Ebook: Recent Advances in Artificial Intelligence Research and Development
Artificial intelligence in all its forms is increasingly interwoven into all our lives, and remains one of the most lively areas of discussion and interest in technology today.
This book presents the proceedings of the 20th International Conference of the Catalan Association for Artificial Intelligence (CCIA’2017): ‘Recent Advances in Artificial Intelligence Research and Development’, held in Deltebre, Terres de l'Ebre, Spain, in October 2017. Despite its title, this annual conference is not only for researchers from the Catalan Countries, but is an international event which attracts participants from countries all around the world. In total, 41 original contributions were submitted to CCIA’2017. Of these, 21 were accepted as long papers for oral presentation and 13 were accepted as short papers to be presented as posters. These 34 submissions appear in this book organized around a number of different topics including: agents and multi-agent systems; artificial vision and image processing; machine learning; artificial neural networks; cognitive modeling; fuzzy logic and reasoning; robotics; and AI applications. The book also includes abstracts of the 3 presentations by invited speakers.
The book offers a representative sample of the current state of the art in the artificial intelligence community, and will be of interest to all those working with AI worldwide.
The Catalan Association for Artificial Intelligence (ACIA), founded in 1994 as a non-profit association, aims primarily at fostering the cooperation among researchers from the Catalan-speaking Artificial Intelligence research community.
It has been a long way since the first edition of the Catalan Conference of Artificial Intelligence in 1998, held in Tarragona, up until today. By that time, the main goal was to create a forum that allowed communication and research collaboration throughout the Catalan AI community. This objective has been achieved and actually has widened at a constant rate over the last 20 years. The International Conference of the Catalan Association for Artificial Intelligence (CCIA), the conference currently organized by ACIA, is an international event not only for researchers in the Catalan Countries, but also involving researchers from other countries all around the world.
The 20th edition of CCIA is this year held in Deltebre, Terres de l'Ebre, Spain, 25–27 October 2017, under joint organization by the Polytechnic University of Catalonia (UPC, BarcelonaTech) and the University of the Balearic Islands (UIB). A total of 41 original contributions were submitted to CCIA'2017. Of these 41 submissions, 21 were accepted as long papers for oral presentation (10-page papers) and 13 were accepted as short papers to be presented as posters (6-page papers). These 34 submissions appear in this book due to their relevance, originality and technical soundness.
All these papers have been organized around a number of different topics such as Agents and multi-agent systems, Artificial vision and Image processing, Machine learning, Artificial neural networks, Cognitive modelling, Fuzzy logic and reasoning, Robotics, and AI Applications, being not only a representative sample of the current state of the art in the Catalan Artificial Intelligence Community, but also showing the collaboration between the ACIA members and the worldwide AI community.
This volume also includes the abstracts of the invited talks by three outstanding researchers:
• Humberto Bustince, Full Professor in the Department of Automatics and Computation at the Public University of Navarra (UPNA),
• Miquel Domènech, Senior Lecturer in Social Psychology at the Autonomous University of Barcelona (UAB), and
• Alex Rayón, Vicedean for External Affairs and Lifelong Learning at the University of Deusto.
We would like to express our sincere gratitude to the authors of the contributed papers, to the invited speakers for their enlightening talks, to all members of the Program and Organizing Committees who have worked hard to make this conference a success, to Angela Nebot, former scientific Program Chair of CCIA'2016, for her help during this past year, and to Josep Pujol, president of ACIA, for her kind support.
Isabel Aguiló, University of the Balearic Islands (UIB)
René Alquézar, Polytechnic University of Catalonia (UPC)
Cecilio Angulo, Polytechnic University of Catalonia (UPC)
Alberto Ortiz, University of the Balearic Islands (UIB)
Joan Torrens, University of the Balearic Islands (UIB)
A 10% of diabetic people may suffer from a disease called Diabetic Retinopathy that destroys the retina causing blindness. An early detection of any sign that indicates a risk of developing DR is important, because there are treatments that can significantly improve the quality of life and minimize the risk of vision loss. Very few work exists on exploiting the information stored in the Electronic Health Record for this purpose. In this work, we propose the integration of two machine learning methods that generate a set of fuzzy classification rules. We show that the combination of two methods improves the classification quality indices.
Diabetes is a major cause of blindness. Patients with history of diabetes are more prone to diabetic retinopathy (DR), which is the damage of the retina due to diabetes. Computer-aided diagnosis (CAD) systems have been used to detect DR early, which can reduce the occurrence of blindness. In this paper we analyze the performance of robust texture analysis methods in order to improve the performance of DR CAD systems. We used six texture analysis methods: the color autocorrelogram, local binary pattern, histogram of oriented gradients, local directional number, co-occurrence matrix features and Gabor filters. We also propose new combinations of different texture analysis methods to further improve the detection results. The proposed CAD system gave promising DR detection results and it is on par with related methods.
The interaction between patients and professionals in complex clinical domains, as in the case of Neurorehabilitation, is always a complex process where crucial decision making in a short period of time is required, and where every decision has a serious impact on the patient. In this situation, deciding which are the most appropriate interventions is not an easy task because these patients simultaneously present several impairments, multiple diagnoses, and required complex interdisciplinary approaches. In this context, a methodology and a tool based on ICF have been developed to explore the relationships between patient impairments and therapeutic goals. The proposed approach, based on graph analysis, was used to analyze a set of 1960 patients that suffered an Acquired Brain Injury. Results achieved show that the proposed methodology is able to find non-explicit relationships. This study constitute a first step to the goal of designing a clinical decision support tool for neurorehabilitation.
In a recent work we have developed an analysis system based on valued abstract argumentation to model and reason about the accepted and rejected tweets of a Twitter discussion. Given a Twitter discussion, the system builds an argument graph where each node denotes a tweet and each edge denotes a criticism relationship between a pair of tweets of the discussion. In the social network Twitter, a tweet always refers to previous tweets in the discussion, so the obtained underlying argument graph is acyclic. In this work we introduce and investigate a natural extension of the system, referred as author-centered analysis system, in which tweets within a discussion are grouped by authors, such that tweets of a same author represent his/her opinion in the discussion with a single node in the graph, and criticism relationships denote controversies between the opinions of Twitter users in the discussion. With this new approach, the interactions between authors can give rise to circular criticism relationships and thus the underlying argument graph can contain cycles. The output of the author-centered analysis system is the set of authors such that their opinions are consistent and are globally accepted within a Twitter discussion. The system can be of special relevance for assessing Twitter discussions in fields where identifying groups of authors whose opinions are globally compatible or consistent, but at the same time are widely accepted, is of particular interest.
According to the sensorimotor approach, cognition is constituted by regularities in the perceptual experiences of an active and situated agent. This perspective rejects traditional inner representational models, stressing instead patterns of sensorimotor dependencies. Those relations are called sensorimotor contingencies (SMC). Many research areas and accounts are working on and related with it. In particular, four distinct kinds of SMCs have been previously introduced for environment, habitat, coordination and strategy using dynamical models from a psychological perspective. As dynamical systems, in this paper we analyze SMCs, for the very first time, from a modern control engineering perspective. We provide equations and block diagrams translating the psychological proposal to control engineering. We also analyze the original toy example proposed from the psychological domain into the modern control engineering point of view, as well as we propose a first approach to this toy example coming from the control engineering domain.
Traversing large search spaces can be done more efficiently by exploiting the dead-ends –in formal terms nogoods– discovered during search. If a previously found nogood appears again, the search process can avoid it, saving some search effort. Storing all found nogoods may require exponential memory, which is unaffordable. However, current memories allow to store a large set of nogoods, to be maintained during the solving process. In many cases, a solution is found before memory is exhausted. In the context of Distributed Constraint Satisfaction, the AWC algorithm allows to compute a solution quickly but, to guarantee completeness, it requires storing all found nogoods. Trading space per time, we develop a new iterative version of the algorithm that delays the exponential effects. We present this new version in the context of distributed SAT, where agents hold several Boolean variables. Taking advantage of existing SAT technology, this version perform calls to external MaxSAT solver. Experimentally, we confirm the benefits of the proposed approach on several benchmarks.
Target-dependent sentiment analysis on Twitter is the problem of identifying the sentiment polarity towards a certain target in a given tweet. All the existing studies of this task assume that the target is known. However, in such tasks, extracting the targets from the text is one of the most important subtasks. In this paper, we propose a model based on Bidirectional Gated Recurrent Units and Conditional Random Fields to identify automatically the targets from the tweets. The model has been evaluated on two benchmarks of tweets, obtaining results which show its superiority over several baseline methods.
Breast cancer can be detected at early stages by radiologists from periodic screening mammography. However, just by viewing the mammogram they cannot discern the subtype of the cancer (Luminal A, Luminal B, Her-2+ and Basal-like), which is a crucial information for the oncologist to decide the appropriate therapy. Consequently, a painful biopsy must be carried out for determining the tumor subtype from cytological and histological analysis of the extracted tissue. In this paper, we aim to design a computer aided diagnosis (CAD) system able to classify the four tumor subtypes just from the image pixels of digital mammography. The proposed strategy is to use a VGGNet-based deep learning convolutional neural network (CNN) that can be trained to learn the underlying micro-texture pattern of image pixels, expected to be characteristic of each subtype. We have collected 716 image samples of 100x100 pixels wide, manually extracted from real tumor image areas that had been labeled in the digital mammography by a radiologist, jointly with the corresponding oncologist diagnose based on histological indicators. Using this ground truth, we have been able to train and test the proposed CNN, which can achieve an accuracy rate of 78% when discerning only Luminal A and Luminal B subtypes. In turn, it yields an accuracy rate of 67% when all four tumor subtypes are considered.
In this paper we first consider the problem of extending a fuzzy preference relation on a set, represented by a fuzzy preorder, to a fuzzy preference relation on subsets, and we characterise different possibilities. Based on their properties, we then semantically define and axiomatize a two-tiered graded modal logic to reason about a notion of fuzzy preferences.
This work presents a content-based recommender system for machine learning classifier algorithms. Given a new data set, a recommendation of what classifier is likely to perform best is made based on classifier performance over similar known data sets. This similarity is measured according to a data set characterization that includes several state-of-the-art metrics taking into account physical structure, statistics, and information theory. A novelty with respect to prior work is the use of a robust approach based on permutation tests to directly assess whether a given learning algorithm is able to exploit the attributes in a data set to predict class labels, and compare it to the more commonly used F-score metric for evaluating classifier performance. To evaluate our approach, we have conducted an extensive experimentation including 8 of the main machine learning classification methods with varying configurations and 65 binary data sets, leading to over 2331 experiments. Our results show that using the information from the permutation test clearly improves the quality of the recommendations.
Breast cancer is one of the most dangerous diseases that threaten women. Computer-aided diagnosis systems can be used to analyse breast images in order to detect breast tumours early. This paper proposes a new method for feature learning from breast cancer images. The learned features are used to discriminate between malignant and benign tumours. We marry the traditional feature extraction methods with deep learning approaches, in the sense that we automatically learn local descriptors from the input images without hand-crafting the feature extraction methods. The proposed technique uses a bio-inspired optimization method, called whale optimization algorithm, to learn local descriptors from the input images. The learned features are tuned to the input images and can describe the local/global arrangement of pixels in them. Unlike the usual deep learning approaches, small datasets can be used to train the proposed method. To validate the proposed feature learning method, we used two breast cancer datasets: mammographic and ultrasound images for benign and malignant cases. The experimental results demonstrate that the proposed method gives good classification results compared with the state of the art ones.
G Protein-Coupled Receptors are a family of cell membrane proteins, whose class C is a current main target for drug development. Their primary sequences are studied as a source of information for the characterization of their behaviour. In previous research, different alignment-free sequence transformations were explored as the basis for the supervised discrimination of the various class C subtypes using Support Vector Machines. We also investigated an alignment-free sequence transformation based on n-gram protein motifs, under the hypothesis that the sequences' extra-cellular N-terminus domain could suffice to retain most of the subtype-discrimination capabilities of the complete receptor, and that a parsimonious selection of n-grams would be responsible for such classification success. In the current study, these previous results are extended by investigating a different classification procedure that now employs a subtype-vs-all the rest of subtypes approach, shifting towards the selection of those sequence motifs that distinguish each class C subtype from the rest. The reported results indicate the adequacy of this new approach, both in terms of discrimination ability and motif selection parsimony.
Precision in predicting people's next location plays a key role in the success of advanced Location-Based Services (LBSs). Models in the literature have achieved satisfactory results, but they lack exploiting of timestamps-sensitive property while combining it with locations sequences. In this paper, we propose a location prediction model in which time encoding scheme is proposed to capture movement behavior characteristics. Embedding learning technique and neural pooling functions are used to extract the semantic information of input data. A set of neural pooling functions are explored in order to extract rich features. Recurrent Neural Network (RNN) is utilized in the proposed model in order to keep track of user movement history which allows to discover more meaningful dependencies. As consequence, the model performance is enhanced. Evaluations on a large real life dataset demonstrate that the proposed model outperforms state-of-the-art models in terms of Recall, Precision and F1-score performance metrics.
How set theory and interval distances can be combined to compare qualitative concepts defined on reference systems at different levels of granularity is studied in this paper. A measure of similarity to compare qualitative concepts and reference systems is presented and implemented using first order logics. A case study with different angle descriptors is used for testing and promising results are obtained.
In this paper, a novel region-based active contour method is proposed based to both correct and segment the intensity inhomogeneous images. A phase stretch transform (PST) kernel is used to compute new intensity means and bias field, which are employed to define a bias fitted image. In the proposed energy function, a new signed pressure force (SPF) function is formulated with a bias image fitted difference, which helps to segment the intensity inhomogeneous objects. A Gaussian kernel is also used to regularize the level set curve, which also removes the computationally expensive re-initialization. Finally, the proposed method is compared with the state-of-the-art both qualitatively and quantitatively using the synthetic and real brain magnetic resonance (MR) images, which shows it yields the best segmentation and correction results.
In human smart nutrition systems, environment based food classification has become popular to help analyzing the food intake based on the nutrition related activity. In this paper, we address the problem of food related environments, which refer to different eating places such as, bars, restaurants, coffee shops, etc. using state-of-the-art convolutional neural networks (CNNs). We collected a new dataset on different food related environments by integrating three publicly available datasets: Places365, ImageNet and SUN397. We have named it “FoodPlaces” and it contains 35 different types of classes. In order to achieve satisfactory results on the food related environment recognition, we fine-tuned several state-of-the-art CNNs, such as VGG16, RsNet50 and InceptionV3 using different transfer learning approaches. The results show that the fully fine-tunned InceptionV3 yields 75.22% classification accuracy among the discussed state-of-the-art CNNs.
Prevention and nutrition are key issues to guarantee a healthy lifestyle and are in the kernel of the new paradigm of patient-centered medicine. A healthy diet protects against risk factors of a large number of chronic diseases and contribute to delay the disease appearance in the general population. More and more, making right diets is becoming essential from the health point of view. However, nutritionists design diets on the specific needs of each patient, using their accumulated experience and there are no well-formalized support mechanisms for such activity. The project Diet4You proposes the creation of an intelligent decision support system oriented to the adaptive and dynamic preparation of personalized diets for the specific individuals of general population, having or not, one or more diseases, by taking into account all the information available on the person. The diets will be built by considering the characteristics of the person, their health conditions, their habits and eventual drugs intake and their genetic information, which until now were not taken into account. The Diet4You tool is a hybrid system interacting several components in a complex way. One of those components is a subsystem able to compose complete menus, for a certain period of time, based on the nutritional prescriptions given by the nutritionists, in terms of balance of different families of nutrients. This component is based on a a Case-Based Reasoning (CBR) engine that adaptively creates menus with breakfast, lunch and dinner on the bases of real food databases containing dishes with their nutritional composition. In this work, this component of the whole system is presented, described and tested.
Corrosion is one of the major causes of the structural defects that affect vessel hulls. For its early detection, intensive inspections of the inner and outer structures of the vessel hull are carried out at a great cost, where the visual assessment plays an important role. In order to reduce the cost of the visual inspections, we present a corrosion detector to identify defective areas in digital images taken from vessel hulls. Two main contributions stand out: on the one hand, a specific detector which combines color and texture features to describe corrosion; on the other hand, a prior stage which implements a generic-defect search based on the concept of saliency, and it is used to boost the specific corrosion detector. Both the original and the saliency-boosted methods provide successful detection rates, but the guidance by means of saliency allows for precision improvements.
In this paper, we rethink the way Natural Deduction systems are usually used in logic, and adapt them in order to solve the MaxSAT problem. By adapting some existing rules and defining some new ones, the Natural Deduction method can be used to provide an elegant and straightforward solution to maximum satisfiability problems of multisets of Boolean and many-valued clauses.
Within the machine learning framework, incremental learning of multivariate spaces is of special interest for on-line applications. In this work, the regression problem for multivariate systems is solved by implementing an efficient probabilistic incremental algorithm. It allows learning high-dimensional redundant non-linear maps by the cumulative acquisition of data from input-output systems. The proposed model is aimed at solving prediction and inference problems. The implementation introduced in this work allows learning from data batches without the need of keeping them in memory afterwards. The learning architecture is built using Incremental Gaussian Mixture Models. The Expectation-Maximization algorithm and general geometric properties of Gaussian distributions are used to train the models. Our current implementation can produce accurate results fitting models in real multivariate systems. Results are shown from testing the algorithm for both situations, one where the incremental learning is demonstrated and the second where the performance to solve the regression problem is evaluated on a toy example.
Several authors have shown that the sounds of anurans can be used as an indicator of climate change. For this purpose anuran sound automatic classification has become an important issue for biologists and other climate scientists. In this paper two approaches have been used to feature every sound frame: Mel Frequency Cepstral Coefficients (MFCC); and parameters based on the MPEG-7 standard. The methods to extract the features and to classify the sounds are described. Up to ten different classification algorithms have also been considered. Their results are compared using several metrics, mainly based on the classification error rate. The main conclusion is that both, the MFCC and the MPEG-7 parameters, are adequate for featuring and classifying anuran sounds, although MFCC get best metrics for more classifiers. When the number of parameters is an important concern, MPEG-7 features show better results. Additionally, from a qualitative point of view, MPEG-7 features are semantically richer than their MFCC counterparts.