowed the prediction of the length of stay (LOS) of each patient in the Internal Medicine specialty, with acuity, considering their demographic data, the information collected at the time of admission and clinical conditions, which may help health professionals in carrying out more assertive planning. For this study, were used data sets from the Centro Hospitalar do Tâmega e Sousa (CHTS), referring to a 5-year period, 2017 to 2021. The GB model achieved an accuracy of ≈96% compared to the DT, RF and KNN, proving that Machine Learning (ML) models, using demographic information simultaneously with the route taken by the patient and clinical data, such as drugs administrated, exams, surgeries and analyses, introduce a greater predictive capacity of the LOS.

Transthyretin Familial Amyloid Polyneuropathy (TTR-FAP) is a neurological genetic illness that inflicts severe symptoms after the onset occurs. Age of onset represents the moment a patient starts to experience the symptoms of a disease. An accurate prediction of this event can improve clinical and operational guidelines that define the work of doctors, nurses, and operational staff. In this work, we transform family trees into compact vectors, that is, embeddings, and handle these as input features to predict the age of onset of patients with TTR-FAP. Our purpose is to evaluate how information present in genealogical trees can be transformed and used to improve a regression-based setting for TTR-FAP age of onset prediction. Our results show that by combining manual and graph-embeddings features there is a decrease in the mean prediction error when there is less information regarding a patient’s family. With this work, we open the way for future work in representation learning for genealogical data, enabling a more effective exploitation of machine learning approaches.

Healthcare environments are generating a deluge of sensitive data. Nonetheless, dealing with large amounts of data is an expensive task, and current solutions resort to the cloud environment. Additionally, the intersection of the cloud environment and healthcare data opens new challenges regarding data privacy.With this in mind, we propose MedCloudCare (MCC), a healthcare application offering medical image viewing and processing tools while integrating cloud computing and AI. Moreover, MCC provides security and privacy features, scalability and high availability. The system is intended for two user groups: health professionals and researchers. The former can remotely view, process and share medical imaging information in the DICOM format. Also, it can use pre-trained Machine Learning (ML) models to aid the analysis of medical images. The latter can remotely add, share, and deploy ML models to perform inference on DICOM images.MCC incorporates a DICOM web viewer enabling users to view and process DICOM studies, which they can also upload and store. Regarding the security and privacy of the data, all sensitive information is encrypted at rest and in transit. Furthermore, MCC is intended for cloud environments. Thus, the system is deployed using Kubernetes, increasing the efficiency, availability and scalability of the ML inference process.

Every year thousands of people get their diagnoses wrongly, and several patients have their health conditions aggravated due to misdiagnosis. This problem is even more challenging when the list of possible diseases is long, as in a general medicine speciality. The development of Artificial Intelligence (AI) medical diagnosis systems could prevent misdiagnosis when clinicians are in doubt. We developed an AI system to help clinicians in their daily practice. They could consult the system to get an immediate opinion and diminish waiting times in triage services since this task could be carried out with minimal human interaction. Our method relies on Machine Learning techniques, more precisely on Active Learning and Neural Networks classifiers. To train this model, we used a data set that relates symptoms to several diseases. We compared our models with other models from the literature, and our results show that it is possible to achieve even better performance with much less data, mainly because of the contribution of the Active Learning component.

Anaemia is often caused by a nutritional problem or by genetic diseases. The world prevalence of anaemia is estimated to be 24.8%, strengthening the need for appropriate discrimination methods between the different types of this disease, an essential step to choosing the best treatment and offering genetic counselling. Several indexes based on haematological features have been proposed to address the challenge of microcytic anaemias classification. However, they have not been tested extensively nor optimised for different countries. Here we test existing binary classification indexes in a Portuguese sample of 390 patients diagnosed with microcytic anaemia and propose novel classification methods to discriminate between the disease classes. We show that existing indexes for the binary classification of Iron Deficiency Anaemia (IDA) and $$\beta $$ β -thalassaemia trait are well adapted to this sample, with RDWI (red cell distribution width index) achieving a median accuracy of 95.4%, a performance we were also able to achieve using Random Forests. The multi-class classification was also developed to discriminate between three microcytic anaemias and healthy subjects, presenting a median accuracy of 93.0%. In addition, we developed a semi-automatic method to identify outliers, which were shown to correspond to subjects with unexpected features given their class and who may correspond to clinical misclassification that require further analysis. The results illustrate that it is possible to achieve excellent performance using just the information obtained through an affordable Complete Blood Count test, thus highlighting the potential of artificial intelligence in classifying microcytic anaemias.

During the last years, deep learning has been used intensively in medical domain making considerable progress in the diagnosis of diseases from radiology images. This is mainly due to the availability of proven algorithms on several computer vision tasks and the publicly accessible medical datasets. However, most approaches that apply deep learning techniques to radiology images transform these images into a format that conforms with the inputs of conventional learning algorithms and deal with the dataset as a set of 2D independent slices instead of volumetric images. In this work we deal with the problem of preparing DICOM CT scans as 3D images for a machine learning/deep learning architecture. We propose a general preprocessing pipeline composed of four stages for volumetric images processing followed by a 3D CNN architecture for 3D images classification. The proposed pipeline is evaluated through a case study for COVID-19 detection from chest CT scans. Experiment results demonstrate the effectiveness of the proposed preprocessing operations.

Complex optimization problems are often associated to large search spaces and consequent prohibitive execution times in finding the optimal results. This is especially relevant when dealing with dynamic real problems, such as those in the field of power and energy systems. Solving this type of problems requires new models that are able to find near-optimal solutions in acceptable times, such as metaheuristic optimization algorithms. The performance of these algorithms is, however, hugely dependent on their correct tuning, including their configuration and parametrization. This is an arduous task, usually done through exhaustive experimentation. This paper contributes to overcome this challenge by proposing the application of sequential model algorithm configuration using Bayesian optimization with Gaussian process and Monte Carlo Markov Chain for the automatic configuration of a genetic algorithm. Results from the application of this model to an electricity market participation optimization problem show that the genetic algorithm automatic configuration enables identifying the ideal tuning of the model, reaching better results when compared to a manual configuration, in similar execution times.

In the upcoming years, European countries have to make a strong bet on solar energy. Small photovoltaic systems are able to provide energy for several applications like housing, traffic and street lighting, among others. This field is expected to have a big growth, thus taking advantage of the largest renewable energy source existing on the planet, the sun. This paper proposes a computational model able to simulate the behavior of a stand-alone photovoltaic system. The developed model allows to predict PV systems behavior, constituted by the panels, storage system, charge controller and inverter, having as input data the solar radiation and the temperature of the installation site. Several tests are presented that validates the reliability of the developed model.

Energy management in buildings can be largely improved by considering adequate forecasting techniques to find load consumption patterns. While these forecasting techniques are relevant, decision making is needed to decide the forecasting technique that suits best each context, thus improving the accuracy of predictions. In this paper, two forecasting methods are used including artificial neural network and k-nearest neighbor. These algorithms are considered to predict the consumption of a building equipped with devices recording consumptions and sensors data. These forecasts are performed from five-to-five minutes and the forecasting technique decision is taken into account as an enhanced factor to improve the accuracy of predictions. This decision making is optimized with the support of the multi-armed bandit, the reinforcement learning algorithm that analyzes the best suitable method in each five minutes. Exploration alternatives are considered in trial and test studies as means to find the best suitable level of unexplored territory that results in higher accumulated rewards. In the case-study, four contexts have been considered to illustrate the application of the proposed methodology.

GPU servers have been responsible for the recent improvements in the accuracy and inference speed of the object detection models targeted to autonomous driving. However, its features, namely, power consumption and dimension, make its integration in autonomous vehicles impractical. Hybrid FPGA-CPU boards emerged as an alternative to server GPUs in the role of edge devices in autonomous vehicles. Despite their energy efficiency, such devices do not offer the same computational power as GPU servers and have fewer resources available. This paper investigates how to deploy deep learning models tailored to object detection in point clouds in edge devices for onboard real-time inference. Different approaches, requiring different levels of expertise in logic programming applied to FPGAs, are explored, resulting in three main solutions: utilization of software tools for model adaptation and compilation for a proprietary hardware IP; design and implementation of a hardware IP optimized for computing traditional convolutions operations; design and implementation of a hardware IP optimized for sparse convolutions operations. The performance of these solutions is compared in the KITTI dataset with computer performances. All the solutions resort to parallelism, quantization and optimized access control to memory to reduce the usage of logical FPGA resources, and improve processing time without significantly sacrificing accuracy. Solutions probed to be effective for real-time inference, power limited and space-constrained purposes.

Spatial data on human activity, including mobile phone data, has the potential to provide patterns of how the citizens use the urban space. The availability of this data boosted research on city dynamics and human behavior. In this context, we address the question: Can we generate a sufficiently accurate picture of the main places of individuals from highly noisy and sparse data generated by mobile phone operators?This paper studies different kinds of anonymized mobile phone data and proposes a model, that uses a density-based clustering algorithm to obtain the geographic profile of customers, by identifying their most visited locations at the antenna level. The individual routine, such as sleeping period and work hours, is dynamically identified according to slots of minimums of activity in the network. Then, based on those slots, areas of Home, Second Home, and Work are inferred. Ground truth is used to validate and evaluate the model.

We consider classical vehicle routing problems with customer costs, vehicle feasibilities, driver profits, and driver responsiveness. We motivate a new template for these new problems, which first returns some feasible matching between drivers and customers and then some feasible plan for routing the vehicles through their matched locations. Thus, by using this template, we show that bounded equitability for drivers and Pareto optimality for customers can always be achieved in isolation but not always in combination. Finally, we give fixed-parameter tractable routing algorithms for fleet equitability and fleet efficiency.

In Cooperative Intelligent Transport Systems, road users and traffic managers share information for coordinating their actions to improve traffic efficiency allowing the driver to adapt to the current traffic situation. Its effectiveness, however, depends on i) the user’s decision-making process, which is the main source of uncertainty in any mobility system, and on ii) the ability of the infrastructure to communicate timely and reliably. To cope with such a complex scenario, this paper proposes a game theory perspective based on the n-Person Prisoner’s Dilemma as a metaphor to represent the uncertainty of cooperation underlined by communication infrastructures in traveller information systems. Results highlight a close relationship between the emergence of cooperation and the network performance, as well as the impact of the communication failure on the loss of cooperation sustainment, which was not recovered after the system was re-established.

The neat and clean surrounding is the main driving force for any city to be called a smart city. In order to address current societal and business challenges, the objective is to provide a solution to enable collection-on-demand of wastes by connecting waste data and users/customers with the waste management system. In that context, the focus is to improve the waste collection process in terms of collection cost, collection time, and CO2. Within the overall objective, an important goal that needs to be solved is waste collection on demand and the present paper addresses this by tackling the optimization problem related to the routing. Application of the presented solution to a case study with real data collected in the municipality of Ålesund, Norway, is presented. This study also shows a comparison of three popular optimization algorithms for solving vehicle routing problems (VRP) and multiple vehicle routing problems (MVRP), to identify a suitable algorithm for the case study, introducing a data-driven model. Five constraints with alternative objectives of distance and cost minimization are considered.

Recent advances in wireless communication technologies have led to the rapid development of branches of engineering, such as those related to the Internet of Things (IoT) paradigm. IoT interconnects devices with the intention of adding value or reducing costs in production processes. In turn, many productive sectors are benefiting from the advances being made in this field, including the agricultural sector. The IoT for Low-power wide-area network (LPWA) is a perfect fit for sectors whose environments are remote (and therefore have limited access to the power grid) and whose facilities may be located at long distances from each other. This research therefore proposes, the design of a LoRaWAN communications module as part of a modular architecture, compatible with environmental parameter measuring devices and irrigation system controllers. The purpose of this module is to improve the management of agricultural facilities and, therefore, boost the competitiveness of companies in this sector.

Lifestyle, poor diet, stress, among other factors, strongly contribute to aggravate people’s health problems, such as diabetes and high blood pressure. Some of these problems could be avoided if some of the essential recommendations for the practice of a healthy lifestyle were followed. The paper proposes a solution designed for diabetic people to find restaurants nearby that are more suitable for their health needs. A diabetic-friendly feature that will use a set of criteria, built through a Multi-Agent System (MAS) that using the user preferences initially recorded, will provide the user with three category recommendations that potentially benefit the user lifestyle and health. The solution proposes the use of Case-Based Reasoning algorithm to enable the solution to evolve and improve in each interaction with the user. Sentiment Analysis was also used for identifying the restaurant reviews score, since this is one of the defined criteria for the solution.

Diabetes is a chronic metabolic disease characterized by high blood sugar levels, which over time leads to body complications that can affect the heart, blood vessels, eyes, kidneys, and nerves. To control this disease, the use of applications for tracking and monitoring vital signs have been used frequently. These support systems improve their quality of life and prevent exacerbations, however they cannot help with nutritional control, so several patients with this disease still use the counting carbohydrates method, but this process is not available to everyone and is a time-consuming and not very rigorous method. This study evaluates three approaches including Support Vector Machine, Convolution Neural Network, and a pre-trained Convolution Neural Network called MobileNetV2, to choose the algorithm with the best performance in meals recognition and makes the control nutritional task more quickly, accurately, and efficiently. The results showed that the pre-trained Convolution Neural Network is the best choice for recognizing meals from an image, with an accuracy of 99%.

Emotion recognition relies heavily on physiological responses and facial expressions. Using current technology, it is possible to use a set of measuring instruments to create a complex sensory network that is able to acquire physiological response and recognize emotions based on the facial features of the user by using facial recognition software. It is also important to automate these devices and the acquisition of sensory data to make it easy to use without the need of further user input. The aim of this work is to describe an experiment, where physiological functions are collected using low-cost, common and non-invasive Internet of Things (IoT) devices, and the sensory data is automatically sent to a server for further processing. From the measured values a dataset is created and in order to understand the data, descriptive statistics is used. The data are visualized with the help of well-known Python libraries such Pandas, Matplotlib or Seaborn.

The aim of this study is to examine the effective communication process during shift changes in a transitional care unit; communication should be clear, concise and relevant. The caregivers should be able to understand their mates needs and wants, and they should understand what is happening. Indeed, it will be study if there is inconsistency in the communication process during shift change, since there are some factors that could be affecting the consistency, such as inadequate time for communication, lack of awareness about the needs of other team members and unclear expectations from managers, a process carried out by interviewing employees who have experience in transitional care. The interviewees are asked about their past experiences and what they think are the most important points for effective communication. No doubt this article will discuss the communication challenges and opportunities that arise when a caregiver transfers care from one shift to the next, with effective communication being the key. On the other hand, in this work is also set a Logic Programming based framework that nurses may use to optimize their insight into how caregivers may make their communication more effective by understanding the needs of both parties involved in the shift change process.

Algorithmic composition (AC) refers to the process of creating music by means of algorithms, either for realising music entirely composed by a computer or with the help of a computer. In this paper, we report on the development of the system $$\textsf{PAUL}$$ PAUL , an algorithmic composer for the automatic creation of short pieces of classical piano music, based on a neural-network architecture. The distinguishing feature of $$\textsf{PAUL}$$ PAUL is that it allows to specify the desired complexity of the output piece in terms of an input parameter, which is a central aspect towards the designated future usage of $$\textsf{PAUL}$$ PAUL as being part of a tutoring system teaching piano students how to sight-read music. $$\textsf{PAUL}$$ PAUL employs a long short-term memory (LSTM) neural network to produce the lead track and a sequence-to-sequence neural network for the realisation of the accompanying track. Although $$\textsf{PAUL}$$ PAUL is still work-in-progress, the obtained results are of reasonable to good quality. In a small-scale study, evaluating the specified vs. the perceived complexity of different pieces generated by $$\textsf{PAUL}$$ PAUL , a clear correlation is observable.

The model-based reinforcement learning paradigm, which uses planning algorithms and neural network models, has recently achieved unprecedented results in diverse applications, leading to what is now known as deep reinforcement learning. These agents are quite complex and involve multiple components, factors that create challenges for research and development of new models. In this work, we propose a new modular software architecture suited for these types of agents, and a set of building blocks that can be easily reused and assembled to construct new model-based reinforcement learning agents. These building blocks include search algorithms, policies, and loss functions (Code available at https://github.com/GaspTO/Modular_MBRL ).We illustrate the use of this architecture by combining several of these building blocks to implement and test agents that are optimized to three different test environments: Cartpole, Minigrid, and Tictactoe. One particular search algorithm, made available in our implementation and not previously used in reinforcement learning, which we called averaged minimax, achieved good results in the three tested environments. Experiments performed with our implementation showed the best combination of search, policy, and loss algorithms to be heavily problem dependent.

This work introduces an interactive machine teaching approach that teaches classification tasks. But instead of assuming perfect knowledge about the learner as most machine teaching approaches do, our adaptive approach—Feature Importance Teaching (FIT)—chooses the samples to show based on a model of the learner updated online using feedback about the weights attributed to the features. We run simulations where there is a mismatch on the prior knowledge and learning model of the student and the ones assumed by the teacher. The results have shown that our teaching approach can mitigate this mismatch and lead to significantly faster learning curves than the ones obtained in conditions where the teacher randomly selects the samples or does not consider this kind of feedback from the student. We tested using data sets from two different application domains and the conclusions were the same. We also tested FIT when the student provides only the most important feature and it still outperformed the other approaches considered. We finally conducted a study with real human users, which confirmed the results obtained in the simulations.

Monitoring the marine ecosystem can be done via observations (either in-situ or satellite) and via deterministic models. However, each of these methods has some drawbacks: observations can be accurate but insufficient in terms of temporal and spatial coverage, while deterministic models cover the whole marine ecosystem but can be inaccurate. This work aims at developing a deep learning model to reproduce the biogeochemical variables in the Mediterranean Sea, integrating observations and the output of an existing deterministic model of the marine ecosystem. In particular, two deep learning architectures will be proposed and tested: first EmuMed, an emulator of the deterministic model, and then InpMed, which consists of an improvement of the latter by the addition of information provided by in-situ and satellite observations. Results show that EmuMed can successfully reproduce the output of the deterministic model, while ImpMed can successfully make use of the additional information provided, thus improving our ability to monitor the biogeochemical variables in the Mediterranean Sea.

This paper proposes an Active Learning algorithm that could detect heart attacks based on different body measures, which requires much less data than the passive learning counterpart while maintaining similar accuracy. To that end, different parameters were tested, namely the batch size and the query strategy used. The initial tests on batch size consisted of varying its value until 50. From these experiments, the conclusion was that the best results were obtained with lower values, which led to the second set of experiments, varying the batch size between 1 and 5 to understand in which value the accuracy was higher. Four query strategies were tested: random sampling, least confident sampling, margin sampling and entropy sampling. The results of each approach were similar, reducing by 57% to 60% the amount of data required to obtain the same results of the passive learning approach.

Matrix-Vector multiplications usually represent the dominant part of computational operations needed to propagate information through a neural network. This number of operations can be reduced if the weight matrices are structured. In this paper, we introduce a training algorithm for neural networks with sequentially semiseparable weight matrices based on the backpropagation algorithm. By exploiting the structures in the weight matrices, the computational complexity for computing the matrix-vector product can be reduced to the subquadratic domain. We show that this can lead to computing time reductions on a microcontroller. Furthermore, we analyze the generalization capabilities of neural networks with sequentially semiseparable matrices. Our experiments show that neural networks with structured weight matrices can outperform standard feed-forward neural networks in terms of test prediction accuracy for several real-world datasets.

The recent advent of machine learning as a transforming technology has sparked fears about human inability to comprehend the rational of gradually more complex approaches. Interpretable Machine Learning (IML) was triggered by such concerns, with the purpose of enabling different actors to grasp the application scenarios, including trustworthiness and decision support in highly regulated sectors as those related to health and public services. YOLO (You Only Look Once) models, as other deep Convolutional Neural Network (CNN) approaches, have recently shown remarkable performance in several tasks dealing with object detection. However, interpretability of these models is still an open issue. Therefore, in this work we extend the LIME (Local Interpretable Model-agnostic Explanations) framework to be used with YOLO models. The main contribution is a public add-on to LIME that can effectively improve YOLO interpretability. Results on complex images show the potential improvement.

We present an approach that is able to automatically choose the best meta-heuristic and configuration for solving a real optimization problem. Our approach allows the researcher to indicate which meta-heuristics to choose from and, for each meta-heuristic, which parameters should be automatically configured to find good solutions for the optimization problem. We show that our approach is sound using ten well know real optimization problems and five meta-heuristics. As a side effect, we were also able to provide an unbiased way of assessing meta-heuristics concerning their performance to address one or more classes of real optimization problems. Our approach improved the results found for all the meta-heuristics in all problems and was also able to find very competitive results for all optimization problems when given the liberty to choose which meta-heuristic to use.

Technologies such as Artificial Intelligence (Analytics and Automation) can harm purposely persons (via fake news of social networks, drones, robots, apps, platforms) without any available regulations and forms of protection. We need not only benefits to offer to everybody but responsible ways to develop all new intelligent systems, and agents without high risks and strange behaviours. In many cases, decisions are not intelligible to humans and easy explanations are not available anywhere. We want diverse technologies to aid people and deliver great advantages to society at large.

We consider the task of using a neural network for controlling a quadrotor drone to perform flight maneuvers. For that, the network must be evaluated with high frequency on the microcontroller of the drone. In order to maintain the evaluation frequency for larger networks, we search for structures in the weight matrices of the trained network. By exploiting structures in the weight matrices, the propagation of information through the network can be made more efficient. In this paper, we focus on four structure classes, namely low rank matrices, matrices of low displacement rank, sequentially semiseparable matrices and products of sparse matrices. We approximate the trained weight matrices with matrices from each structure class and analyze the flying capabilities of the approximated neural network controller. Our results show that there is structure in the weight matrices, which can be exploited to speed up the inference, while still being able to perform the flight maneuvers in the real world. The best results were obtained with products of sparse matrices, which could even outperform non-approximated networks with the same number of parameters in some cases.

This paper describes a full navigation architecture which includes a set of available Deep Learning-based modules, focused on “speech to text” and “text to speech” translation, and face recognition, for enabling natural interaction between a smart mobile assistant robot and its human users, in a context of autonomous navigation. The system is novel because it allows complex spoken commands to be syntactically analyzed in Spanish and transformed into motion plans, ready to be executed by the robot, by using the well-known Navigation stack included in the ROS ecosystem. A novel computationally efficient approach (to semantically label the free space from raw data provided by a low cost laser scanner device), enables the generation of a labelled polygonal map, which enhances fixed and mobile obstacle avoidance and local planning in real indoor environments. Robot configuration, learned maps, and Deep Learning models adapted to each scenario are safely and privately stored in Google Cloud, allowing the robot to adapt its behavior to different users and settings. The tests which demonstrate the performance of the system, both in simulation and real environments, are also described.

Most smart city sensors generate time series records and forecasting such data can provide valuable insights for citizens and city managers. Within this context, the adoption of Automated Time Series Forecasting (AutoTSF) tools is a key issue, since it facilitates the design and deployment of multiple TSF models. In this work, we adapt and compare eight recent AutoTSF tools (Pmdarima, Prophet, Ludwig, DeepAR, TFT, FEDOT, AutoTs and Sktime) using nine freely available time series that can be related with the smart city concept (e.g., temperature, energy consumption, city traffic). An extensive experimentation was carried out by using a realistic rolling window with several training and testing iterations. Also, the AutoTSF tools were evaluated by considering both the predictive performances and required computational effort. Overall, the FEDOT tool presented the best overall performance.

This work proposes a spatiotemporal clustering approach for the analysis of coastal upwelling from Sea Surface Temperature (SST) grid maps derived from satellite images. The algorithm, Core-Shell clustering, models the upwelling as an evolving cluster whose core points are constant during a certain time window while the shell points move through an in-and-out binary sequence. The least squares minimization of clustering criterion allows to derive key parameters in an automated way. The algorithm is initialized with an extension of Seeded Region Growing offering self-tuning thresholding, the STSEC algorithm, that is able to precisely delineate the upwelling region at each SST instant map. Yet, the application of STSEC to the SST grid maps as temporal data puts the business of finding relatively stable “time windows”, here called “time ranges”, for obtaining the core clusters onto an automated footing. The experiments conducted with three yearly collections of SST data of the Portuguese coast shown that the core-shell clusters precisely recognize the upwelling regions taking as ground-truth the STSEC segmentations with Kulczynski similarity score values higher than 98%. Also, the extracted time series of upwelling features presented consistent regularities among the three independent upwelling seasons.

AutoML appeared in the last few years as the ultimate challenge in the field of machine learning and data science. However, despite the advances on hyper-parameter optimization, the data preparation step continues to face great difficulties, mainly due to the inability to incorporate human expertise on variables reengineering. In this paper, we present an algorithm able to automate the trivial preparation tasks and to generate features using domain knowledge, represented through entity-relationship (ER) diagrams. Along with the algorithm, we define a set of operators that can be applied to distinct kinds of data, with small human intervention. The algorithm is evaluated over a small set of public datasets, for which we designed basic ER models. The new method shows results comparable to the ones achieved with other automation tools, such as AutoSklearn [4], but with much lower processing times.

In hospital and after ICU discharge deaths are usual, given the severity of the condition under which many of them are admitted to these wings. Because of this, there is an urge to identify and follow these cases closely. Furthermore, as ICU data is usually composed of variables measured in varying time intervals, there is a need for a method that can capture causal relationships in this type of data. To solve this problem, we propose ItsPC, a causal Bayesian network that can model irregular multivariate time-series data. The preliminary results show that ItsPC creates smaller and more concise networks while maintaining the temporal properties. Moreover, its irregular approach to time-series can capture more relationships with the target than the Dynamic Bayesian Networks.

Competitive Intelligence allows an organization to keep up with market trends and foresee business opportunities. This practice is mainly performed by analysts scanning for any piece of valuable information in a myriad of dispersed and unstructured sources. Here we present MapIntel, a system for acquiring intelligence from vast collections of text data by representing each document as a multidimensional vector that captures its own semantics. The system is designed to handle complex Natural Language queries and visual exploration of the corpus, potentially aiding overburdened analysts in finding meaningful insights to help decision-making. The system searching module uses a retriever and re-ranker engine that first finds the closest neighbors to the query embedding and then sifts the results through a cross-encoder model that identifies the most relevant documents. The browsing module also leverages the embeddings by projecting them onto two dimensions while preserving the original landscape, resulting in a map where semantically related documents form topical clusters which we capture using topic modeling. This map aims at promoting a fast overview of the corpus while allowing a more detailed exploration and interactive information encountering process. In this work, we evaluate the system and its components on the 20 newsgroups dataset and demonstrate the superiority of Transformer-based components.

The repair process of devices is an important part of the business of many original equipment manufacturers. The consumption of spare parts, during the repair process, is driven by the defects found during inspection of the devices, and these parts are a big part of the costs in the repair process. But current Supply Chain Control Tower solutions do not provide support for the automatic check of spare parts consumption in the repair process.In this paper, we investigate a multi-label classification problem and present a learning-to-rank approach, where we simulate the passage of time while training hundreds of Logistic Regression Machine Learning models to provide an automatic check in the consumption of spare parts.The results show that the trained models can achieve a mean NDCG@20 score of 81% when ranking the expected parts, while also marking a low volume of 10% of the consumed parts for alert generation. We briefly discuss how these marked parts can be aggregated and combined with additional data to generate more fine-grained alerts.

Churn and how to deal with it is an essential issue in the telecommunications sector. Within the scope of actionable knowledge, we argue that it is crucial to find effective personalized interventions that can lead to a reduction in dropouts and that, at the same time, make it possible to determine the causal effect of these interventions. Considering an intervention that encourages clients to opt for a longer-term contract for benefits, we used Uplift modeling and the Transformed Outcome Approach as a machine learning-based technique for individual-level prediction. The result is actionable profiles of persuadable customers that increase retention and strike the right balance between the campaign budget.

The development of analytical systems imposes several challenges related not only to the amount and heterogeneity of the involved data but also to the constant need to readapt and evolve to overcome new business challenges. The data modelling layer represents the mapping between the domain and technical knowledge, however, to organize raw data into a form that can be used for analytics, specific Extract, Transform and Load (ETL) processes should be applied. ETL systems are recognized as a critical and tightly coupled system component that encapsulates data-level requirements that are hard to implement and maintain. In a Big Data era, adaptability and extensibility are important characteristics to hold when developing analytical systems. Thus, to provide more consistent, reliable, flexible, and reusable ETL processes, a service-oriented implementation for ETL development is proposed.

Analyzing the way individuals move is fundamental to understand the dynamics of humanity. Transportation mode plays a significant role in human behavior as it changes how individuals travel, how far, and how often they can move. The identification of transportation modes can be used in many applications and it is a key component of the internet of things (IoT) and the Smart Cities concept as it helps to organize traffic control and transport management. In this paper, we propose the use of ensemble methods to infer the transportation modes using raw GPS data. From latitude, longitude, and timestamp we perform feature engineering in order to obtain more discriminative fields for the classification. We test our features in several machine learning algorithms and among those with the best results we perform feature selection using the Boruta method in order to boost our accuracy results and decrease the amount of data, processing time, and noise in the model. We assess the validity of our approach on a real-world dataset with several different transportation modes and the results show the efficacy of our approach.

Concept refinement operators have been introduced to describe and compute generalisations and specialisations of concepts, with, amongst others, applications in concept learning and ontology repair through axiom weakening. We here provide a probabilistic proof of almost-certain termination for iterated refinements, thus for an axiom weakening procedure for the fine-grained repair of $$\mathcal {ALC}$$ ALC ontologies. We determine the computational complexity of refinement membership, and discuss performance aspects of a prototypical implementation, verifying that almost-certain termination means actual termination in practice.

In this paper we present DeMaxSAT, a memetic algorithm for solving the non-partial MaxSAT problem. It combines the evolutionary algorithm of Differential Evolution with GSAT and RandomWalk, two MaxSAT-specific local search heuristics. An implementation of the algorithm has been used to solve the benchmarks for non-partial MaxSAT included in the MaxSAT Evaluation 2021. The performance of DeMaxSAT has reached results that are comparable, both in computing time and quality of the solutions, to the best solvers presented in MaxSAT Evaluation 2021, reaching the state of the art for non-partial problems.

This paper studies belief correction and state transition for ontic actions in a multi-agent epistemic framework. When a full observer agent observes the execution of an action, he will correct his (possibly wrong) initial belief about the precondition of the action as well as his belief about his own observability. The paper shows that correcting beliefs about precondition and observability is vital for observing the effect of the action and robust state transition, highlighting the risk of yielding counter-intuitive results. The paper proposes a state transition function for ontic actions which integrates correcting beliefs for precondition, observability and realizing the effect of the action. This novel transition function does not require event update models. The paper investigates several properties of the transition function, assessing its robustness in ensuring that beliefs of agents change consistently with their degree of observability of action occurrences. Sample scenarios are provided to illustrate the novel transition function.

Multi-adjoint lattice logic (MLL) has been introduced as an axiomatization of multi-adjoint algebras on lattices. This paper highlights the interest of MLL introducing new relevant properties and some interesting examples of how to reasoning with this logic.

Solving long-horizon problems is a desirable property in autonomous agents. Learning reusable behaviours can equip the agent with this property, allowing it to adapt them when performing various real-world tasks. Our approach for learning these behaviours is composed of three modules, operating in two separate timescales and it uses a hierarchical model with both discrete and continuous variables. This modular structure allows an independent training process for each stage. These stages are organized using a two-level temporal hierarchy. The first level contains the planner, responsible for issuing the skills that should be executed, while the second level executes the skill. In this latter level, to achieve the desired skill behaviour, the discrete skill is converted to a continuous vector that contains information regarding which environment change must occur. With this approach, we aimed to solve long-horizon sequential tasks with delayed rewards. Contrary to existing work, our method uses both variable types to allow an agent to learn high-level behaviours consisting of an interpretable set of skills. This method allows to compose the discrete skills easily, while keeping the flexibility, provided by the continuous representations, to execute them in several different ways. Using a 2D scenario where the agent has to catch a set of objects in a specific order, we demonstrate that our approach is scalable to scenarios with increasingly longer tasks.

We consider a fair division model in which agents have general valuations for bundles of indivisible items. We propose two new approximate properties for envy freeness of allocations in this model: DEFX and DEF1. We compare these with two existing axiomatic properties: EFX and EF1. For example, we give the first result confirming that EFX allocations may not exist with general but identical valuations. However, even when they do exist in such problems, we prove that DEFX (and, therefore DEF1) and PO allocations exist whereas EFX and PO allocations may not exist. Our results assert eloquently that DEFX and DEF1 approximate fairness better than EFX and EF1.

A key goal of ad hoc teamwork is to develop a learning agent that cooperates with unknown teams, without resorting to any pre-coordination protocol. Despite a vast number of ad hoc teamwork algorithms in the literature, most of them cannot address the problem of learning to cooperate with a completely unknown team, unless it learns from scratch. This article presents a novel approach that uses transfer learning alongside the state-of-the-art PLASTIC-Policy to adapt to completely unknown teammates quickly. We test our solution within the Half Field Offense simulator with five different teammates. The teammates were designed independently by developers from different countries and at different times. Our empirical evaluation shows that it is advantageous for an ad hoc agent to leverage its past knowledge when adapting to a new team instead of learning how to cooperate with it from scratch.

The digital learning transformation brings the extension of the traditional libraries to online repositories. Learning object repositories are employed to deliver several functionalities related to the learning object’s lifecycle. However, these educational resources usually are not described effectively, lacking, for example, educational metadata and learning goals. Then, metadata incompleteness limits the quality of the services, such as search and recommendation, resulting in educational objects that do not have a proper role in teaching/learning environments. This work proposes to bring an active role to all educational resources, acting on the analysis generated from the usage statistics. To achieve this goal, we created a multi-agent architecture that complements the common repository’s functionalities to improve learning and teaching experiences. We intend to use this architecture on a repository focused on ocean literacy learning objects. This paper presents some steps toward this goal by enhancing, when needed, the repository to adapt itself.

Named Entity Recognition (NER) is a challenging Natural Language Processing task for a language as rich as Portuguese. When applied in a scenario appropriate to informal language and short texts, the task acquires a new layer of complexity, handling a particular lexicon to the domain in question. In this paper, we expanded the UlyssesNER-Br corpus for NER task with Brazilian Portuguese comments about bills. Additionally, we enriched the annotated set with a formal corpora, in order to analyze whether the combination of formal and informal texts from the same domain could improve NER. Finally, we carry out experiments with a Conditional Random Fields (CRF) model, a Bidirectional LSTM-CRF (BiLSTM-CRF) model, and subsequently, we realized fine-tuning of a language model BERT on NER task with our dataset. We concluded that formal texts helped identification of entities in informal texts. The best model was the fine-tuned BERT which achieved an F1-score of 73.90%, beating the benchmark of related works.

Question Generation (QG) is an important and challenging problem that has attracted attention from the natural language processing (NLP) community over the last years. QG aims to automatically generate questions given an input. Recent studies in this field typically use widely available question-answering (QA) datasets (in English) and neural models to train and build these QG systems. As lower-resourced languages (e.g. Portuguese) lack large-scale quality QA data, it becomes a significant challenge to experiment with recent neural techniques. This study uses a Portuguese machine-translated version of the SQuAD v1.1 dataset to perform a preliminary analysis of a neural approach to the QG task for Portuguese. We frame our approach as a sequence-to-sequence problem by fine-tuning a pre-trained language model – T5 for generating factoid (or wh)-questions. Despite the evident issues that a machine-translated dataset may bring when using it for training neural models, the automatic evaluation of our Portuguese neural QG models presents results in line with those obtained for English. To the best of our knowledge, this is the first study addressing Neural QG for Portuguese. The code and models are publicly available at https://github.com/bernardoleite/question-generation-t5-pytorch-lightning .

In this work, we targeted the search engine of a sports-related website that presented an opportunity for search result quality improvement. We reframed the engine as a Federated Search instance, where each collection represented a searchable entity type within the system, using Apache Solr for querying each resource and a Python Flask server to merge results. We extend previous work on individual search term weighing, making use of past search terms as a relevance indicator for user selected documents. To incorporate term weights we define four strategies combining two binary variables: integration with default relevance (linear scaling or linear combination) and search term frequency (raw value or log-smoothed). To evaluate our solution, we extracted two query sets from search logs: one with frequently submitted queries, and another with ambiguous result access patterns. We used click-through information as a relevance proxy and tried to mitigate its limitations by evaluating under distinct IR metrics, including MRR, MAP and NDCG. Moreover, we also measured Spearman rank correlation coefficients to test similarities between produced rankings and reference orderings according to user access patterns. Results show consistency across all metrics in both sets. Previous search terms were key to obtaining a higher effectiveness, with runs that used pure search term frequency performing best. Compared to the baseline, our best strategies were able to maintain quality on frequent queries and improve retrieval effectiveness on ambiguous queries, with up to $$\sim $$ ∼ six percentage points better performance on most metrics.